Merge changes from github.
PiperOrigin-RevId: 189231636
diff --git a/README.md b/README.md
index ef5bdc6..3cdb6e4 100644
--- a/README.md
+++ b/README.md
@@ -22,6 +22,10 @@
networks research. The system is general enough to be applicable in a wide
variety of other domains, as well.
+Keep up to date with release announcements and security updates by
+subscribing to
+[announce@tensorflow.org](https://groups.google.com/a/tensorflow.org/forum/#!forum/announce).
+
## Installation
*See [Installing TensorFlow](https://www.tensorflow.org/get_started/os_setup.html) for instructions on how to install our release binaries or how to build from source.*
diff --git a/SECURITY.md b/SECURITY.md
index fea24b2..2aaa920 100644
--- a/SECURITY.md
+++ b/SECURITY.md
@@ -6,7 +6,7 @@
## TensorFlow models are programs
-TensorFlow's runtime system interprets and executes programs. What machine
+TensorFlow's runtime system interprets and executes programs. What machine
learning practitioners term
[**models**](https://developers.google.com/machine-learning/glossary/#model) are
expressed as programs that TensorFlow executes. TensorFlow programs are encoded
@@ -28,12 +28,12 @@
inference on the data.
**TensorFlow models are programs, and need to be treated as such from a security
-perspective.**
+perspective.**
## Running untrusted models
As a general rule: **Always** execute untrusted models inside a sandbox (e.g.,
-[nsjail](https://github.com/google/nsjail)).
+[nsjail](https://github.com/google/nsjail)).
There are several ways in which a model could become untrusted. Obviously, if an
untrusted party supplies TensorFlow kernels, arbitrary code may be executed.
@@ -109,11 +109,11 @@
graphs using untrusted inputs as described above, but they would not be able to
execute arbitrary graphs. It is possible to safely expose a `ModelServer`
directly to an untrusted network, **but only if the graphs it is configured to
-use have been carefully audited to be safe**.
+use have been carefully audited to be safe**.
Similar to best practices for other servers, we recommend running any
`ModelServer` with appropriate privileges (i.e., using a separate user with
-reduced permisisons). In the spirit of defense in depth, we recommend
+reduced permissions). In the spirit of defense in depth, we recommend
authenticating requests to any TensorFlow server connected to an untrusted
network, as well as sandboxing the server to minimize the adverse effects of
any breach.
@@ -133,7 +133,7 @@
can perform arbitrary computations means that they may read and write files,
communicate via the network, produce deadlocks and infinite loops, or run out
of memory. It is only when these behaviors are outside the specifications of the
-operations involved that such behavior is a vulnerability.
+operations involved that such behavior is a vulnerability.
A `FileWriter` writing a file is not unexpected behavior and therefore is not a
vulnerability in TensorFlow. A `MatMul` allowing arbitrary binary code execution
@@ -168,7 +168,7 @@
Please use a descriptive subject line for your report email. After the initial
reply to your report, the security team will endeavor to keep you informed of
-the progress being made towards a fix and announcement.
+the progress being made towards a fix and announcement.
If you believe that an existing (public) issue is security-related, please send
an email to `security@tensorflow.org`. The email should include the issue ID and
diff --git a/configure.py b/configure.py
index 97f4675..d14edef 100644
--- a/configure.py
+++ b/configure.py
@@ -1048,7 +1048,10 @@
for lib_file in possible_files:
if is_compatible(lib_file, cuda_ver, cudnn_ver):
- ver_str = nvinfer_pattern.search(lib_file).group(1)
+ matches = nvinfer_pattern.search(lib_file)
+ if len(matches.groups()) == 0:
+ continue
+ ver_str = matches.group(1)
ver = convert_version_to_int(ver_str) if len(ver_str) else 0
if ver > highest_ver[0]:
highest_ver = [ver, ver_str, lib_file]
@@ -1377,7 +1380,7 @@
# environment variables.
environ_cp = dict(os.environ)
- check_bazel_version('0.5.4')
+ check_bazel_version('0.10.0')
reset_tf_configure_bazelrc(args.workspace)
cleanup_makefile()
diff --git a/tensorflow/compiler/xla/service/generic_transfer_manager.cc b/tensorflow/compiler/xla/service/generic_transfer_manager.cc
index 78dc0ad..a99e2b7 100644
--- a/tensorflow/compiler/xla/service/generic_transfer_manager.cc
+++ b/tensorflow/compiler/xla/service/generic_transfer_manager.cc
@@ -38,14 +38,7 @@
GenericTransferManager::GenericTransferManager(se::Platform::Id platform_id,
size_t pointer_size)
- : platform_id_(platform_id), pointer_size_(pointer_size) {
- // We currently only support kHostPlatformId for CPU, kCudaPlatformId for
- // GPU and kInterpreterPlatformId for Interpreter. Before supporting other
- // platforms, we need to test this transfer manager on them.
- CHECK(platform_id_ == se::host::kHostPlatformId ||
- platform_id_ == se::interpreter::kInterpreterPlatformId ||
- platform_id_ == se::cuda::kCudaPlatformId);
-}
+ : platform_id_(platform_id), pointer_size_(pointer_size) {}
se::Platform::Id GenericTransferManager::PlatformId() const {
return platform_id_;
diff --git a/tensorflow/compiler/xla/tests/convolution_test.cc b/tensorflow/compiler/xla/tests/convolution_test.cc
index 99640f5..7271539 100644
--- a/tensorflow/compiler/xla/tests/convolution_test.cc
+++ b/tensorflow/compiler/xla/tests/convolution_test.cc
@@ -723,7 +723,7 @@
);
#endif
-TEST_F(ConvolutionTest, Convolve_bf16_1x1x1x2_1x1x1x2_Valid) {
+XLA_TEST_F(ConvolutionTest, Convolve_bf16_1x1x1x2_1x1x1x2_Valid) {
ComputationBuilder builder(client_, TestName());
Shape input_shape = ShapeUtil::MakeShape(BF16, {1, 1, 1, 2});
Shape filter_shape = ShapeUtil::MakeShape(BF16, {1, 1, 1, 2});
diff --git a/tensorflow/contrib/BUILD b/tensorflow/contrib/BUILD
index bab37e8..986b61b 100644
--- a/tensorflow/contrib/BUILD
+++ b/tensorflow/contrib/BUILD
@@ -8,6 +8,7 @@
load("//third_party/mpi:mpi.bzl", "if_mpi")
load("@local_config_cuda//cuda:build_defs.bzl", "if_cuda")
load("@local_config_tensorrt//:build_defs.bzl", "if_tensorrt")
+load("//tensorflow:tensorflow.bzl", "if_not_windows")
py_library(
name = "contrib_py",
@@ -51,7 +52,6 @@
"//tensorflow/contrib/image:single_image_random_dot_stereograms_py",
"//tensorflow/contrib/input_pipeline:input_pipeline_py",
"//tensorflow/contrib/integrate:integrate_py",
- "//tensorflow/contrib/kafka",
"//tensorflow/contrib/keras",
"//tensorflow/contrib/kernel_methods",
"//tensorflow/contrib/kfac",
@@ -63,7 +63,6 @@
"//tensorflow/contrib/linalg:linalg_py",
"//tensorflow/contrib/linear_optimizer:sdca_estimator_py",
"//tensorflow/contrib/linear_optimizer:sdca_ops_py",
- "//tensorflow/contrib/lite/python:lite",
"//tensorflow/contrib/lookup:lookup_py",
"//tensorflow/contrib/losses:losses_py",
"//tensorflow/contrib/losses:metric_learning_py",
@@ -110,6 +109,10 @@
"//tensorflow/python:util",
] + if_mpi(["//tensorflow/contrib/mpi_collectives:mpi_collectives_py"]) + if_tensorrt([
"//tensorflow/contrib/tensorrt:init_py",
+ ]) + if_not_windows([
+ "//tensorflow/contrib/ffmpeg:ffmpeg_ops_py", # unix dependency, need to fix code
+ "//tensorflow/contrib/lite/python:lite", # unix dependency, need to fix code
+ "//tensorflow/contrib/kafka", # has some linking issue on opensssl.
]),
)
@@ -121,6 +124,7 @@
"//tensorflow/contrib/coder:all_kernels",
"//tensorflow/contrib/cudnn_rnn:cudnn_rnn_kernels",
"//tensorflow/contrib/data/kernels:dataset_kernels",
+ "//tensorflow/contrib/kafka:dataset_kernels",
"//tensorflow/contrib/factorization/kernels:all_kernels",
"//tensorflow/contrib/input_pipeline:input_pipeline_ops_kernels",
"//tensorflow/contrib/layers:sparse_feature_cross_op_kernel",
@@ -147,7 +151,7 @@
"//tensorflow/contrib/factorization:all_ops",
"//tensorflow/contrib/framework:all_ops",
"//tensorflow/contrib/input_pipeline:input_pipeline_ops_op_lib",
- "//tensorflow/contrib/kafka:kafka_ops_op_lib",
+ "//tensorflow/contrib/kafka:dataset_ops_op_lib",
"//tensorflow/contrib/layers:sparse_feature_cross_op_op_lib",
"//tensorflow/contrib/nccl:nccl_ops_op_lib",
"//tensorflow/contrib/nearest_neighbor:nearest_neighbor_ops_op_lib",
diff --git a/tensorflow/contrib/__init__.py b/tensorflow/contrib/__init__.py
index 4f6f539..669d611 100644
--- a/tensorflow/contrib/__init__.py
+++ b/tensorflow/contrib/__init__.py
@@ -18,6 +18,8 @@
from __future__ import division
from __future__ import print_function
+import os
+
# Add projects here, they will show up under tf.contrib.
from tensorflow.contrib import batching
from tensorflow.contrib import bayesflow
@@ -83,7 +85,8 @@
from tensorflow.contrib import training
from tensorflow.contrib import util
from tensorflow.contrib.eager.python import tfe as eager
-from tensorflow.contrib.lite.python import lite
+if os.name != 'nt':
+ from tensorflow.contrib.lite.python import lite
from tensorflow.contrib.receptive_field import receptive_field_api as receptive_field
from tensorflow.contrib.remote_fused_graph import pylib as remote_fused_graph
from tensorflow.contrib.specs import python as specs
@@ -92,6 +95,7 @@
from tensorflow.python.util.lazy_loader import LazyLoader
ffmpeg = LazyLoader("ffmpeg", globals(),
"tensorflow.contrib.ffmpeg")
+del os
del LazyLoader
del absolute_import
diff --git a/tensorflow/contrib/boosted_trees/lib/utils/batch_features.h b/tensorflow/contrib/boosted_trees/lib/utils/batch_features.h
index da5e744..7815fa0 100644
--- a/tensorflow/contrib/boosted_trees/lib/utils/batch_features.h
+++ b/tensorflow/contrib/boosted_trees/lib/utils/batch_features.h
@@ -48,9 +48,9 @@
Status GetFeatureColumnSizes(int64* const num_dense_float_features,
int64* const num_sparse_float_features,
int64* const num_sparse_int_features) const {
- QCHECK_NE(num_dense_float_features, nullptr);
- QCHECK_NE(num_sparse_float_features, nullptr);
- QCHECK_NE(num_sparse_int_features, nullptr);
+ QCHECK_NE(num_dense_float_features, (int64*) nullptr);
+ QCHECK_NE(num_sparse_float_features, (int64*) nullptr);
+ QCHECK_NE(num_sparse_int_features, (int64*) nullptr);
*num_dense_float_features = dense_float_feature_columns_.size();
*num_sparse_float_features = sparse_float_feature_columns_.size();
*num_sparse_int_features = sparse_int_feature_columns_.size();
diff --git a/tensorflow/contrib/cmake/README.md b/tensorflow/contrib/cmake/README.md
index 8f85a75..fe83bb3 100644
--- a/tensorflow/contrib/cmake/README.md
+++ b/tensorflow/contrib/cmake/README.md
@@ -26,7 +26,7 @@
example C++ binary, and a PIP package containing the runtime and Python
bindings.
-### Pre-requisites
+### Prerequisites
* CMake version 3.5 or later.
@@ -34,14 +34,16 @@
* [SWIG](http://www.swig.org/download.html)
-* Additional pre-requisites for Microsoft Windows:
+* Additional prerequisites for Microsoft Windows:
- Visual Studio 2015
- Python 3.5
- - NumPy 1.11.0 or later
-* Additional pre-requisites for Linux:
+* Additional prerequisites for Linux:
- Python 2.7 or later
- [Docker](https://www.docker.com/) (for automated testing)
+
+* Python dependencies:
+ - wheel
- NumPy 1.11.0 or later
### Known-good configurations
@@ -102,7 +104,7 @@
Step-by-step Windows build
==========================
-1. Install the pre-requisites detailed above, and set up your environment.
+1. Install the prerequisites detailed above, and set up your environment.
* The following commands assume that you are using the Windows Command
Prompt (`cmd.exe`). You will need to set up your environment to use the
diff --git a/tensorflow/contrib/cmake/external/grpc.cmake b/tensorflow/contrib/cmake/external/grpc.cmake
index a9f43a3..17f6599 100644
--- a/tensorflow/contrib/cmake/external/grpc.cmake
+++ b/tensorflow/contrib/cmake/external/grpc.cmake
@@ -35,6 +35,7 @@
set(grpc_STATIC_LIBRARIES
${CMAKE_CURRENT_BINARY_DIR}/grpc/src/grpc/libgrpc++_unsecure.a
${CMAKE_CURRENT_BINARY_DIR}/grpc/src/grpc/libgrpc_unsecure.a
+ ${CMAKE_CURRENT_BINARY_DIR}/grpc/src/grpc/third_party/cares/cares/lib/libcares.a
${CMAKE_CURRENT_BINARY_DIR}/grpc/src/grpc/libgpr.a)
endif()
diff --git a/tensorflow/contrib/cmake/external/protobuf.cmake b/tensorflow/contrib/cmake/external/protobuf.cmake
index aba8a52..ab464bc 100644
--- a/tensorflow/contrib/cmake/external/protobuf.cmake
+++ b/tensorflow/contrib/cmake/external/protobuf.cmake
@@ -16,7 +16,7 @@
set(PROTOBUF_INCLUDE_DIRS ${CMAKE_CURRENT_BINARY_DIR}/protobuf/src/protobuf/src)
set(PROTOBUF_URL https://github.com/google/protobuf.git)
-set(PROTOBUF_TAG 396336eb961b75f03b25824fe86cf6490fb75e3a)
+set(PROTOBUF_TAG b04e5cba356212e4e8c66c61bbe0c3a20537c5b9)
if(WIN32)
if(${CMAKE_GENERATOR} MATCHES "Visual Studio.*")
diff --git a/tensorflow/contrib/cmake/tf_tests.cmake b/tensorflow/contrib/cmake/tf_tests.cmake
index 9f96a4b..b3e5b30 100644
--- a/tensorflow/contrib/cmake/tf_tests.cmake
+++ b/tensorflow/contrib/cmake/tf_tests.cmake
@@ -476,6 +476,10 @@
"${tensorflow_source_dir}/tensorflow/core/profiler/internal/advisor/*_test.cc"
)
+ list(REMOVE_ITEM tf_test_src_simple
+ ${tf_core_profiler_test_srcs}
+ )
+
set(tf_test_lib tf_test_lib)
add_library(${tf_test_lib} STATIC ${tf_src_testlib})
diff --git a/tensorflow/contrib/data/__init__.py b/tensorflow/contrib/data/__init__.py
index f09d156..9212b69 100644
--- a/tensorflow/contrib/data/__init__.py
+++ b/tensorflow/contrib/data/__init__.py
@@ -40,6 +40,7 @@
@@rejection_resample
@@scan
@@shuffle_and_repeat
+@@sliding_window_batch
@@sloppy_interleave
@@unbatch
@@ -72,6 +73,9 @@
from tensorflow.contrib.data.python.ops.resampling import rejection_resample
from tensorflow.contrib.data.python.ops.scan_ops import scan
from tensorflow.contrib.data.python.ops.shuffle_ops import shuffle_and_repeat
+from tensorflow.contrib.data.python.ops.sliding import sliding_window_batch
+from tensorflow.python.data.ops.iterator_ops import Iterator
+from tensorflow.python.ops.parsing_ops import parse_single_example_v2 as parse_single_example
# pylint: enable=unused-import
from tensorflow.python.util.all_util import remove_undocumented
diff --git a/tensorflow/contrib/data/python/kernel_tests/BUILD b/tensorflow/contrib/data/python/kernel_tests/BUILD
index 22418b3..2c4d4ad 100644
--- a/tensorflow/contrib/data/python/kernel_tests/BUILD
+++ b/tensorflow/contrib/data/python/kernel_tests/BUILD
@@ -498,6 +498,23 @@
],
)
+tf_py_test(
+ name = "slide_dataset_op_test",
+ size = "small",
+ srcs = ["slide_dataset_op_test.py"],
+ additional_deps = [
+ "//tensorflow/contrib/data/python/ops:dataset_ops",
+ "//tensorflow/contrib/data/python/ops:transformation_ops",
+ "//tensorflow/python:array_ops",
+ "//tensorflow/python:client_testlib",
+ "//tensorflow/python:dtypes",
+ "//tensorflow/python:errors",
+ "//tensorflow/python:math_ops",
+ "//tensorflow/python:sparse_tensor",
+ "//third_party/py/numpy",
+ ],
+)
+
filegroup(
name = "all_files",
srcs = glob(
diff --git a/tensorflow/contrib/data/python/kernel_tests/slide_dataset_op_test.py b/tensorflow/contrib/data/python/kernel_tests/slide_dataset_op_test.py
new file mode 100644
index 0000000..33c48e2
--- /dev/null
+++ b/tensorflow/contrib/data/python/kernel_tests/slide_dataset_op_test.py
@@ -0,0 +1,242 @@
+# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Tests for the experimental input pipeline ops."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+
+from tensorflow.contrib.data.python.ops import sliding
+from tensorflow.python.data.ops import dataset_ops
+from tensorflow.python.framework import dtypes
+from tensorflow.python.framework import errors
+from tensorflow.python.framework import sparse_tensor
+from tensorflow.python.ops import array_ops
+from tensorflow.python.ops import math_ops
+from tensorflow.python.platform import test
+
+
+class SlideDatasetTest(test.TestCase):
+
+ def testSlideDataset(self):
+ """Test an dataset that maps a TF function across its input elements."""
+ components = (np.arange(7),
+ np.array([[1, 2, 3]]) * np.arange(7)[:, np.newaxis],
+ np.array(37.0) * np.arange(7))
+
+ count = array_ops.placeholder(dtypes.int64, shape=[])
+ window_size = array_ops.placeholder(dtypes.int64, shape=[])
+ stride = array_ops.placeholder(dtypes.int64, shape=[])
+
+ def _map_fn(x, y, z):
+ return math_ops.square(x), math_ops.square(y), math_ops.square(z)
+
+ # The pipeline is TensorSliceDataset -> MapDataset(square_3) ->
+ # RepeatDataset(count) -> _SlideDataset(window_size, stride).
+ iterator = (dataset_ops.Dataset.from_tensor_slices(components)
+ .map(_map_fn)
+ .repeat(count)
+ .apply(sliding.sliding_window_batch(window_size, stride))
+ .make_initializable_iterator())
+ init_op = iterator.initializer
+ get_next = iterator.get_next()
+
+ self.assertEqual([[None] + list(c.shape[1:]) for c in components],
+ [t.shape.as_list() for t in get_next])
+
+ with self.test_session() as sess:
+ # Slide over a finite input, where the window_size divides the
+ # total number of elements.
+ sess.run(init_op, feed_dict={count: 20, window_size: 14, stride: 7})
+ # Same formula with convolution layer.
+ num_batches = (20 * 7 - 14) // 7 + 1
+ for i in range(num_batches):
+ result = sess.run(get_next)
+ for component, result_component in zip(components, result):
+ for j in range(14):
+ self.assertAllEqual(component[(i*7 + j) % 7]**2,
+ result_component[j])
+ with self.assertRaises(errors.OutOfRangeError):
+ sess.run(get_next)
+
+ # Slide over a finite input, where the window_size does not
+ # divide the total number of elements.
+ sess.run(init_op, feed_dict={count: 20, window_size: 17, stride: 9})
+
+ num_batches = (20 * 7 - 17) // 9 + 1
+ for i in range(num_batches):
+ result = sess.run(get_next)
+ for component, result_component in zip(components, result):
+ for j in range(17):
+ self.assertAllEqual(component[(i*9 + j) % 7]**2,
+ result_component[j])
+ with self.assertRaises(errors.OutOfRangeError):
+ sess.run(get_next)
+
+ # Slide over a finite input, which is less than window_size,
+ # should fail straight away.
+ sess.run(init_op, feed_dict={count: 1, window_size: 10, stride: 4})
+ with self.assertRaises(errors.OutOfRangeError):
+ sess.run(get_next)
+
+ sess.run(init_op, feed_dict={count: 1, window_size: 10, stride: 8})
+ with self.assertRaises(errors.OutOfRangeError):
+ sess.run(get_next)
+
+ # Slide over an empty input should fail straight away.
+ sess.run(init_op, feed_dict={count: 0, window_size: 8, stride: 4})
+ with self.assertRaises(errors.OutOfRangeError):
+ sess.run(get_next)
+
+ # Empty window_size should be an initialization time error.
+ with self.assertRaises(errors.InvalidArgumentError):
+ sess.run(init_op, feed_dict={count: 14, window_size: 0, stride: 0})
+
+ # Invalid stride should be an initialization time error.
+ with self.assertRaises(errors.InvalidArgumentError):
+ sess.run(init_op, feed_dict={count: 14, window_size: 3, stride: 0})
+ with self.assertRaises(errors.InvalidArgumentError):
+ sess.run(init_op, feed_dict={count: 14, window_size: 3, stride: 3})
+ with self.assertRaises(errors.InvalidArgumentError):
+ sess.run(init_op, feed_dict={count: 14, window_size: 3, stride: 5})
+
+ def assertSparseValuesEqual(self, a, b):
+ self.assertAllEqual(a.indices, b.indices)
+ self.assertAllEqual(a.values, b.values)
+ self.assertAllEqual(a.dense_shape, b.dense_shape)
+
+ def testSlideSparse(self):
+
+ def _sparse(i):
+ return sparse_tensor.SparseTensorValue(
+ indices=[[0]], values=(i * [1]), dense_shape=[1])
+
+ iterator = dataset_ops.Dataset.range(10).map(_sparse).apply(
+ sliding.sliding_window_batch(5, 3)).make_initializable_iterator()
+ init_op = iterator.initializer
+ get_next = iterator.get_next()
+
+ with self.test_session() as sess:
+ sess.run(init_op)
+ num_batches = (10 - 5) // 3 + 1
+ for i in range(num_batches):
+ actual = sess.run(get_next)
+ expected = sparse_tensor.SparseTensorValue(
+ indices=[[0, 0], [1, 0], [2, 0], [3, 0], [4, 0]],
+ values=[i * 3, i * 3 + 1, i * 3 + 2, i * 3 + 3, i * 3 + 4],
+ dense_shape=[5, 1])
+ self.assertTrue(sparse_tensor.is_sparse(actual))
+ self.assertSparseValuesEqual(actual, expected)
+ with self.assertRaises(errors.OutOfRangeError):
+ sess.run(get_next)
+
+ def testSlideSparseWithDifferentDenseShapes(self):
+
+ def _sparse(i):
+ return sparse_tensor.SparseTensorValue(
+ indices=array_ops.expand_dims(
+ math_ops.range(i, dtype=dtypes.int64), 1),
+ values=array_ops.fill([math_ops.to_int32(i)], i),
+ dense_shape=[i])
+
+ iterator = dataset_ops.Dataset.range(10).map(_sparse).apply(
+ sliding.sliding_window_batch(5, 3)).make_initializable_iterator()
+ init_op = iterator.initializer
+ get_next = iterator.get_next()
+
+ with self.test_session() as sess:
+ sess.run(init_op)
+ num_batches = (10 - 5) // 3 + 1
+ for i in range(num_batches):
+ actual = sess.run(get_next)
+ expected_indices = []
+ expected_values = []
+ for j in range(5):
+ for k in range(i * 3 + j):
+ expected_indices.append([j, k])
+ expected_values.append(i * 3 + j)
+ expected = sparse_tensor.SparseTensorValue(
+ indices=expected_indices,
+ values=expected_values,
+ dense_shape=[5, i * 3 + 5 - 1])
+ self.assertTrue(sparse_tensor.is_sparse(actual))
+ self.assertSparseValuesEqual(actual, expected)
+ with self.assertRaises(errors.OutOfRangeError):
+ sess.run(get_next)
+
+ def testNestedSlideSparse(self):
+
+ def _sparse(i):
+ return sparse_tensor.SparseTensorValue(
+ indices=[[0]], values=(i * [1]), dense_shape=[1])
+
+ iterator = (dataset_ops.Dataset.range(10)
+ .map(_sparse)
+ .apply(sliding.sliding_window_batch(4, 2))
+ .apply(sliding.sliding_window_batch(3, 1))
+ .make_initializable_iterator())
+ init_op = iterator.initializer
+ get_next = iterator.get_next()
+
+ with self.test_session() as sess:
+ sess.run(init_op)
+ # Slide: 1st batch.
+ actual = sess.run(get_next)
+ expected = sparse_tensor.SparseTensorValue(
+ indices=[[0, 0, 0], [0, 1, 0], [0, 2, 0], [0, 3, 0],
+ [1, 0, 0], [1, 1, 0], [1, 2, 0], [1, 3, 0],
+ [2, 0, 0], [2, 1, 0], [2, 2, 0], [2, 3, 0]],
+ values=[0, 1, 2, 3, 2, 3, 4, 5, 4, 5, 6, 7],
+ dense_shape=[3, 4, 1])
+ self.assertTrue(sparse_tensor.is_sparse(actual))
+ self.assertSparseValuesEqual(actual, expected)
+ # Slide: 2nd batch.
+ actual = sess.run(get_next)
+ expected = sparse_tensor.SparseTensorValue(
+ indices=[[0, 0, 0], [0, 1, 0], [0, 2, 0], [0, 3, 0],
+ [1, 0, 0], [1, 1, 0], [1, 2, 0], [1, 3, 0],
+ [2, 0, 0], [2, 1, 0], [2, 2, 0], [2, 3, 0]],
+ values=[2, 3, 4, 5, 4, 5, 6, 7, 6, 7, 8, 9],
+ dense_shape=[3, 4, 1])
+ self.assertTrue(sparse_tensor.is_sparse(actual))
+ self.assertSparseValuesEqual(actual, expected)
+ with self.assertRaises(errors.OutOfRangeError):
+ sess.run(get_next)
+
+ def testSlideShapeError(self):
+
+ def generator():
+ yield [1.0, 2.0, 3.0]
+ yield [4.0, 5.0, 6.0]
+ yield [7.0, 8.0, 9.0, 10.0]
+
+ iterator = (dataset_ops.Dataset.from_generator(generator, dtypes.float32,
+ output_shapes=[None])
+ .apply(sliding.sliding_window_batch(3, 1))
+ .make_initializable_iterator())
+ next_element = iterator.get_next()
+
+ with self.test_session() as sess:
+ sess.run(iterator.initializer)
+ with self.assertRaisesRegexp(
+ errors.InvalidArgumentError,
+ r"Cannot batch tensors with different shapes in component 0. "
+ r"First element had shape \[3\] and element 2 had shape \[4\]."):
+ sess.run(next_element)
+
+
+if __name__ == "__main__":
+ test.main()
diff --git a/tensorflow/contrib/data/python/ops/BUILD b/tensorflow/contrib/data/python/ops/BUILD
index f03430c..c3331e9 100644
--- a/tensorflow/contrib/data/python/ops/BUILD
+++ b/tensorflow/contrib/data/python/ops/BUILD
@@ -106,6 +106,7 @@
"interleave_ops.py",
"resampling.py",
"scan_ops.py",
+ "sliding.py",
"stats_ops.py",
"threadpool.py",
"unique.py",
diff --git a/tensorflow/contrib/data/python/ops/sliding.py b/tensorflow/contrib/data/python/ops/sliding.py
new file mode 100644
index 0000000..19cc3cb
--- /dev/null
+++ b/tensorflow/contrib/data/python/ops/sliding.py
@@ -0,0 +1,102 @@
+# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Sliding dataset transformations."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from tensorflow.python.data.ops import dataset_ops
+from tensorflow.python.data.util import nest
+from tensorflow.python.data.util import sparse
+from tensorflow.python.framework import dtypes
+from tensorflow.python.framework import ops
+from tensorflow.python.framework import tensor_shape
+from tensorflow.python.ops import gen_dataset_ops
+
+
+class _SlideDataset(dataset_ops.Dataset):
+ """A `Dataset` that passes a sliding window over its input."""
+
+ def __init__(self, input_dataset, window_size, stride=1):
+ """See `sliding_window_batch` for details."""
+ super(_SlideDataset, self).__init__()
+ self._input_dataset = input_dataset
+ self._window_size = ops.convert_to_tensor(
+ window_size, dtype=dtypes.int64, name="window_size")
+ self._stride = ops.convert_to_tensor(
+ stride, dtype=dtypes.int64, name="stride")
+
+ def _as_variant_tensor(self):
+ return gen_dataset_ops.slide_dataset(
+ self._input_dataset._as_variant_tensor(), # pylint: disable=protected-access
+ window_size=self._window_size,
+ stride=self._stride,
+ output_shapes=nest.flatten(
+ sparse.as_dense_shapes(self.output_shapes, self.output_classes)),
+ output_types=nest.flatten(
+ sparse.as_dense_types(self.output_types, self.output_classes)))
+
+ @property
+ def output_classes(self):
+ return self._input_dataset.output_classes
+
+ @property
+ def output_shapes(self):
+ input_shapes = self._input_dataset.output_shapes
+ return nest.pack_sequence_as(input_shapes, [
+ tensor_shape.vector(None).concatenate(s)
+ for s in nest.flatten(self._input_dataset.output_shapes)
+ ])
+
+ @property
+ def output_types(self):
+ return self._input_dataset.output_types
+
+
+def sliding_window_batch(window_size, stride=1):
+ """A sliding window with size of `window_size` and step of `stride`.
+
+ This transformation passes a sliding window over this dataset. The
+ window size is `window_size` and step size is `stride`. If the left
+ elements cannot fill up the sliding window, this transformation will
+ drop the final smaller element. For example:
+
+ ```python
+ # NOTE: The following examples use `{ ... }` to represent the
+ # contents of a dataset.
+ a = { [1], [2], [3], [4], [5], [6] }
+
+ a.apply(tf.contrib.data.sliding_window_batch(window_size=3, stride=2)) ==
+ {
+ [[1], [2], [3]],
+ [[3], [4], [5]],
+ }
+ ```
+
+ Args:
+ window_size: A `tf.int64` scalar `tf.Tensor`, representing the number of
+ elements in the sliding window.
+ stride: (Optional.) A `tf.int64` scalar `tf.Tensor`, representing the
+ steps moving the sliding window forward for one iteration. The default
+ is `1`. It must be in `[1, window_size)`.
+
+ Returns:
+ A `Dataset` transformation function, which can be passed to
+ @{tf.data.Dataset.apply}.
+ """
+ def _apply_fn(dataset):
+ return _SlideDataset(dataset, window_size, stride)
+
+ return _apply_fn
diff --git a/tensorflow/contrib/distributions/BUILD b/tensorflow/contrib/distributions/BUILD
index 6bd3f5f..1bd73ee 100644
--- a/tensorflow/contrib/distributions/BUILD
+++ b/tensorflow/contrib/distributions/BUILD
@@ -454,6 +454,7 @@
"//tensorflow/python:framework_test_lib",
"//tensorflow/python:platform_test",
],
+ tags = ["no_windows"], # TODO: needs investigation on Windows
)
cuda_py_test(
@@ -1143,6 +1144,7 @@
"//tensorflow/python:math_ops",
"//tensorflow/python:platform_test",
],
+ tags = ["no_windows"], # TODO: needs investigation on Windows
)
cuda_py_test(
diff --git a/tensorflow/contrib/eager/python/BUILD b/tensorflow/contrib/eager/python/BUILD
index 2f9bc68..384ef7f 100644
--- a/tensorflow/contrib/eager/python/BUILD
+++ b/tensorflow/contrib/eager/python/BUILD
@@ -266,7 +266,10 @@
"//tensorflow/python/eager:test",
"//tensorflow/python/keras",
],
- tags = ["notsan"],
+ tags = [
+ "no_windows", # TODO: needs investigation on Windows
+ "notsan",
+ ],
)
filegroup(
diff --git a/tensorflow/contrib/eager/python/examples/linear_regression/BUILD b/tensorflow/contrib/eager/python/examples/linear_regression/BUILD
index f86331a..2f6cfdf 100644
--- a/tensorflow/contrib/eager/python/examples/linear_regression/BUILD
+++ b/tensorflow/contrib/eager/python/examples/linear_regression/BUILD
@@ -22,6 +22,7 @@
":linear_regression",
"//tensorflow:tensorflow_py",
],
+ tags = ["no_windows"], # TODO: needs investigation on Windows
)
cuda_py_test(
diff --git a/tensorflow/contrib/factorization/BUILD b/tensorflow/contrib/factorization/BUILD
index 90f10f1..ad8568a 100644
--- a/tensorflow/contrib/factorization/BUILD
+++ b/tensorflow/contrib/factorization/BUILD
@@ -224,7 +224,10 @@
srcs = ["python/ops/kmeans_test.py"],
shard_count = 4,
srcs_version = "PY2AND3",
- tags = ["notsan"], # b/67512932
+ tags = [
+ "nomac", # b/73741358
+ "notsan", # b/67512932
+ ],
deps = [
":factorization_py",
":factorization_py_CYCLIC_DEPENDENCIES_THAT_NEED_TO_GO",
diff --git a/tensorflow/contrib/ffmpeg/default/ffmpeg_lib.cc b/tensorflow/contrib/ffmpeg/default/ffmpeg_lib.cc
index e61221a..3534140 100644
--- a/tensorflow/contrib/ffmpeg/default/ffmpeg_lib.cc
+++ b/tensorflow/contrib/ffmpeg/default/ffmpeg_lib.cc
@@ -256,6 +256,9 @@
if (p != std::string::npos) {
string rgb24 = line.substr(p + 9, line.find(" ", p + 9));
rgb24 = rgb24.substr(0, rgb24.find(","));
+ // Strip anything after " ", in case the format is
+ // `640x360 [SAR 1:1 DAR 16:9]`
+ rgb24 = rgb24.substr(0, rgb24.find(" "));
string rgb24_width = rgb24.substr(0, rgb24.find("x"));
string rgb24_height = rgb24.substr(rgb24_width.length() + 1);
if (strings::safe_strtou32(rgb24_width, &width_value) &&
@@ -270,8 +273,10 @@
// We only look for the first stream mapping to have the number of the
// frames.
// Once processed we will not further process stream mapping section.
- if (line.find("frame= ") == 0) {
- string number = line.substr(8, line.find(" ", 8));
+ if (line.find("frame=") == 0) {
+ // The format might be `frame= 166 ` or `frame=12488 `
+ string number = line.substr(6);
+ number = number.substr(number.find_first_not_of(" "));
number = number.substr(0, number.find(" "));
if (strings::safe_strtou32(number, &frames_value)) {
in_mapping = false;
diff --git a/tensorflow/contrib/gan/BUILD b/tensorflow/contrib/gan/BUILD
index 0eb0e3c..ff6f3b7 100644
--- a/tensorflow/contrib/gan/BUILD
+++ b/tensorflow/contrib/gan/BUILD
@@ -354,6 +354,7 @@
name = "classifier_metrics_test",
srcs = ["python/eval/python/classifier_metrics_test.py"],
srcs_version = "PY2AND3",
+ tags = ["no_windows"], # TODO: needs investigation on Windows
deps = [
":classifier_metrics",
"//tensorflow/core:protos_all_py",
diff --git a/tensorflow/contrib/kafka/BUILD b/tensorflow/contrib/kafka/BUILD
index efb4034..1c39748 100644
--- a/tensorflow/contrib/kafka/BUILD
+++ b/tensorflow/contrib/kafka/BUILD
@@ -1,66 +1,93 @@
-package(
- default_visibility = ["//visibility:private"],
-)
+package(default_visibility = ["//tensorflow:internal"])
licenses(["notice"]) # Apache 2.0
exports_files(["LICENSE"])
-load("//tensorflow:tensorflow.bzl", "tf_gen_op_libs")
-load("//tensorflow:tensorflow.bzl", "tf_gen_op_wrapper_py")
-load("//tensorflow:tensorflow.bzl", "tf_kernel_library")
-load("//tensorflow:tensorflow.bzl", "tf_py_test")
-
-tf_kernel_library(
- name = "kafka_kernels",
- srcs = ["kernels/kafka_dataset_ops.cc"],
- visibility = ["//visibility:public"],
- deps = [
- "//tensorflow/core:framework",
- "//tensorflow/core:lib",
- "//tensorflow/core:lib_internal",
- "//tensorflow/core/kernels:bounds_check_lib",
- "//tensorflow/core/kernels:dataset",
- "//third_party/eigen3",
- "@kafka",
- ],
-)
-
-tf_gen_op_libs(
- op_lib_names = ["kafka_ops"],
- deps = [
- "//tensorflow/core:lib",
- ],
-)
-
-tf_gen_op_wrapper_py(
- name = "gen_kafka_ops",
- out = "python/ops/gen_kafka_ops.py",
- require_shape_functions = True,
- deps = [":kafka_ops_op_lib"],
+load(
+ "//tensorflow:tensorflow.bzl",
+ "tf_gen_op_wrapper_py",
+ "tf_kernel_library",
+ "tf_custom_op_library",
+ "tf_custom_op_py_library",
+ "tf_gen_op_libs",
+ "tf_py_test",
)
py_library(
name = "kafka",
+ srcs = ["__init__.py"],
+ srcs_version = "PY2AND3",
+ deps = [
+ ":dataset_ops",
+ ],
+)
+
+tf_custom_op_library(
+ name = "_dataset_ops.so",
+ srcs = ["ops/dataset_ops.cc"],
+ deps = [":dataset_kernels"],
+)
+
+tf_gen_op_libs(
+ op_lib_names = ["dataset_ops"],
+)
+
+cc_library(
+ name = "dataset_kernels",
+ srcs = ["kernels/kafka_dataset_ops.cc"],
+ deps = [
+ "//tensorflow/core:framework_headers_lib",
+ "//third_party/eigen3",
+ "@kafka",
+ "@protobuf_archive//:protobuf_headers",
+ ],
+ alwayslink = 1,
+)
+
+py_library(
+ name = "dataset_ops",
srcs = [
- "__init__.py",
"python/ops/kafka_dataset_ops.py",
],
srcs_version = "PY2AND3",
- visibility = ["//visibility:public"],
deps = [
- ":gen_kafka_ops",
- "//tensorflow/contrib/util:util_py",
- "//tensorflow/python:array_ops",
- "//tensorflow/python:control_flow_ops",
- "//tensorflow/python:framework",
- "//tensorflow/python:framework_for_generated_wrappers",
- "//tensorflow/python:platform",
- "//tensorflow/python:state_ops",
- "//tensorflow/python:training",
+ ":kafka_op_loader",
+ "//tensorflow/python:dataset_ops_gen",
+ "//tensorflow/python:util",
"//tensorflow/python/data/ops:dataset_ops",
- "//tensorflow/python/data/ops:iterator_ops",
- "//tensorflow/python/data/ops:readers",
+ "//tensorflow/python/data/util:nest",
+ ],
+)
+
+tf_gen_op_wrapper_py(
+ name = "gen_dataset_ops",
+ out = "python/ops/gen_dataset_ops.py",
+ deps = ["//tensorflow/contrib/kafka:dataset_ops_op_lib"],
+)
+
+tf_kernel_library(
+ name = "dataset_ops_kernels",
+ deps = [
+ ":dataset_kernels",
+ "//tensorflow/core:framework",
+ ],
+ alwayslink = 1,
+)
+
+tf_custom_op_py_library(
+ name = "kafka_op_loader",
+ srcs = ["python/ops/kafka_op_loader.py"],
+ dso = ["//tensorflow/contrib/kafka:_dataset_ops.so"],
+ kernels = [
+ ":dataset_ops_kernels",
+ "//tensorflow/contrib/kafka:dataset_ops_op_lib",
+ ],
+ srcs_version = "PY2AND3",
+ deps = [
+ ":gen_dataset_ops",
+ "//tensorflow/contrib/util:util_py",
+ "//tensorflow/python:platform",
],
)
@@ -88,6 +115,7 @@
],
tags = [
"manual",
+ "no_windows",
"notap",
],
)
@@ -95,7 +123,9 @@
filegroup(
name = "all_files",
srcs = glob(
- ["**/*"],
+ include = [
+ "**/*",
+ ],
exclude = [
"**/METADATA",
"**/OWNERS",
diff --git a/tensorflow/contrib/kafka/kernels/kafka_dataset_ops.cc b/tensorflow/contrib/kafka/kernels/kafka_dataset_ops.cc
index 88ef5f3..a4cd4a2 100644
--- a/tensorflow/contrib/kafka/kernels/kafka_dataset_ops.cc
+++ b/tensorflow/contrib/kafka/kernels/kafka_dataset_ops.cc
@@ -13,9 +13,7 @@
limitations under the License.
==============================================================================*/
-#include "tensorflow/core/kernels/dataset.h"
-
-#include "tensorflow/core/framework/tensor.h"
+#include "tensorflow/core/framework/dataset.h"
#include "src-cpp/rdkafkacpp.h"
diff --git a/tensorflow/contrib/kafka/ops/dataset_ops.cc b/tensorflow/contrib/kafka/ops/dataset_ops.cc
new file mode 100644
index 0000000..8cdf161
--- /dev/null
+++ b/tensorflow/contrib/kafka/ops/dataset_ops.cc
@@ -0,0 +1,44 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/core/framework/common_shape_fns.h"
+#include "tensorflow/core/framework/op.h"
+#include "tensorflow/core/framework/shape_inference.h"
+
+namespace tensorflow {
+
+REGISTER_OP("KafkaDataset")
+ .Input("topics: string")
+ .Input("servers: string")
+ .Input("group: string")
+ .Input("eof: bool")
+ .Input("timeout: int64")
+ .Output("handle: variant")
+ .SetIsStateful()
+ .SetShapeFn(shape_inference::ScalarShape)
+ .Doc(R"doc(
+Creates a dataset that emits the messages of one or more Kafka topics.
+
+topics: A `tf.string` tensor containing one or more subscriptions,
+ in the format of [topic:partition:offset:length],
+ by default length is -1 for unlimited.
+servers: A list of bootstrap servers.
+group: The consumer group id.
+eof: If True, the kafka reader will stop on EOF.
+timeout: The timeout value for the Kafka Consumer to wait
+ (in millisecond).
+)doc");
+
+} // namespace tensorflow
diff --git a/tensorflow/contrib/kafka/python/ops/kafka_dataset_ops.py b/tensorflow/contrib/kafka/python/ops/kafka_dataset_ops.py
index 8e51d27..a162461 100644
--- a/tensorflow/contrib/kafka/python/ops/kafka_dataset_ops.py
+++ b/tensorflow/contrib/kafka/python/ops/kafka_dataset_ops.py
@@ -17,8 +17,9 @@
from __future__ import division
from __future__ import print_function
-from tensorflow.contrib.kafka.python.ops import gen_kafka_ops
-from tensorflow.python.data.ops.readers import Dataset
+from tensorflow.contrib.kafka.python.ops import kafka_op_loader # pylint: disable=unused-import
+from tensorflow.contrib.kafka.python.ops import gen_dataset_ops
+from tensorflow.python.data.ops.dataset_ops import Dataset
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import ops
from tensorflow.python.framework import tensor_shape
@@ -58,8 +59,8 @@
timeout, dtype=dtypes.int64, name="timeout")
def _as_variant_tensor(self):
- return gen_kafka_ops.kafka_dataset(self._topics, self._servers, self._group,
- self._eof, self._timeout)
+ return gen_dataset_ops.kafka_dataset(self._topics, self._servers,
+ self._group, self._eof, self._timeout)
@property
def output_classes(self):
diff --git a/tensorflow/contrib/kafka/python/ops/kafka_op_loader.py b/tensorflow/contrib/kafka/python/ops/kafka_op_loader.py
new file mode 100644
index 0000000..ec2fdea
--- /dev/null
+++ b/tensorflow/contrib/kafka/python/ops/kafka_op_loader.py
@@ -0,0 +1,24 @@
+# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Python helper for loading kafka ops and kernels."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from tensorflow.contrib.util import loader
+from tensorflow.python.platform import resource_loader
+
+_dataset_ops = loader.load_op_library(
+ resource_loader.get_path_to_datafile("../../_dataset_ops.so"))
diff --git a/tensorflow/contrib/kfac/python/kernel_tests/BUILD b/tensorflow/contrib/kfac/python/kernel_tests/BUILD
index 146ae8b..d1c4494 100644
--- a/tensorflow/contrib/kfac/python/kernel_tests/BUILD
+++ b/tensorflow/contrib/kfac/python/kernel_tests/BUILD
@@ -114,6 +114,7 @@
name = "utils_test",
srcs = ["utils_test.py"],
srcs_version = "PY2AND3",
+ tags = ["no_windows"], # TODO: needs investigation on Windows
deps = [
"//tensorflow/contrib/kfac/python/ops:utils",
"//tensorflow/contrib/tpu",
diff --git a/tensorflow/contrib/labeled_tensor/BUILD b/tensorflow/contrib/labeled_tensor/BUILD
index 894e6f6..544065d 100644
--- a/tensorflow/contrib/labeled_tensor/BUILD
+++ b/tensorflow/contrib/labeled_tensor/BUILD
@@ -70,6 +70,7 @@
"python/ops/core_test.py",
],
srcs_version = "PY2AND3",
+ tags = ["no_windows"], # TODO: needs investigation on Windows
deps = [
":_typecheck",
":core",
diff --git a/tensorflow/contrib/layers/BUILD b/tensorflow/contrib/layers/BUILD
index 852d06e..cc7bbab 100644
--- a/tensorflow/contrib/layers/BUILD
+++ b/tensorflow/contrib/layers/BUILD
@@ -188,6 +188,7 @@
size = "small",
srcs = ["python/layers/normalization_test.py"],
srcs_version = "PY2AND3",
+ tags = ["no_windows"], # TODO: needs investigation on Windows
deps = [
":layers_py",
"//tensorflow/contrib/framework:framework_py",
@@ -353,6 +354,7 @@
size = "small",
srcs = ["python/ops/sparse_ops_test.py"],
srcs_version = "PY2AND3",
+ tags = ["no_windows"], # TODO: needs investigation on Windows
deps = [
":layers_py",
"//tensorflow/python:array_ops",
diff --git a/tensorflow/contrib/layers/python/layers/embedding_ops.py b/tensorflow/contrib/layers/python/layers/embedding_ops.py
index b62e305..ffa2085 100644
--- a/tensorflow/contrib/layers/python/layers/embedding_ops.py
+++ b/tensorflow/contrib/layers/python/layers/embedding_ops.py
@@ -470,7 +470,7 @@
ids = ops.convert_to_tensor(ids)
shape = array_ops.shape(ids)
ids_flat = array_ops.reshape(
- ids, math_ops.reduce_prod(shape, keep_dims=True))
+ ids, math_ops.reduce_prod(shape, keepdims=True))
unique_ids, idx = array_ops.unique(ids_flat)
unique_embeddings = embedding_ops.embedding_lookup(params, unique_ids)
embeds_flat = array_ops.gather(unique_embeddings, idx)
diff --git a/tensorflow/contrib/learn/BUILD b/tensorflow/contrib/learn/BUILD
index f837ca3..b05f5ee 100644
--- a/tensorflow/contrib/learn/BUILD
+++ b/tensorflow/contrib/learn/BUILD
@@ -5,6 +5,8 @@
exports_files(["LICENSE"])
+load("//tensorflow:tensorflow.bzl", "py_test")
+
package(default_visibility = [
"//engedu/ml/tf_from_scratch:__pkg__",
"//tensorflow:internal",
@@ -115,6 +117,7 @@
size = "small",
srcs = ["python/learn/learn_io/data_feeder_test.py"],
srcs_version = "PY2AND3",
+ tags = ["no_windows"], # TODO: needs investigation on Windows
deps = [
":learn",
"//tensorflow/python:client_testlib",
@@ -170,6 +173,7 @@
"//tensorflow/python:variables",
"//tensorflow/python/estimator",
],
+ tags = ["no_windows"], # TODO: needs investigation on Windows
)
py_test(
@@ -188,6 +192,7 @@
size = "small",
srcs = ["python/learn/graph_actions_test.py"],
srcs_version = "PY2AND3",
+ tags = ["no_windows"], # TODO: needs investigation on Windows
deps = [
":learn",
"//tensorflow/contrib/framework:framework_py",
@@ -426,7 +431,10 @@
size = "medium",
srcs = ["python/learn/estimators/kmeans_test.py"],
srcs_version = "PY2AND3",
- tags = ["noasan"],
+ tags = [
+ "noasan", # b/73741358
+ "nomac",
+ ],
deps = [
":learn",
"//tensorflow/python:array_ops",
@@ -585,6 +593,7 @@
size = "small",
srcs = ["python/learn/learn_io/io_test.py"],
srcs_version = "PY2AND3",
+ tags = ["no_windows"], # TODO: needs investigation on Windows
deps = [
":learn",
"//tensorflow/contrib/learn/python/learn/datasets",
@@ -814,6 +823,7 @@
size = "small",
srcs = ["python/learn/utils/saved_model_export_utils_test.py"],
srcs_version = "PY2AND3",
+ tags = ["no_windows"], # TODO: needs investigation on Windows
deps = [
":learn",
"//tensorflow/contrib/layers:layers_py",
diff --git a/tensorflow/contrib/learn/python/learn/ops/embeddings_ops.py b/tensorflow/contrib/learn/python/learn/ops/embeddings_ops.py
index b3b067b..8f9811c 100644
--- a/tensorflow/contrib/learn/python/learn/ops/embeddings_ops.py
+++ b/tensorflow/contrib/learn/python/learn/ops/embeddings_ops.py
@@ -61,7 +61,7 @@
ids = ops.convert_to_tensor(ids)
shape = array_ops_.shape(ids)
ids_flat = array_ops_.reshape(
- ids, math_ops.reduce_prod(shape, keep_dims=True))
+ ids, math_ops.reduce_prod(shape, keepdims=True))
embeds_flat = nn.embedding_lookup(params, ids_flat, name)
embed_shape = array_ops_.concat([shape, [-1]], 0)
embeds = array_ops_.reshape(embeds_flat, embed_shape)
diff --git a/tensorflow/contrib/lite/Makefile b/tensorflow/contrib/lite/Makefile
index 7f31629..b4504f2 100644
--- a/tensorflow/contrib/lite/Makefile
+++ b/tensorflow/contrib/lite/Makefile
@@ -27,10 +27,10 @@
GENDIR := $(MAKEFILE_DIR)/gen/obj/
# Settings for the host compiler.
-CXX := $(CC_PREFIX) gcc
+CXX := $(CC_PREFIX)gcc
CXXFLAGS := --std=c++11 -O3 -DNDEBUG
-CC := $(CC_PREFIX) gcc
-CFLAGS :=
+CC := $(CC_PREFIX)gcc
+CFLAGS := -O3 -DNDEBUG
LDOPTS :=
LDOPTS += -L/usr/local/lib
ARFLAGS := -r
@@ -57,10 +57,11 @@
# If we're on Linux, also link in the dl library.
ifeq ($(HOST_OS),LINUX)
- LIBS += -ldl -lpthread
+ LIBS += -ldl
endif
include $(MAKEFILE_DIR)/ios_makefile.inc
+include $(MAKEFILE_DIR)/rpi_makefile.inc
# This library is the main target for this makefile. It will contain a minimal
# runtime that can be linked in to other programs.
diff --git a/tensorflow/contrib/lite/arena_planner.h b/tensorflow/contrib/lite/arena_planner.h
index 58bc164..f84b3da 100644
--- a/tensorflow/contrib/lite/arena_planner.h
+++ b/tensorflow/contrib/lite/arena_planner.h
@@ -33,7 +33,7 @@
// each tensor needs to be allocated and deallocated, and preallocates all the
// necessary memory (the PlanAllocations phase). It then assigns portions of
// this memory buffer to each tensor (the ExecuteAllocations phase). Tensors may
-// share some of the bufer if a tensor B is to be allocated after another tensor
+// share some of the buffer if a tensor B is to be allocated after another tensor
// A has been deallocated.
//
// If dynamic tensors are used the planning steps can be repeated during model
diff --git a/tensorflow/contrib/lite/build_rpi_lib.sh b/tensorflow/contrib/lite/build_rpi_lib.sh
new file mode 100755
index 0000000..3824b16
--- /dev/null
+++ b/tensorflow/contrib/lite/build_rpi_lib.sh
@@ -0,0 +1,22 @@
+#!/bin/bash -x
+# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
+set -e
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+cd "$SCRIPT_DIR/../../.."
+
+CC_PREFIX=arm-linux-gnueabihf- make -j 3 -f tensorflow/contrib/lite/Makefile TARGET=RPI TARGET_ARCH=armv7
diff --git a/tensorflow/contrib/lite/builtin_ops.h b/tensorflow/contrib/lite/builtin_ops.h
index 2218ea8..ea3ae34 100644
--- a/tensorflow/contrib/lite/builtin_ops.h
+++ b/tensorflow/contrib/lite/builtin_ops.h
@@ -24,7 +24,7 @@
#endif // __cplusplus
// The enum for builtin operators.
-// Note: CUSTOM and DELEGATE are 2 special ops which are not real biultin
+// Note: CUSTOM and DELEGATE are 2 special ops which are not real builtin
// ops.
typedef enum {
kTfLiteBuiltinAdd = 0,
diff --git a/tensorflow/contrib/lite/error_reporter.h b/tensorflow/contrib/lite/error_reporter.h
index da193d2..3c5f805 100644
--- a/tensorflow/contrib/lite/error_reporter.h
+++ b/tensorflow/contrib/lite/error_reporter.h
@@ -30,7 +30,7 @@
// va_list args;
// foo.Report("test %d", args); // where args is va_list
//
-// Sublclass ErrorReporter to provide another reporting destination.
+// Subclass ErrorReporter to provide another reporting destination.
// For example, if you have a GUI program, you might redirect to a buffer
// that drives a GUI error log box.
class ErrorReporter {
diff --git a/tensorflow/contrib/lite/g3doc/rpi.md b/tensorflow/contrib/lite/g3doc/rpi.md
new file mode 100644
index 0000000..7a3a231
--- /dev/null
+++ b/tensorflow/contrib/lite/g3doc/rpi.md
@@ -0,0 +1,50 @@
+# TensorFlow Lite for Raspberry Pi
+
+## Cross compiling
+### Installing toolchian
+This has been tested on Ubuntu 16.04.3 64bit and Tensorflow devel docker image [tensorflow/tensorflow:nightly-devel](https://hub.docker.com/r/tensorflow/tensorflow/tags/).
+
+To cross compiling TensorFlow Lite. First you should install the toolchain and libs.
+```bash
+sudo apt-get update
+sudo apt-get install crossbuild-essential-armhf
+```
+> If you are using docker, you may not use `sudo`
+
+### Building
+Clone this Tensorflow repository, Run this script at the root of the repository to download all the dependencies:
+> The Tensorflow repository is in `/tensorflow` if you are using `tensorflow/tensorflow:nightly-devel` docker image, just try it.
+```bash
+./tensorflow/contrib/lite/download_dependencies.sh
+```
+Note than you only need to to this once.
+
+You should then be able to compile:
+```bash
+./tensorflow/contrib/lite/build_rpi_lib.sh
+```
+
+This should compile a static library in:
+`tensorflow/contrib/lite/gen/lib/rpi_armv7/libtensorflow-lite.a`.
+
+## Native compiling
+This has been tested on Raspberry Pi 3b, Raspbian GNU/Linux 9.1 (stretch), gcc version 6.3.0 20170516 (Raspbian 6.3.0-18+rpi1).
+
+Log in to you RPI, install the toolchain.
+```bash
+sudo apt-get instal build-essential
+```
+
+First, clone this TensorFlow repository. Run this at the root of the repository:
+```bash
+./tensorflow/contrib/lite/download_dependencies.sh
+```
+Note than you only need to to this once.
+
+You should then be able to compile:
+```bash
+./tensorflow/contrib/lite/build_rpi_lib.sh
+```
+
+This should compile a static library in:
+`tensorflow/contrib/lite/gen/lib/rpi_armv7/libtensorflow-lite.a`.
diff --git a/tensorflow/contrib/lite/interpreter.h b/tensorflow/contrib/lite/interpreter.h
index af14337..3749869 100644
--- a/tensorflow/contrib/lite/interpreter.h
+++ b/tensorflow/contrib/lite/interpreter.h
@@ -481,7 +481,7 @@
// During Invoke(), Interpreter will allocate input tensors first, which are
// known to be fixed size. Then it will allocate outputs from nodes as many
// as possible. When there is a node that produces dynamic sized tensor.
- // Intepreter will stop allocating tensors, set the value of next allocate
+ // Interpreter will stop allocating tensors, set the value of next allocate
// node id, and execute the node to generate the output tensor before continue
// to allocate successors. This process repeats until all nodes are executed.
// NOTE: this relies on the order of nodes that is in topological order.
diff --git a/tensorflow/contrib/lite/interpreter_test.cc b/tensorflow/contrib/lite/interpreter_test.cc
index 7a029c7..72d4ace 100644
--- a/tensorflow/contrib/lite/interpreter_test.cc
+++ b/tensorflow/contrib/lite/interpreter_test.cc
@@ -40,7 +40,7 @@
ASSERT_EQ(interpreter.Invoke(), kTfLiteOk);
}
-// Test size accesser functions.
+// Test size accessor functions.
TEST(BasicInterpreter, TestSizeFunctions) {
Interpreter interpreter;
int base_index;
diff --git a/tensorflow/contrib/lite/kernels/conv.cc b/tensorflow/contrib/lite/kernels/conv.cc
index b91ba1a..e0cd12f 100644
--- a/tensorflow/contrib/lite/kernels/conv.cc
+++ b/tensorflow/contrib/lite/kernels/conv.cc
@@ -64,7 +64,7 @@
TfLitePaddingValues padding;
// The scaling factor from input to output (aka the 'real multiplier') can
- // be represented as a fixed point multipler plus a left shift.
+ // be represented as a fixed point multiplier plus a left shift.
int32_t output_multiplier;
int output_shift;
// The range of the fused activation layer. For example for kNone and
diff --git a/tensorflow/contrib/lite/kernels/depthwise_conv.cc b/tensorflow/contrib/lite/kernels/depthwise_conv.cc
index 15dbfe0..cad9ce1 100644
--- a/tensorflow/contrib/lite/kernels/depthwise_conv.cc
+++ b/tensorflow/contrib/lite/kernels/depthwise_conv.cc
@@ -52,7 +52,7 @@
struct OpData {
TfLitePaddingValues padding;
// The scaling factor from input to output (aka the 'real multiplier') can
- // be represented as a fixed point multipler plus a left shift.
+ // be represented as a fixed point multiplier plus a left shift.
int32_t output_multiplier;
int output_shift;
// The range of the fused activation layer. For example for kNone and
diff --git a/tensorflow/contrib/lite/kernels/fully_connected.cc b/tensorflow/contrib/lite/kernels/fully_connected.cc
index a77fe94..888e679 100644
--- a/tensorflow/contrib/lite/kernels/fully_connected.cc
+++ b/tensorflow/contrib/lite/kernels/fully_connected.cc
@@ -48,7 +48,7 @@
struct OpData {
// The scaling factor from input to output (aka the 'real multiplier') can
- // be represented as a fixed point multipler plus a left shift.
+ // be represented as a fixed point multiplier plus a left shift.
int32_t output_multiplier;
int output_shift;
// The range of the fused activation layer. For example for kNone and
diff --git a/tensorflow/contrib/lite/kernels/kernel_util.h b/tensorflow/contrib/lite/kernels/kernel_util.h
index 28f53b9..21da1da 100644
--- a/tensorflow/contrib/lite/kernels/kernel_util.h
+++ b/tensorflow/contrib/lite/kernels/kernel_util.h
@@ -58,7 +58,7 @@
}
// Determines whether tensor is dynamic. Note that a tensor can be non-const and
-// not dynamic. This function specificially checks for a dynamic tensor.
+// not dynamic. This function specifically checks for a dynamic tensor.
inline bool IsDynamicTensor(TfLiteTensor* tensor) {
return tensor->allocation_type == kTfLiteDynamic;
}
diff --git a/tensorflow/contrib/lite/kernels/lsh_projection.cc b/tensorflow/contrib/lite/kernels/lsh_projection.cc
index 5f73b56..0ee3577 100644
--- a/tensorflow/contrib/lite/kernels/lsh_projection.cc
+++ b/tensorflow/contrib/lite/kernels/lsh_projection.cc
@@ -13,7 +13,7 @@
limitations under the License.
==============================================================================*/
-// LSH Projection projects an input to a bit vector via locality senstive
+// LSH Projection projects an input to a bit vector via locality sensitive
// hashing.
//
// Options:
diff --git a/tensorflow/contrib/lite/kernels/lstm.cc b/tensorflow/contrib/lite/kernels/lstm.cc
index b9255b2..8cf1165 100644
--- a/tensorflow/contrib/lite/kernels/lstm.cc
+++ b/tensorflow/contrib/lite/kernels/lstm.cc
@@ -213,9 +213,9 @@
// present.
// 2) If projection weight is present, then projection bias is optional.
// TODO(ghodrat): make sure this is correct.
- const bool projecton_tensors_consistent =
+ const bool projection_tensors_consistent =
((projection_weights != nullptr) || (projection_bias == nullptr));
- TF_LITE_ENSURE(context, projecton_tensors_consistent == true);
+ TF_LITE_ENSURE(context, projection_tensors_consistent == true);
return kTfLiteOk;
}
@@ -357,7 +357,7 @@
const int n_output = recurrent_to_output_weights->dims->data[1];
// Since we have already checked that weights are all there or none, we can
- // check the existense of only one to the get the condition.
+ // check the existence of only one to get the condition.
const bool use_cifg = (input_to_input_weights == nullptr);
const bool use_peephole = (cell_to_output_weights != nullptr);
diff --git a/tensorflow/contrib/lite/kernels/reshape.cc b/tensorflow/contrib/lite/kernels/reshape.cc
index f3e6ddc..438f70d 100644
--- a/tensorflow/contrib/lite/kernels/reshape.cc
+++ b/tensorflow/contrib/lite/kernels/reshape.cc
@@ -49,20 +49,20 @@
TfLiteIntArray* output_size = TfLiteIntArrayCreate(params->num_dimensions);
int num_output_elements = 1;
- int strech_dim = -1;
+ int stretch_dim = -1;
for (int i = 0; i < params->num_dimensions; ++i) {
int value = params->shape[i];
if (value == -1) {
- TF_LITE_ENSURE_EQ(context, strech_dim, -1);
- strech_dim = i;
+ TF_LITE_ENSURE_EQ(context, stretch_dim, -1);
+ stretch_dim = i;
} else {
num_output_elements *= value;
output_size->data[i] = value;
}
}
- if (strech_dim != -1) {
- output_size->data[strech_dim] = num_input_elements / num_output_elements;
- num_output_elements *= output_size->data[strech_dim];
+ if (stretch_dim != -1) {
+ output_size->data[stretch_dim] = num_input_elements / num_output_elements;
+ num_output_elements *= output_size->data[stretch_dim];
}
TF_LITE_ENSURE_EQ(context, num_input_elements, num_output_elements);
diff --git a/tensorflow/contrib/lite/kernels/reshape_test.cc b/tensorflow/contrib/lite/kernels/reshape_test.cc
index 0fbcf6e..aecbd03 100644
--- a/tensorflow/contrib/lite/kernels/reshape_test.cc
+++ b/tensorflow/contrib/lite/kernels/reshape_test.cc
@@ -60,7 +60,7 @@
TEST(ReshapeOpTest, TooManySpecialDimensions) {
EXPECT_DEATH(ReshapeOpModel({1, 2, 4, 1}, {-1, -1, 2, 4}),
- "strech_dim != -1");
+ "stretch_dim != -1");
}
TEST(ReshapeOpTest, SimpleTest) {
diff --git a/tensorflow/contrib/lite/kernels/test_util.cc b/tensorflow/contrib/lite/kernels/test_util.cc
index 373310b..0bb28b5 100644
--- a/tensorflow/contrib/lite/kernels/test_util.cc
+++ b/tensorflow/contrib/lite/kernels/test_util.cc
@@ -141,8 +141,8 @@
void SingleOpModel::SetCustomOp(
const string& name, const std::vector<uint8_t>& custom_option,
- const std::function<TfLiteRegistration*()>& registeration) {
- custom_registrations_[name] = registeration;
+ const std::function<TfLiteRegistration*()>& registration) {
+ custom_registrations_[name] = registration;
opcodes_.push_back(
CreateOperatorCodeDirect(builder_, BuiltinOperator_CUSTOM, name.data()));
operators_.push_back(CreateOperator(
diff --git a/tensorflow/contrib/lite/kernels/unidirectional_sequence_lstm.cc b/tensorflow/contrib/lite/kernels/unidirectional_sequence_lstm.cc
index 508a570..42941a9 100644
--- a/tensorflow/contrib/lite/kernels/unidirectional_sequence_lstm.cc
+++ b/tensorflow/contrib/lite/kernels/unidirectional_sequence_lstm.cc
@@ -360,7 +360,7 @@
const int n_output = recurrent_to_output_weights->dims->data[1];
// Since we have already checked that weights are all there or none, we can
- // check the existense of only one to the get the condition.
+ // check the existence of only one to get the condition.
const bool use_cifg = (input_to_input_weights == nullptr);
const bool use_peephole = (cell_to_output_weights != nullptr);
diff --git a/tensorflow/contrib/lite/memory_planner.h b/tensorflow/contrib/lite/memory_planner.h
index 5cd6c20..0294ec8 100644
--- a/tensorflow/contrib/lite/memory_planner.h
+++ b/tensorflow/contrib/lite/memory_planner.h
@@ -34,8 +34,8 @@
// [first_node, last_node].
virtual TfLiteStatus ExecuteAllocations(int first_node, int last_node) = 0;
- // Invalidates allocations made earliers. This is called when tensors sizes
- // have change. All planned allocations remain, but can't be used until
+ // Invalidates allocations made earlier. This is called when tensors sizes
+ // have changed. All planned allocations remain, but can't be used until
// ExecuteAllocations() is called.
virtual TfLiteStatus ResetAllocations() = 0;
};
diff --git a/tensorflow/contrib/lite/model.h b/tensorflow/contrib/lite/model.h
index a467df5..8dc1c79 100644
--- a/tensorflow/contrib/lite/model.h
+++ b/tensorflow/contrib/lite/model.h
@@ -64,7 +64,7 @@
const tflite::Model* model_spec,
ErrorReporter* error_reporter = DefaultErrorReporter());
- // Releases memory or unmaps mmaped meory.
+ // Releases memory or unmaps mmaped memory.
~FlatBufferModel();
// Copying or assignment is disallowed to simplify ownership semantics.
diff --git a/tensorflow/contrib/lite/nnapi/NeuralNetworksShim.h b/tensorflow/contrib/lite/nnapi/NeuralNetworksShim.h
index 7603277..bd49d32 100644
--- a/tensorflow/contrib/lite/nnapi/NeuralNetworksShim.h
+++ b/tensorflow/contrib/lite/nnapi/NeuralNetworksShim.h
@@ -569,7 +569,7 @@
ANEURALNETWORKS_LOGISTIC = 14,
/**
- * Projects an input to a bit vector via locality senstive hashing.
+ * Projects an input to a bit vector via locality sensitive hashing.
*
* Inputs:
* * 0: Hash functions. Dim.size == 2, DataType: Float.
diff --git a/tensorflow/contrib/lite/rpi_makefile.inc b/tensorflow/contrib/lite/rpi_makefile.inc
new file mode 100644
index 0000000..832ef58
--- /dev/null
+++ b/tensorflow/contrib/lite/rpi_makefile.inc
@@ -0,0 +1,33 @@
+# Settings for Raspberry Pi.
+ifeq ($(TARGET), RPI)
+ ifeq ($(TARGET_ARCH), armv7)
+ CXXFLAGS += \
+ -march=armv7-a \
+ -mfpu=neon-vfpv4 \
+ -funsafe-math-optimizations \
+ -ftree-vectorize
+
+ CCFLAGS += \
+ -march=armv7-a \
+ -mfpu=neon-vfpv4 \
+ -funsafe-math-optimizations \
+ -ftree-vectorize
+
+ LDFLAGS := \
+ -Wl,--no-export-dynamic \
+ -Wl,--exclude-libs,ALL \
+ -Wl,--gc-sections \
+ -Wl,--as-needed
+ endif
+
+ LIBS := \
+ -lstdc++ \
+ -lpthread \
+ -lm \
+ -ldl
+
+ OBJDIR := $(OBJDIR)rpi_$(TARGET_ARCH)/
+ LIBDIR := $(LIBDIR)rpi_$(TARGET_ARCH)/
+ BINDIR := $(BINDIR)rpi_$(TARGET_ARCH)/
+ DEPDIR := $(DEPDIR)rpi_$(TARGET_ARCH)/
+endif
diff --git a/tensorflow/contrib/lite/schema/builtin_ops_header/generator.cc b/tensorflow/contrib/lite/schema/builtin_ops_header/generator.cc
index 08bcfe4..640972d 100644
--- a/tensorflow/contrib/lite/schema/builtin_ops_header/generator.cc
+++ b/tensorflow/contrib/lite/schema/builtin_ops_header/generator.cc
@@ -46,7 +46,7 @@
#endif // __cplusplus
// The enum for builtin operators.
-// Note: CUSTOM and DELEGATE are 2 special ops which are not real biultin
+// Note: CUSTOM and DELEGATE are 2 special ops which are not real builtin
// ops.
typedef enum {
)";
diff --git a/tensorflow/contrib/lite/simple_memory_arena.cc b/tensorflow/contrib/lite/simple_memory_arena.cc
index 4aab244..2f2004f 100644
--- a/tensorflow/contrib/lite/simple_memory_arena.cc
+++ b/tensorflow/contrib/lite/simple_memory_arena.cc
@@ -113,21 +113,21 @@
underlying_buffer_size_ = required_size;
underlying_buffer_aligned_ptr_ = new_underlying_buffer_aligned_ptr;
}
- commited_ = true;
+ committed_ = true;
return underlying_buffer_ != nullptr ? kTfLiteOk : kTfLiteError;
}
TfLiteStatus SimpleMemoryArena::ResolveAlloc(TfLiteContext* context,
const ArenaAlloc& alloc,
char** output_ptr) {
- TF_LITE_ENSURE(context, commited_);
+ TF_LITE_ENSURE(context, committed_);
TF_LITE_ENSURE(context, output_ptr != nullptr);
*output_ptr = underlying_buffer_aligned_ptr_ + alloc.offset;
return kTfLiteOk;
}
TfLiteStatus SimpleMemoryArena::Clear() {
- commited_ = false;
+ committed_ = false;
high_water_mark_ = 0;
allocs_.clear();
return kTfLiteOk;
diff --git a/tensorflow/contrib/lite/simple_memory_arena.h b/tensorflow/contrib/lite/simple_memory_arena.h
index 0535522..5faf78b 100644
--- a/tensorflow/contrib/lite/simple_memory_arena.h
+++ b/tensorflow/contrib/lite/simple_memory_arena.h
@@ -22,7 +22,7 @@
namespace tflite {
// This little structure holds the offset and the size for a dynamic memory
-// allocation in the memory arena. When the arena is commited and the
+// allocation in the memory arena. When the arena is committed and the
// underlying buffer is set, the alloc can be resolved into an actual memory
// pointer.
struct ArenaAlloc {
@@ -43,7 +43,7 @@
class SimpleMemoryArena {
public:
explicit SimpleMemoryArena(size_t arena_alignment)
- : commited_(false),
+ : committed_(false),
arena_alignment_(arena_alignment),
high_water_mark_(0),
underlying_buffer_size_(0),
@@ -73,7 +73,7 @@
}
private:
- bool commited_;
+ bool committed_;
size_t arena_alignment_;
size_t high_water_mark_;
std::unique_ptr<char[]> underlying_buffer_;
diff --git a/tensorflow/contrib/lookup/BUILD b/tensorflow/contrib/lookup/BUILD
index 8ca03f4..0a6edc3 100644
--- a/tensorflow/contrib/lookup/BUILD
+++ b/tensorflow/contrib/lookup/BUILD
@@ -46,6 +46,7 @@
"//tensorflow/python:variables",
],
grpc_enabled = True,
+ tags = ["no_windows"], # TODO: needs investigation on Windows
)
filegroup(
diff --git a/tensorflow/contrib/makefile/README.md b/tensorflow/contrib/makefile/README.md
index 995230d..6c3b02e 100644
--- a/tensorflow/contrib/makefile/README.md
+++ b/tensorflow/contrib/makefile/README.md
@@ -194,6 +194,8 @@
srcs = glob(["libs/arm64-v8a/*.so"]),
```
+If you are building for Android TV (Shield TV devices), replace "portrait" with "landscape" for android:screenOrientation in all four activities in tensorflow/examples/android/AndroidManifest.xml
+
Then run:
```bash
# Create dir for native libs
diff --git a/tensorflow/contrib/makefile/build_all_ios.sh b/tensorflow/contrib/makefile/build_all_ios.sh
index 2d99791..9b14868 100755
--- a/tensorflow/contrib/makefile/build_all_ios.sh
+++ b/tensorflow/contrib/makefile/build_all_ios.sh
@@ -80,10 +80,9 @@
fi
else
echo "${PRNT_SLCTV_BIN} found. Using it"
- ${PRNT_SLCTV_BIN} --graphs=${OPTIMIZE_FOR_GRAPH} > ${TOP_SRCDIR}/tensorflow/core/framework/ops_to_register.h
-
fi
+ ${PRNT_SLCTV_BIN} --graphs=${OPTIMIZE_FOR_GRAPH} > ${TOP_SRCDIR}/tensorflow/core/framework/ops_to_register.h
fi
if [[ "${ONLY_MAKE_TENSORFLOW}" != "true" ]]; then
diff --git a/tensorflow/contrib/mpi/mpi_utils.h b/tensorflow/contrib/mpi/mpi_utils.h
index fa297c2..df055ff 100644
--- a/tensorflow/contrib/mpi/mpi_utils.h
+++ b/tensorflow/contrib/mpi/mpi_utils.h
@@ -24,6 +24,8 @@
#include "tensorflow/core/lib/strings/str_util.h"
+// Skip MPI C++ bindings support, this matches the usage in other places
+#define OMPI_SKIP_MPICXX
#include "third_party/mpi/mpi.h"
#define MPI_CHECK(cmd) \
do { \
diff --git a/tensorflow/contrib/predictor/predictor_factories.py b/tensorflow/contrib/predictor/predictor_factories.py
index 04b5d5b..6e77e93 100644
--- a/tensorflow/contrib/predictor/predictor_factories.py
+++ b/tensorflow/contrib/predictor/predictor_factories.py
@@ -53,7 +53,7 @@
`Estimator`.
"""
if isinstance(estimator, core_estimator.Estimator):
- raise TypeError('Espected estimator to be of type '
+ raise TypeError('Expected estimator to be of type '
'tf.contrib.learn.Estimator, but got type '
'tf.python.estimator.Estimator. You likely want to call '
'from_estimator.')
@@ -88,7 +88,7 @@
`Estimator`.
"""
if isinstance(estimator, contrib_estimator.Estimator):
- raise TypeError('Espected estimator to be of type '
+ raise TypeError('Expected estimator to be of type '
'tf.python.estimator.Estimator, but got type '
'tf.contrib.learn.Estimator. You likely want to call '
'from_contrib_estimator.')
diff --git a/tensorflow/contrib/py2tf/converters/BUILD b/tensorflow/contrib/py2tf/converters/BUILD
index f624c42..4bb6f76 100644
--- a/tensorflow/contrib/py2tf/converters/BUILD
+++ b/tensorflow/contrib/py2tf/converters/BUILD
@@ -81,6 +81,7 @@
name = "builtin_functions_test",
srcs = ["builtin_functions_test.py"],
srcs_version = "PY2AND3",
+ tags = ["no_windows"], # TODO: needs investigation on Windows
deps = [
":test_lib",
"//tensorflow/python:client_testlib",
@@ -91,6 +92,7 @@
name = "call_trees_test",
srcs = ["call_trees_test.py"],
srcs_version = "PY2AND3",
+ tags = ["no_windows"], # TODO: needs investigation on Windows
deps = [
":test_lib",
"//tensorflow/contrib/py2tf/impl",
diff --git a/tensorflow/contrib/py2tf/converters/single_return.py b/tensorflow/contrib/py2tf/converters/single_return.py
index 90bc220..1194b98 100644
--- a/tensorflow/contrib/py2tf/converters/single_return.py
+++ b/tensorflow/contrib/py2tf/converters/single_return.py
@@ -212,7 +212,7 @@
def __init__(self):
self.cant_return = False
- super(gast.NodeVisitor, self).__init__()
+ super(DetectReturnInUnsupportedControlFlow, self).__init__()
def visit_While(self, node):
self.cant_return = True
diff --git a/tensorflow/contrib/py2tf/utils/BUILD b/tensorflow/contrib/py2tf/utils/BUILD
index d029289..8bc338e 100644
--- a/tensorflow/contrib/py2tf/utils/BUILD
+++ b/tensorflow/contrib/py2tf/utils/BUILD
@@ -83,6 +83,7 @@
name = "py_func_test",
srcs = ["py_func_test.py"],
srcs_version = "PY2AND3",
+ tags = ["no_windows"], # TODO: needs investigation on Windows
deps = [
":utils",
"//tensorflow/python:client_testlib",
diff --git a/tensorflow/contrib/quantize/python/fold_batch_norms.py b/tensorflow/contrib/quantize/python/fold_batch_norms.py
index b278265..1afcbb8 100644
--- a/tensorflow/contrib/quantize/python/fold_batch_norms.py
+++ b/tensorflow/contrib/quantize/python/fold_batch_norms.py
@@ -237,7 +237,7 @@
# The batch variance used during forward and backward prop is biased,
# i.e it is calculated as: V=sum(x(k)-mu)^2/N. For the moving average
# calculation, the variance is corrected by the term N/N-1 (Bessel's
- # correction). The variance tensor read from FuseBatchNorm has bessel's
+ # correction). The variance tensor read from FuseBatchNorm has Bessel's
# correction applied, so we undo it here.
scope, sep, _ = bn_op.name.rpartition('/')
g = ops.get_default_graph()
@@ -306,7 +306,7 @@
Args:
context: The scope under which we look for batch norm params
- match: Object containg required batch norm tensors for correction
+ match: Object containing required batch norm tensors for correction
computation.
freeze_batch_norm_delay: Delay in steps at which computation switches
from regular batch norm to frozen mean and variance.
diff --git a/tensorflow/contrib/quantize/python/quant_ops.py b/tensorflow/contrib/quantize/python/quant_ops.py
index 0a8e350..a4f7b1b 100644
--- a/tensorflow/contrib/quantize/python/quant_ops.py
+++ b/tensorflow/contrib/quantize/python/quant_ops.py
@@ -282,8 +282,8 @@
Args:
inputs: a tensor containing values to be quantized.
min_var: a variable containing quantization range lower end(s).
- max_var: a variable containing quantization range lupper end(s).
- per_channel: a boolean specifying whether to use per-channel quantizatioh.
+ max_var: a variable containing quantization range upper end(s).
+ per_channel: a boolean specifying whether to use per-channel quantization.
num_bits: Number of bits to use for quantization, must be between 2 and 8.
narrow_range: Whether to use the narrow quantization range
[1; 2^num_bits - 1] or wide range [0; 2^num_bits - 1].
diff --git a/tensorflow/contrib/quantize/python/quantize.py b/tensorflow/contrib/quantize/python/quantize.py
index 0608ab9..ec721af 100644
--- a/tensorflow/contrib/quantize/python/quantize.py
+++ b/tensorflow/contrib/quantize/python/quantize.py
@@ -267,7 +267,7 @@
"""Inserts a quant op between a producer op and (multiple) consumer ops.
Args:
- context: Context w,here producer and consumer operations are nested.
+ context: Context where producer and consumer operations are nested.
name: Name for the new quantization op within the context.
producer: Producer operation of the pairs where quantization will be
inserted.
diff --git a/tensorflow/contrib/quantize/python/quantize_graph.py b/tensorflow/contrib/quantize/python/quantize_graph.py
index 5a3a74c..5abdcd2 100644
--- a/tensorflow/contrib/quantize/python/quantize_graph.py
+++ b/tensorflow/contrib/quantize/python/quantize_graph.py
@@ -158,7 +158,7 @@
often fail.
Args:
- input_graph: The tf.Graph to be transformed,if None then defaults to the
+ input_graph: The tf.Graph to be transformed, if None then defaults to the
default graph.
weight_bits: Number of bits to use for quantizing weights.
activation_bits: Number of bits to use for quantizing activations.
diff --git a/tensorflow/contrib/quantize/python/quantize_parameterized_test.py b/tensorflow/contrib/quantize/python/quantize_parameterized_test.py
index 0624cc8..db745aa 100644
--- a/tensorflow/contrib/quantize/python/quantize_parameterized_test.py
+++ b/tensorflow/contrib/quantize/python/quantize_parameterized_test.py
@@ -419,7 +419,7 @@
normalizer_params=self._BatchNormParams(fused_batch_norm),
scope=scope)
- # Manually add a bypass (optionaly) and an activation.
+ # Manually add a bypass (optional) and an activation.
if with_bypass:
node = math_ops.add(inputs, node, name='test/Add')
@@ -470,7 +470,7 @@
normalizer_params=self._BatchNormParams(fused_batch_norm),
scope=scope)
- # Manually add a bypass (optionaly) and an activation.
+ # Manually add a bypass (optional) and an activation.
if with_bypass:
node = math_ops.add(inputs, node, name='test/Add')
@@ -526,7 +526,7 @@
normalizer_params=self._BatchNormParams(fused_batch_norm),
scope=scope)
- # Manually add a bypass (optionaly) and an activation.
+ # Manually add a bypass (optional) and an activation.
if with_bypass:
node = math_ops.add(inputs, node, name='test/Add')
@@ -565,7 +565,7 @@
stddev: Standard deviation of normal variable.
Returns:
- An initialized that initialzes with a truncated normal variable.
+ An initialized that initializes with a truncated normal variable.
"""
return init_ops.truncated_normal_initializer(stddev=stddev)
diff --git a/tensorflow/contrib/quantize/python/quantize_test.py b/tensorflow/contrib/quantize/python/quantize_test.py
index ef59475..b2e5707 100644
--- a/tensorflow/contrib/quantize/python/quantize_test.py
+++ b/tensorflow/contrib/quantize/python/quantize_test.py
@@ -144,7 +144,7 @@
stddev: Standard deviation of normal variable.
Returns:
- An initialized that initialzes with a truncated normal variable.
+ An initialized that initializes with a truncated normal variable.
"""
return init_ops.truncated_normal_initializer(stddev=stddev)
diff --git a/tensorflow/contrib/remote_fused_graph/pylib/BUILD b/tensorflow/contrib/remote_fused_graph/pylib/BUILD
index 27f0a7f..54c6627 100644
--- a/tensorflow/contrib/remote_fused_graph/pylib/BUILD
+++ b/tensorflow/contrib/remote_fused_graph/pylib/BUILD
@@ -38,7 +38,6 @@
size = "small",
srcs = ["python/ops/remote_fused_graph_ops_test.py"],
srcs_version = "PY2AND3",
- tags = ["no_windows"],
deps = [
":remote_fused_graph_ops_py",
"//tensorflow/core:protos_all_py",
diff --git a/tensorflow/contrib/rnn/python/ops/rnn_cell.py b/tensorflow/contrib/rnn/python/ops/rnn_cell.py
index 73f2607..358b2eb 100644
--- a/tensorflow/contrib/rnn/python/ops/rnn_cell.py
+++ b/tensorflow/contrib/rnn/python/ops/rnn_cell.py
@@ -2133,7 +2133,7 @@
def __init__(self, name="conv_1d_lstm_cell", **kwargs):
"""Construct Conv1DLSTM. See `ConvLSTMCell` for more details."""
- super(Conv1DLSTMCell, self).__init__(conv_ndims=1, **kwargs)
+ super(Conv1DLSTMCell, self).__init__(conv_ndims=1, name=name, **kwargs)
class Conv2DLSTMCell(ConvLSTMCell):
@@ -2144,7 +2144,7 @@
def __init__(self, name="conv_2d_lstm_cell", **kwargs):
"""Construct Conv2DLSTM. See `ConvLSTMCell` for more details."""
- super(Conv2DLSTMCell, self).__init__(conv_ndims=2, **kwargs)
+ super(Conv2DLSTMCell, self).__init__(conv_ndims=2, name=name, **kwargs)
class Conv3DLSTMCell(ConvLSTMCell):
@@ -2155,7 +2155,7 @@
def __init__(self, name="conv_3d_lstm_cell", **kwargs):
"""Construct Conv3DLSTM. See `ConvLSTMCell` for more details."""
- super(Conv3DLSTMCell, self).__init__(conv_ndims=3, **kwargs)
+ super(Conv3DLSTMCell, self).__init__(conv_ndims=3, name=name, **kwargs)
def _conv(args, filter_size, num_features, bias, bias_start=0.0):
diff --git a/tensorflow/contrib/saved_model/BUILD b/tensorflow/contrib/saved_model/BUILD
index 245fe07..b10757d 100644
--- a/tensorflow/contrib/saved_model/BUILD
+++ b/tensorflow/contrib/saved_model/BUILD
@@ -53,6 +53,7 @@
size = "small",
srcs = ["python/saved_model/reader_test.py"],
srcs_version = "PY2AND3",
+ tags = ["no_windows"], # TODO: needs investigation on Windows
visibility = ["//visibility:private"],
deps = [
":saved_model_py",
diff --git a/tensorflow/contrib/seq2seq/python/ops/beam_search_decoder.py b/tensorflow/contrib/seq2seq/python/ops/beam_search_decoder.py
index 03fe31a..6adbb8b 100644
--- a/tensorflow/contrib/seq2seq/python/ops/beam_search_decoder.py
+++ b/tensorflow/contrib/seq2seq/python/ops/beam_search_decoder.py
@@ -299,12 +299,13 @@
"""
finished, start_inputs = self._finished, self._start_inputs
+ dtype = nest.flatten(self._initial_cell_state)[0].dtype
log_probs = array_ops.one_hot( # shape(batch_sz, beam_sz)
array_ops.zeros([self._batch_size], dtype=dtypes.int32),
depth=self._beam_width,
- on_value=0.0,
- off_value=-np.Inf,
- dtype=nest.flatten(self._initial_cell_state)[0].dtype)
+ on_value=ops.convert_to_tensor(0.0, dtype=dtype),
+ off_value=ops.convert_to_tensor(-np.Inf, dtype=dtype),
+ dtype=dtype)
initial_state = BeamSearchDecoderState(
cell_state=self._initial_cell_state,
diff --git a/tensorflow/contrib/session_bundle/BUILD b/tensorflow/contrib/session_bundle/BUILD
index 67011c8..3ad88a8 100644
--- a/tensorflow/contrib/session_bundle/BUILD
+++ b/tensorflow/contrib/session_bundle/BUILD
@@ -165,6 +165,7 @@
name = "gc_test",
srcs = ["gc_test.py"],
srcs_version = "PY2AND3",
+ tags = ["no_windows"], # TODO: needs investigation on Windows
visibility = ["//visibility:private"],
deps = [
":gc",
diff --git a/tensorflow/contrib/slim/python/slim/data/BUILD b/tensorflow/contrib/slim/python/slim/data/BUILD
index 5daabbd..7aa1684 100644
--- a/tensorflow/contrib/slim/python/slim/data/BUILD
+++ b/tensorflow/contrib/slim/python/slim/data/BUILD
@@ -61,6 +61,7 @@
name = "dataset_data_provider_test",
srcs = ["dataset_data_provider_test.py"],
srcs_version = "PY2AND3",
+ tags = ["no_windows"], # TODO: needs investigation on Windows
deps = [
":dataset",
":dataset_data_provider",
diff --git a/tensorflow/contrib/tensor_forest/BUILD b/tensorflow/contrib/tensor_forest/BUILD
index 1e4cc3f..07b6b1f 100644
--- a/tensorflow/contrib/tensor_forest/BUILD
+++ b/tensorflow/contrib/tensor_forest/BUILD
@@ -553,7 +553,6 @@
srcs = ["client/random_forest_test.py"],
srcs_version = "PY2AND3",
tags = [
- "no_windows",
"nomac", # b/63258195
"notsan",
],
diff --git a/tensorflow/contrib/tensorboard/BUILD b/tensorflow/contrib/tensorboard/BUILD
index d833744..db2e000 100644
--- a/tensorflow/contrib/tensorboard/BUILD
+++ b/tensorflow/contrib/tensorboard/BUILD
@@ -9,6 +9,7 @@
# For platform specific build config
load("//tensorflow/core:platform/default/build_config.bzl", "tf_proto_library")
+load("//tensorflow:tensorflow.bzl", "py_test")
tf_proto_library(
name = "protos_all",
diff --git a/tensorflow/contrib/tensorrt/BUILD b/tensorflow/contrib/tensorrt/BUILD
index c832c6f..906cc3f 100644
--- a/tensorflow/contrib/tensorrt/BUILD
+++ b/tensorflow/contrib/tensorrt/BUILD
@@ -83,6 +83,7 @@
"kernels/trt_engine_op.h",
],
copts = tf_copts(),
+ visibility = ["//visibility:public"],
deps = [
":trt_logging",
":trt_resources",
@@ -154,6 +155,7 @@
deps = [
":trt_convert_py",
":trt_ops_py",
+ "//tensorflow/python:errors",
],
)
diff --git a/tensorflow/contrib/tensorrt/README.md b/tensorflow/contrib/tensorrt/README.md
index dfcce0f..461e627 100644
--- a/tensorflow/contrib/tensorrt/README.md
+++ b/tensorflow/contrib/tensorrt/README.md
@@ -2,7 +2,8 @@
============================
This module provides necessary bindings and introduces TRT_engine_op
-operator that wraps a subgraph in TensorRT.
+operator that wraps a subgraph in TensorRT. This is still a work in progress
+but should be useable with most common graphs.
Compilation
-----------
@@ -15,26 +16,10 @@
automatically. If installed from tar packages, user has to set path to
location where the library is installed during configuration.
-
-```
+```shell
bazel build --config=cuda --config=opt //tensorflow/tools/pip_package:build_pip_package
bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/
```
After the installation of tensorflow package, TensorRT transformation
-will be available. An example use is shown below.
-
-```python
-import tensorflow as tf
-import tensorflow.contrib.tensorrt as trt
-#... create and train or load model
-gdef = sess.graph.as_graph_def()
-trt_gdef = trt.create_inference_graph(
- gdef, #original graph_def
- ["output"], #name of output node(s)
- max_batch_size, #maximum batch size to run the inference
- max_workspace_size_bytes) # max memory for TensorRT to use
-tf.reset_default_graph()
-tf.import_graph_def(graph_def=trt_gdef)
-#...... run inference
-```
+will be available. An example use can be found in test/test_tftrt.py directory
diff --git a/tensorflow/contrib/tensorrt/__init__.py b/tensorflow/contrib/tensorrt/__init__.py
index fd551d70..140ad48 100644
--- a/tensorflow/contrib/tensorrt/__init__.py
+++ b/tensorflow/contrib/tensorrt/__init__.py
@@ -18,6 +18,18 @@
from __future__ import division
from __future__ import print_function
-# pylint: disable=unused-import,wildcard-import
-from tensorflow.contrib.tensorrt.python import *
-# pylint: enable=unused-import,wildcard-import
+from tensorflow.python.framework import errors
+
+# pylint: disable=unused-import,wildcard-import,g-import-not-at-top
+try:
+ from tensorflow.contrib.tensorrt.python import *
+except errors.NotFoundError as e:
+ no_trt_message = (
+ '**** Failed to initialize TensorRT. This is either because the TensorRT'
+ ' installation path is not in LD_LIBRARY_PATH, or because you do not have'
+ ' it installed. If not installed, please go to'
+ ' https://developer.nvidia.com/tensorrt to download and install'
+ ' TensorRT ****')
+ print(no_trt_message)
+ raise e
+# pylint: enable=unused-import,wildcard-import,g-import-not-at-top
diff --git a/tensorflow/contrib/tensorrt/convert/convert_graph.cc b/tensorflow/contrib/tensorrt/convert/convert_graph.cc
index 970f810..eea8c8e 100644
--- a/tensorflow/contrib/tensorrt/convert/convert_graph.cc
+++ b/tensorflow/contrib/tensorrt/convert/convert_graph.cc
@@ -15,6 +15,7 @@
#include "tensorflow/contrib/tensorrt/convert/convert_graph.h"
+#include <list>
#include <map>
#include <set>
#include <unordered_map>
@@ -48,13 +49,29 @@
namespace convert {
namespace {
-static bool IsTensorRTCandidate(const tensorflow::NodeDef& node_def) {
+bool IsTensorRTCandidate(const tensorflow::NodeDef& node_def) {
// LINT.IfChange
// TODO(jie): Segmentation shouldn't associated with op name.
// Split it into a registration for each kernel.
static const std::set<string> candidate_ops = {
- "Identity", "Const", "Conv2D", "MaxPool", "BiasAdd", "Relu",
- "Add", "Mul", "Sub", "Rsqrt", "Pad" // "Placeholder" ,"Mean"
+ "Identity",
+ "Const",
+ "Conv2D",
+ "MaxPool",
+ "BiasAdd",
+ "Relu",
+ "Add",
+ "Mul",
+ "Sub",
+ "Rsqrt",
+ "Pad",
+ "Mean",
+ "AvgPool",
+ "ConcatV2",
+ "DepthwiseConv2dNative",
+ "FusedBatchNorm",
+ "FusedBatchNormV2",
+ // TODO(ben,jie): ...
};
// LINT.ThenChange(//tensorflow/contrib/tensorrt/convert/convert_nodes.h)
return candidate_ops.count(node_def.op());
@@ -69,6 +86,8 @@
if (!subgraph_node_ids.count(edge->src()->id()) &&
!edge->src()->IsSource()) {
incoming_edges->insert(edge);
+ } else {
+ VLOG(2) << edge->src()->name() << " N, ";
}
}
}
@@ -82,7 +101,10 @@
for (const tensorflow::Edge* edge : node->out_edges()) {
if (!subgraph_node_ids.count(edge->dst()->id()) &&
!edge->dst()->IsSink()) {
+ VLOG(2) << edge->dst()->name() << " Y, ";
outgoing_edges->insert(edge);
+ } else {
+ VLOG(2) << edge->dst()->name() << " N, ";
}
}
}
@@ -109,74 +131,150 @@
}
return result;
}
-
-tensorflow::Status ConvertSubGraphToTensorRT(
- const std::vector<string>& output_names,
- const std::set<int>& subgraph_node_ids,
- size_t max_batch_size, // Max batch size that engine will be created for
- // Max amount of memory that engine will be allowed to consume, in bytes
- size_t max_workspace_size_bytes,
- const tensorflow::grappler::GraphProperties& graph_properties,
- tensorflow::Graph* graph) {
- tensorflow::EdgeSet subgraph_incoming_edges;
- GetSubGraphIncomingEdges(*graph, subgraph_node_ids, &subgraph_incoming_edges);
-
+// TODO(sami): convert references to pointers
+struct ConvertGraphParams {
+ ConvertGraphParams(
+ tensorflow::Graph& inp_graph,
+ const std::vector<string>& output_node_names,
+ const std::set<int>& subgraph_node_id_numbers,
+ size_t max_supported_batch_size, size_t max_consumed_workspace_size_bytes,
+ const tensorflow::grappler::GraphProperties& current_graph_properties,
+ std::unordered_map<string, std::pair<int, string>>* output_edges,
+ int engine_precision_mode)
+ : graph(inp_graph),
+ output_names(output_node_names),
+ subgraph_node_ids(subgraph_node_id_numbers),
+ max_batch_size(max_supported_batch_size),
+ max_workspace_size_bytes(max_consumed_workspace_size_bytes),
+ graph_properties(current_graph_properties),
+ output_edge_map(output_edges),
+ precision_mode(engine_precision_mode) {}
+ tensorflow::Graph& graph;
+ const std::vector<string>& output_names;
+ const std::set<int>& subgraph_node_ids;
+ size_t max_batch_size;
+ size_t max_workspace_size_bytes;
+ const tensorflow::grappler::GraphProperties& graph_properties;
+ std::unordered_map<string, std::pair<int, string>>* output_edge_map;
+ int precision_mode;
std::vector<std::pair<int, int>> subgraph_inputs;
+ std::vector<std::pair<int, int>> subgraph_outputs;
+ tensorflow::EdgeSet subgraph_incoming_edges;
+ tensorflow::EdgeSet subgraph_outgoing_edges;
+};
- // Collect inputs by looking for incoming edges
- for (const tensorflow::Edge* edge : subgraph_incoming_edges) {
- subgraph_inputs.push_back({edge->src()->id(), edge->src_output()});
+static tensorflow::Status FillSubGraphEdgeSets(ConvertGraphParams* p) {
+ GetSubGraphIncomingEdges(p->graph, p->subgraph_node_ids,
+ &p->subgraph_incoming_edges);
+ for (const tensorflow::Edge* edge : p->subgraph_incoming_edges) {
+ p->subgraph_inputs.push_back({edge->src()->id(), edge->src_output()});
}
+ auto output_name_to_index_map = BuildTensorNameMap(p->output_names);
std::set<std::pair<int, int>> subgraph_outputs_set;
// Collect outputs referenced from output_names
- auto output_name_to_index_map = BuildTensorNameMap(output_names);
- for (int node_id : subgraph_node_ids) {
- tensorflow::Node* node = graph->FindNodeId(node_id);
+ for (int node_id : p->subgraph_node_ids) {
+ tensorflow::Node* node = p->graph.FindNodeId(node_id);
if (output_name_to_index_map.count(node->name())) {
for (int index : output_name_to_index_map.at(node->name())) {
subgraph_outputs_set.insert({node_id, index});
}
}
}
- // Collect outputs referenced from outgoing edges
- tensorflow::EdgeSet subgraph_outgoing_edges;
- GetSubGraphOutgoingEdges(*graph, subgraph_node_ids, &subgraph_outgoing_edges);
- for (const tensorflow::Edge* edge : subgraph_outgoing_edges) {
+ GetSubGraphOutgoingEdges(p->graph, p->subgraph_node_ids,
+ &p->subgraph_outgoing_edges);
+ for (const tensorflow::Edge* edge : p->subgraph_outgoing_edges) {
subgraph_outputs_set.insert({edge->src()->id(), edge->src_output()});
}
- // Impose an ordering on the outputs
- std::vector<std::pair<int, int>> subgraph_outputs(
- subgraph_outputs_set.begin(), subgraph_outputs_set.end());
- // Build TensorRT node and add it to the graph
+ p->subgraph_outputs.reserve(subgraph_outputs_set.size());
+ p->subgraph_outputs.insert(p->subgraph_outputs.begin(),
+ subgraph_outputs_set.begin(),
+ subgraph_outputs_set.end());
+ return tensorflow::Status::OK();
+};
+
+tensorflow::Status GetCalibNode(ConvertGraphParams* params) {
+ TF_RETURN_IF_ERROR(FillSubGraphEdgeSets(params));
tensorflow::NodeDef trt_node_def;
- TF_RETURN_IF_ERROR(ConvertSubGraphToTensorRTNodeDef(
- *graph, subgraph_node_ids, subgraph_inputs, subgraph_outputs,
- max_batch_size, max_workspace_size_bytes, graph_properties,
- &trt_node_def));
+ SubGraphParams s(params->graph, params->subgraph_node_ids,
+ params->subgraph_inputs, params->subgraph_outputs,
+ params->max_batch_size, params->max_workspace_size_bytes,
+ params->graph_properties, params->output_edge_map,
+ &trt_node_def, params->precision_mode);
+ TF_RETURN_IF_ERROR(InjectCalibrationNode(s));
tensorflow::Status status;
- tensorflow::Node* trt_node = graph->AddNode(trt_node_def, &status);
+ tensorflow::Node* trt_node = params->graph.AddNode(trt_node_def, &status);
+
+ TF_RETURN_IF_ERROR(status);
+
+ for (auto in_edge :
+ params->subgraph_incoming_edges) { // loop over incoming edges and
+ // attach them to calib node
+ // tensorflow::Node* src_node = in_edge->src();
+ auto src_output = in_edge->src_output();
+ auto dst_node = in_edge->dst();
+ auto dst_input = in_edge->dst_input();
+ VLOG(1) << " update edge " << trt_node->name() << ":" << src_output
+ << " -> " << dst_node->name() << ":" << dst_input;
+ TF_RETURN_IF_ERROR(
+ params->graph.UpdateEdge(trt_node, src_output, dst_node, dst_input));
+ }
+ return tensorflow::Status::OK();
+}
+
+tensorflow::Status ConvertSubGraphToTensorRT(ConvertGraphParams* params) {
+ TF_RETURN_IF_ERROR(FillSubGraphEdgeSets(params));
+ tensorflow::NodeDef trt_node_def;
+
+ SubGraphParams s(params->graph, params->subgraph_node_ids,
+ params->subgraph_inputs, params->subgraph_outputs,
+ params->max_batch_size, params->max_workspace_size_bytes,
+ params->graph_properties, params->output_edge_map,
+ &trt_node_def, params->precision_mode);
+ TF_RETURN_IF_ERROR(ConvertSubGraphToTensorRTNodeDef(s));
+ tensorflow::Status status;
+ tensorflow::Node* trt_node = params->graph.AddNode(trt_node_def, &status);
+
+ // AddNode does not wire edges.
+ // Re-map incoming edges to use the new TRT node instead of the orig subgraph
+ std::map<std::pair<int, int>, int> subgraph_edge_to_input_map;
+ for (size_t i = 0; i < params->subgraph_inputs.size(); ++i) {
+ subgraph_edge_to_input_map.insert({params->subgraph_inputs.at(i), i});
+ }
+ for (const tensorflow::Edge* edge : params->subgraph_incoming_edges) {
+ std::pair<int, int> old_src = {edge->src()->id(), edge->src_output()};
+ int new_src_output = subgraph_edge_to_input_map.at(old_src);
+ params->graph.AddEdge(edge->src(), edge->src_output(), trt_node,
+ new_src_output);
+ params->graph.RemoveEdge(edge);
+ }
+
+ VLOG(2) << "new wiring edges: " << trt_node->in_edges().size();
+ for (const tensorflow::Edge* edge : trt_node->in_edges()) {
+ VLOG(2) << edge->src()->name() << " port: " << edge->src_output();
+ }
+
TF_RETURN_IF_ERROR(status);
// Re-map outgoing edges to use the new TRT node instead of the orig subgraph
std::map<std::pair<int, int>, int> subgraph_edge_to_output_map;
- for (size_t i = 0; i < subgraph_outputs.size(); ++i) {
- subgraph_edge_to_output_map.insert({subgraph_outputs.at(i), i});
+ for (size_t i = 0; i < params->subgraph_outputs.size(); ++i) {
+ subgraph_edge_to_output_map.insert({params->subgraph_outputs.at(i), i});
}
TF_RETURN_IF_ERROR(status);
- for (const tensorflow::Edge* edge : subgraph_outgoing_edges) {
+ for (const tensorflow::Edge* edge : params->subgraph_outgoing_edges) {
std::pair<int, int> old_src = {edge->src()->id(), edge->src_output()};
int new_src_output = subgraph_edge_to_output_map.at(old_src);
- TF_RETURN_IF_ERROR(graph->UpdateEdge(trt_node, new_src_output, edge->dst(),
- edge->dst_input()));
+ TF_RETURN_IF_ERROR(params->graph.UpdateEdge(
+ trt_node, new_src_output, edge->dst(), edge->dst_input()));
}
// Remove the original subgraph
- for (int node_id : subgraph_node_ids) {
- tensorflow::Node* node = graph->FindNodeId(node_id);
+ for (int node_id : params->subgraph_node_ids) {
+ tensorflow::Node* node = params->graph.FindNodeId(node_id);
// Don't remove the input placeholders
if (node->type_string() == "Placeholder") {
continue;
}
- graph->RemoveNode(node);
+ params->graph.RemoveNode(node);
}
return tensorflow::Status::OK();
}
@@ -194,12 +292,39 @@
}
} // namespace
+tensorflow::Status ConvertCalibGraphToInferGraph(
+ const tensorflow::GraphDef& graph_def, tensorflow::GraphDef* infer_graph) {
+ VLOG(0) << "Starting Calib Conversion";
+ tensorflow::Graph graph(tensorflow::OpRegistry::Global());
+ TF_RETURN_IF_ERROR(tensorflow::ConvertGraphDefToGraph(
+ tensorflow::GraphConstructorOptions(), graph_def, &graph));
+ // get calib nodes
+ std::vector<tensorflow::Node*> calib_nodes;
+ for (auto node : graph.op_nodes()) {
+ if (node->type_string() == "TRTCalibOp") {
+ VLOG(1) << "Found Calib Node";
+ calib_nodes.push_back(node);
+ }
+ }
+ VLOG(0) << "Num Calib nodes in graph= " << calib_nodes.size();
+ if (calib_nodes.size() == 0)
+ return tensorflow::errors::FailedPrecondition(
+ "Graph doesn't contain any calibration nodes!."
+ " Please generate calibration graph and run calibration first");
+ for (auto n : calib_nodes) {
+ TF_RETURN_IF_ERROR(
+ tensorrt::convert::ConvertCalibrationNodeToEngineNode(graph, n));
+ }
+ graph.ToGraphDef(infer_graph);
+ return tensorflow::Status::OK();
+}
tensorflow::Status ConvertGraphDefToTensorRT(
const tensorflow::GraphDef& graph_def,
const std::vector<string>& output_names, size_t max_batch_size,
- size_t max_workspace_size_bytes, tensorflow::GraphDef* new_graph_def) {
- // Optimization pass
+ size_t max_workspace_size_bytes, tensorflow::GraphDef* new_graph_def,
+ int precision_mode = FP32MODE, int minimum_segment_size = 3) {
+ // optimization pass
tensorflow::grappler::GrapplerItem item;
item.fetch = output_names;
tensorflow::GraphDef gdef;
@@ -209,16 +334,23 @@
tensorflow::grappler::LayoutOptimizer optimizer;
tensorflow::grappler::Cluster* cluster;
- // Virtual cluster
+ // virtual cluster
tensorflow::DeviceProperties device_properties;
+
device_properties.set_type("GPU");
device_properties.mutable_environment()->insert({"architecture", "6"});
cluster =
new tensorflow::grappler::VirtualCluster({{"/GPU:0", device_properties}});
+ // single machine
+ int num_cpu_cores = tensorflow::grappler::GetNumAvailableLogicalCPUCores();
+ int num_gpus = tensorflow::grappler::GetNumAvailableGPUs();
+ VLOG(2) << "cpu_cores: " << num_cpu_cores;
+ VLOG(2) << "gpus: " << num_gpus;
+
TF_RETURN_IF_ERROR(optimizer.Optimize(cluster, item, &gdef));
- // Constant folding
+ // constant folding
item.graph = gdef;
tensorflow::grappler::ConstantFolding fold(nullptr);
TF_RETURN_IF_ERROR(fold.Optimize(nullptr, item, &gdef));
@@ -226,7 +358,6 @@
// AJ refactoring shape inference through grappler/GraphProperties.
tensorflow::grappler::GraphProperties static_graph_properties(item);
TF_RETURN_IF_ERROR(static_graph_properties.InferStatically(false));
-
// Build full graph
tensorflow::FunctionLibraryDefinition flib(tensorflow::OpRegistry::Global(),
gdef.library());
@@ -243,7 +374,7 @@
}
// TODO(sami): this should be passed as a knob!!!!
- segment_options.minimum_segment_size = 2;
+ segment_options.minimum_segment_size = minimum_segment_size;
tensorflow::tensorrt::segment::SegmentNodesVector segments;
TF_RETURN_IF_ERROR(tensorrt::segment::SegmentGraph(
gdef, IsTensorRTCandidate, segment_options, &segments));
@@ -252,14 +383,37 @@
}
std::unordered_map<string, tensorflow::Node*> node_map;
TF_RETURN_IF_ERROR(BuildNodeMap(graph, &node_map));
+ std::unordered_map<string, std::pair<int, string>> output_edge_map;
+ int count = 0;
+ float total_num_nodes_in_segments = 0.;
+ for (auto s : segments) {
+ total_num_nodes_in_segments += s.size();
+ }
for (const std::set<string>& subgraph_node_names : segments) {
std::set<int> subgraph_node_ids;
+ size_t max_mem_per_engine =
+ max_workspace_size_bytes *
+ ((float)subgraph_node_names.size() / total_num_nodes_in_segments);
+ std::stringstream oss;
for (const string& node_name : subgraph_node_names) {
+ oss << " " << node_name;
subgraph_node_ids.insert(node_map.at(node_name)->id());
}
- TF_RETURN_IF_ERROR(ConvertSubGraphToTensorRT(
- output_names, subgraph_node_ids, max_batch_size,
- max_workspace_size_bytes, static_graph_properties, &graph));
+ VLOG(2) << "Subgraph nodes" << oss.str();
+ ConvertGraphParams p(graph, output_names, subgraph_node_ids, max_batch_size,
+ max_mem_per_engine, static_graph_properties,
+ &output_edge_map, precision_mode);
+ if (precision_mode == INT8MODE) {
+ TF_RETURN_IF_ERROR(GetCalibNode(&p));
+ } else {
+ tensorflow::Status status = ConvertSubGraphToTensorRT(&p);
+ if (status != tensorflow::Status::OK()) {
+ LOG(WARNING) << "subgraph conversion error for subgraph_index:" << count
+ << " due to: \n"
+ << status.ToString() << " SKIPPING......";
+ }
+ count++;
+ }
}
graph.ToGraphDef(new_graph_def);
return tensorflow::Status::OK();
diff --git a/tensorflow/contrib/tensorrt/convert/convert_graph.h b/tensorflow/contrib/tensorrt/convert/convert_graph.h
index 154ad3f..e1596e8 100644
--- a/tensorflow/contrib/tensorrt/convert/convert_graph.h
+++ b/tensorflow/contrib/tensorrt/convert/convert_graph.h
@@ -28,6 +28,11 @@
namespace tensorrt {
namespace convert {
+// This method converts an already generated calibration graph which was used in
+// calibration runs to an inference graph
+tensorflow::Status ConvertCalibGraphToInferGraph(
+ const tensorflow::GraphDef& graph_def, tensorflow::GraphDef* new_graph_def);
+
// max_batch_size: maximum batch size which can be used for inference for
// optimization targets inference run with max batch size.
// max_workspace_size_bytes: The upper bound of memory allowence for
@@ -35,7 +40,8 @@
tensorflow::Status ConvertGraphDefToTensorRT(
const tensorflow::GraphDef& graph_def,
const std::vector<string>& output_names, size_t max_batch_size,
- size_t max_workspace_size_bytes, tensorflow::GraphDef* new_graph_def);
+ size_t max_workspace_size_bytes, tensorflow::GraphDef* new_graph_def,
+ int precision_mode, int minimum_segment_size);
} // namespace convert
} // namespace tensorrt
diff --git a/tensorflow/contrib/tensorrt/convert/convert_nodes.cc b/tensorflow/contrib/tensorrt/convert/convert_nodes.cc
index 9ee717d..75a3c3d 100644
--- a/tensorflow/contrib/tensorrt/convert/convert_nodes.cc
+++ b/tensorflow/contrib/tensorrt/convert/convert_nodes.cc
@@ -24,6 +24,10 @@
#include <utility>
#include <vector>
+#include "tensorflow/contrib/tensorrt/log/trt_logger.h"
+#include "tensorflow/contrib/tensorrt/resources/trt_resource_manager.h"
+#include "tensorflow/contrib/tensorrt/resources/trt_resources.h"
+#include "tensorflow/core/framework/node_def.pb.h" // NOLINT
#include "tensorflow/core/framework/node_def_builder.h"
#include "tensorflow/core/framework/tensor_shape.pb.h" // NOLINT
#include "tensorflow/core/framework/types.h"
@@ -32,6 +36,7 @@
#include "tensorflow/core/graph/graph_constructor.h"
#include "tensorflow/core/lib/core/errors.h"
#include "tensorflow/core/lib/core/status.h"
+#include "tensorflow/core/lib/strings/str_util.h"
#include "tensorflow/core/lib/strings/strcat.h"
#include "tensorflow/core/platform/logging.h"
#include "tensorflow/core/platform/tensor_coding.h"
@@ -39,7 +44,6 @@
#if GOOGLE_CUDA
#if GOOGLE_TENSORRT
-#include "tensorflow/contrib/tensorrt/log/trt_logger.h"
#include "tensorrt/include/NvInfer.h"
// Check if the types are equal. Cast to int first so that failure log message
@@ -49,6 +53,7 @@
namespace tensorflow {
namespace tensorrt {
namespace convert {
+using ::tensorflow::strings::StrCat;
namespace {
@@ -65,7 +70,8 @@
*trt_dtype = nvinfer1::DataType::kHALF;
break;
default:
- return tensorflow::errors::InvalidArgument("Unsupported data type");
+ return tensorflow::errors::InvalidArgument(
+ "Unsupported data type " + tensorflow::DataTypeString(tf_dtype));
}
return tensorflow::Status::OK();
}
@@ -112,6 +118,18 @@
return padding;
}
+string GetCommonNameScope(const string& op_name_a, const string& op_name_b) {
+ size_t last_scope_separator = 0;
+ for (size_t i = 0; i < std::min(op_name_a.size(), op_name_b.size()); ++i) {
+ if (op_name_a[i] != op_name_b[i]) {
+ break;
+ } else if (op_name_a[i] == '/') {
+ last_scope_separator = i + 1;
+ }
+ }
+ return op_name_a.substr(0, last_scope_separator);
+}
+
class TRT_ShapedWeights {
public:
TRT_ShapedWeights(tensorflow::DataType type, const void* values,
@@ -245,6 +263,11 @@
}
template <>
+std::vector<string> TFAttrs::get<std::vector<string>>(string key) const {
+ auto attr = this->at(key)->list().s();
+ return std::vector<string>(attr.begin(), attr.end());
+}
+template <>
nvinfer1::Dims TFAttrs::get<nvinfer1::Dims>(string key) const {
auto values = this->get<std::vector<int>>(key);
nvinfer1::Dims dims;
@@ -266,6 +289,17 @@
return this->at(key)->type();
}
+template <>
+float TFAttrs::get<float>(string key) const {
+ return this->at(key)->f();
+}
+
+template <>
+bool TFAttrs::get<bool>(string key) const {
+ return this->at(key)->b();
+}
+
+// TODO(jie): reorder4 & reorder2 should be merged?
template <typename T>
void Reorder4(nvinfer1::DimsNCHW shape, const T* idata,
nvinfer1::DimsNCHW istrides, T* odata,
@@ -283,29 +317,87 @@
}
}
+template <typename T>
+void Reorder2(nvinfer1::DimsHW shape, const T* idata, nvinfer1::DimsHW istrides,
+ T* odata, nvinfer1::DimsHW ostrides) {
+ for (int h = 0; h < shape.h(); ++h) {
+ for (int w = 0; w < shape.w(); ++w) {
+ odata[h * ostrides.h() + w * ostrides.w()] =
+ idata[h * ostrides.h() + w * ostrides.w()];
+ }
+ }
+}
+
+// TODO(jie): fallback to tensorflow!!
+void ReorderCKtoKC(const TRT_ShapedWeights& iweights,
+ TRT_ShapedWeights* oweights) {
+ int c = iweights.shape_.d[0];
+ int k = iweights.shape_.d[1];
+ oweights->shape_.d[0] = k;
+ oweights->shape_.d[1] = c;
+ nvinfer1::DimsHW istrides = {1, k};
+ nvinfer1::DimsHW ostrides = {c, 1};
+ switch (iweights.type_) {
+ case tensorflow::DataType::DT_FLOAT: {
+ Reorder2({k, c}, static_cast<float const*>(iweights.GetValues()),
+ istrides,
+ static_cast<float*>(const_cast<void*>(oweights->GetValues())),
+ ostrides);
+ break;
+ }
+ case tensorflow::DataType::DT_HALF: {
+ Reorder2(
+ {k, c}, static_cast<Eigen::half const*>(iweights.GetValues()),
+ istrides,
+ static_cast<Eigen::half*>(const_cast<void*>(oweights->GetValues())),
+ ostrides);
+ break;
+ }
+ default:
+ LOG(FATAL) << "Unsupported type in reorder expected fp32 or fp16 but got "
+ << DataTypeString(iweights.type_);
+ }
+}
+
void ReorderRSCKToKCRS(const TRT_ShapedWeights& iweights,
- TRT_ShapedWeights* oweights) {
+ TRT_ShapedWeights* oweights, int num_groups) {
CHECK_EQ(iweights.type_, oweights->type_);
CHECK_EQ(iweights.size_bytes(), oweights->size_bytes());
int r = iweights.shape_.d[0];
int s = iweights.shape_.d[1];
- int c = iweights.shape_.d[2];
- int k = iweights.shape_.d[3];
- oweights->shape_.d[0] = k;
- oweights->shape_.d[1] = c;
+ // TRT requires GKcRS, while TF depthwise has RSCK
+ // where c=1, C=G
+ VLOG(2) << "num_groups: " << num_groups;
+ int c = iweights.shape_.d[2] / num_groups;
+ VLOG(2) << "c" << iweights.shape_.d[2] << " then " << c;
+ int k = iweights.shape_.d[3] * num_groups;
+ VLOG(2) << "k" << iweights.shape_.d[3] << " then " << k;
+ oweights->shape_.d[0] = k / num_groups;
+ oweights->shape_.d[1] = c * num_groups;
oweights->shape_.d[2] = r;
oweights->shape_.d[3] = s;
nvinfer1::DimsNCHW istrides = {1, k, s * k * c, c * k};
nvinfer1::DimsNCHW ostrides = {c * r * s, r * s, s, 1};
switch (iweights.type_) {
- case tensorflow::DataType::DT_FLOAT:
+ case tensorflow::DataType::DT_FLOAT: {
Reorder4({k, c, r, s}, static_cast<float const*>(iweights.GetValues()),
istrides,
static_cast<float*>(const_cast<void*>(oweights->GetValues())),
ostrides);
break;
+ }
+ case tensorflow::DataType::DT_HALF: {
+ Reorder4(
+ {k, c, r, s}, static_cast<Eigen::half const*>(iweights.GetValues()),
+ istrides,
+ static_cast<Eigen::half*>(const_cast<void*>(oweights->GetValues())),
+ ostrides);
+ break;
+ }
+
default:
- LOG(FATAL) << "!!!!!!!!!!!!!!!!!!!!!!!!broke!!!!!!!!!!!!";
+ LOG(FATAL) << "Unsupported type, expected fp32 or fp16 but got "
+ << DataTypeString(iweights.type_);
}
}
@@ -323,12 +415,11 @@
return std::shared_ptr<T>(obj, InferDeleter());
}
-// Logger for GIE info/warning/errors
class Converter;
using OpConverter =
std::function<tensorflow::Status(Converter&, const tensorflow::NodeDef&,
- std::vector<TRT_TensorOrWeights> const&,
+ const std::vector<TRT_TensorOrWeights>&,
std::vector<TRT_TensorOrWeights>*)>;
class Converter {
@@ -336,34 +427,57 @@
std::unordered_map<string, OpConverter> op_registry_;
nvinfer1::INetworkDefinition* trt_network_;
std::list<std::vector<uint8_t>> temp_bufs_;
-
+ tensorflow::tensorrt::TRTWeightStore* weight_store_;
+ bool fp16_;
void register_op_converters();
-
std::vector<TRT_TensorOrWeights> get_inputs(
const tensorflow::NodeDef& node_def) {
std::vector<TRT_TensorOrWeights> inputs;
- for (const auto& input_name : node_def.input()) {
- VLOG(2) << "Retrieve input: " << input_name;
- inputs.push_back(trt_tensors_.at(input_name));
+ for (auto const& input_name : node_def.input()) {
+ /*************************************************************************
+ * TODO(jie) handle case 1) here
+ * Normalizes the inputs and extracts associated metadata:
+ * 1) Inputs can contain a colon followed by a suffix of characters.
+ * That suffix may be a single number (e.g. inputName:1) or several
+ * word characters separated from a number by a colon
+ * (e.g. inputName:foo:1). The
+ * latter case is used to denote inputs and outputs of functions.
+ * 2) Control dependency inputs contain caret at the beginning and we
+ * remove this and annotate the edge as a control dependency.
+ ************************************************************************/
+ string name = input_name[0] == '^' ? input_name.substr(1) : input_name;
+ auto first = name.find_first_of(':');
+ if (first != string::npos && first + 2 == name.size() &&
+ name[first + 1] == '0')
+ name.erase(first);
+
+ VLOG(2) << "retrieve input: " << name;
+ if (trt_tensors_.count(name)) {
+ inputs.push_back(trt_tensors_.at(name));
+ } else {
+ LOG(FATAL) << "input: " << name << " not availabled for node at, "
+ << node_def.name();
+ }
}
return inputs;
}
public:
- explicit Converter(nvinfer1::INetworkDefinition* trt_network)
- : trt_network_(trt_network) {
+ explicit Converter(nvinfer1::INetworkDefinition* trt_network,
+ tensorflow::tensorrt::TRTWeightStore* ws, bool fp16)
+ : trt_network_(trt_network), weight_store_(ws), fp16_(fp16) {
this->register_op_converters();
}
-
+ tensorflow::tensorrt::TRTWeightStore* weight_store() { return weight_store_; }
TRT_ShapedWeights get_temp_weights(tensorflow::DataType type,
nvinfer1::Dims shape) {
TRT_ShapedWeights weights(type, nullptr, shape);
// TODO(jie): check weights size_bytes. 0 means type error
- temp_bufs_.push_back(std::vector<uint8_t>(weights.size_bytes()));
- weights.SetValues(temp_bufs_.back().data());
+ weight_store_->store_.push_back(std::vector<uint8_t>(weights.size_bytes()));
+ weights.SetValues(weight_store_->store_.back().data());
return weights;
}
-
+ bool isFP16() { return fp16_; };
TRT_ShapedWeights get_temp_weights_like(const TRT_ShapedWeights& weights) {
return this->get_temp_weights(weights.type_, weights.shape_);
}
@@ -382,7 +496,7 @@
TRT_TensorOrWeights output = outputs.at(i);
// TODO(jie): tf protobuf seems to be omitting the :0 suffix
string output_name = node_def.name();
- if (i != 0) output_name = output_name + ":" + std::to_string(i);
+ if (i != 0) output_name = StrCat(output_name, ":", i);
if (output.is_tensor()) {
output.tensor()->setName(output_name.c_str());
}
@@ -448,7 +562,7 @@
switch (op) {
case OP_CATEGORY::RSQRT: {
VLOG(2) << "RSQRT GETS DONE";
- return [](T t) -> T { return 1.0 / std::sqrt(t); };
+ return [](T t) -> T { return 1.0 / sqrt(t); };
}
case OP_CATEGORY::NEG:
return [](T t) -> T { return -t; };
@@ -534,6 +648,22 @@
}
};
+template <>
+std::function<Eigen::half(Eigen::half)> LambdaFactory::unary<Eigen::half>() {
+ switch (op) {
+ case OP_CATEGORY::RSQRT: {
+ VLOG(2) << "RSQRT GETS DONE";
+ return [](Eigen::half t) -> Eigen::half {
+ return Eigen::half(1.0 / sqrt(float(t)));
+ };
+ }
+ case OP_CATEGORY::NEG:
+ return [](Eigen::half t) -> Eigen::half { return -t; };
+ default:
+ VLOG(2) << "Not supported op for unary: " << static_cast<int>(op);
+ return nullptr;
+ }
+}
tensorflow::Status UnaryCompute(const TRT_ShapedWeights& iweights,
TRT_ShapedWeights* oweights,
LambdaFactory unary_op) {
@@ -545,6 +675,14 @@
std::transform(inp, inp + iweights.count(), oup, unary_op.unary<float>());
break;
}
+ case tensorflow::DataType::DT_HALF: {
+ auto inp = static_cast<Eigen::half const*>(iweights.GetValues());
+ auto oup =
+ static_cast<Eigen::half*>(const_cast<void*>(oweights->GetValues()));
+ std::transform(inp, inp + iweights.count(), oup,
+ unary_op.unary<Eigen::half>());
+ break;
+ }
default:
return tensorflow::errors::Unimplemented(
"Data type not supported: " +
@@ -588,6 +726,32 @@
}
break;
}
+ case tensorflow::DataType::DT_HALF: {
+ auto inp_l = static_cast<const Eigen::half*>(iweights_l.GetValues());
+ auto inp_r = static_cast<const Eigen::half*>(iweights_r.GetValues());
+ auto oup =
+ static_cast<Eigen::half*>(const_cast<void*>(oweights->GetValues()));
+
+ if (iweights_l.count() != iweights_r.count()) {
+ // We only supports broadcast of RankZero
+ if (iweights_l.count() == 1) {
+ VLOG(2) << "I bet it is not working!" << (*inp_l);
+ std::transform(inp_r, inp_r + iweights_r.count(), oup,
+ binary_op.broadcast_l<Eigen::half>(*inp_l));
+ } else if (iweights_r.count() == 1) {
+ VLOG(2) << "I bet it is not working!" << (*inp_r);
+ std::transform(inp_l, inp_l + iweights_l.count(), oup,
+ binary_op.broadcast_r<Eigen::half>(*inp_r));
+ } else {
+ return tensorflow::errors::Unimplemented(
+ "Binary op with non-rankZero broadcast not supported");
+ }
+ } else {
+ std::transform(inp_l, inp_l + iweights_l.count(), inp_r, oup,
+ binary_op.binary<Eigen::half>());
+ }
+ break;
+ }
default:
return tensorflow::errors::Unimplemented(
"Data type not supported: " +
@@ -599,7 +763,7 @@
tensorflow::Status ConstantFoldUnary(
Converter& ctx, const tensorflow::NodeDef& node_def,
- std::vector<TRT_TensorOrWeights> const& inputs,
+ const std::vector<TRT_TensorOrWeights>& inputs,
std::vector<TRT_TensorOrWeights>* outputs) {
TRT_ShapedWeights weights_input = inputs.at(0).weights();
@@ -613,13 +777,12 @@
CHECK_EQ(weights_input.type_,
TFAttrs(node_def).get<tensorflow::DataType>("T"));
- // Maybe I should do a switch
LambdaFactory unary_op;
if (node_def.op() == "Rsqrt") {
// Compute rsqrt
unary_op.op = LambdaFactory::OP_CATEGORY::RSQRT;
auto ret = UnaryCompute(weights_input, &weights_output, unary_op);
- // PAss the output
+ // Pass the output
if (ret == tensorflow::Status::OK()) {
outputs->push_back(TRT_TensorOrWeights(weights_output));
}
@@ -631,11 +794,11 @@
}
// TODO(jie,ben) broadcast is needed yet not implemented
-// Let's get the simple stuff working first. Maybe we should fall bakc to TF
+// Let's get the simple stuff working first. Maybe we should fall back to TF
// approach for constant folding
tensorflow::Status ConstantFoldBinary(
Converter& ctx, const tensorflow::NodeDef& node_def,
- std::vector<TRT_TensorOrWeights> const& inputs,
+ const std::vector<TRT_TensorOrWeights>& inputs,
std::vector<TRT_TensorOrWeights>* outputs) {
TRT_ShapedWeights weights_input_l = inputs.at(0).weights();
TRT_ShapedWeights weights_input_r = inputs.at(1).weights();
@@ -648,12 +811,12 @@
"Binary op implicit broadcast not supported: " + node_def.op());
// TODO(jie): constant fold should really fall back to TF.
- int nb_dims = weights_input_l.shape_.nbDims;
+ int num_dims = weights_input_l.shape_.nbDims;
nvinfer1::Dims output_shape;
- output_shape.nbDims = nb_dims;
- VLOG(2) << "nb_dims: " << nb_dims
+ output_shape.nbDims = num_dims;
+ VLOG(2) << "nb_dims: " << num_dims
<< ", the other: " << weights_input_r.shape_.nbDims;
- for (int i = 0; i < nb_dims; i++) {
+ for (int i = 0; i < num_dims; i++) {
if (weights_input_l.shape_.d[i] == weights_input_r.shape_.d[i]) {
output_shape.d[i] = weights_input_l.shape_.d[i];
} else if (weights_input_l.shape_.d[i] == 1 ||
@@ -678,7 +841,6 @@
// Allocate output weights
TRT_ShapedWeights weights_output = ctx.get_temp_weights(dtype, output_shape);
- // Maybe I should do a switch
LambdaFactory binary_op;
if (node_def.op() == "Sub") {
binary_op.op = LambdaFactory::OP_CATEGORY::SUB;
@@ -712,48 +874,90 @@
// Maybe this part has to be moved into the block of rsqrt later
// Check type consistency
- auto dtype = TFAttrs(node_def).get<nvinfer1::DataType>("T");
- CHECK_EQ_TYPE(tensor->getType(), dtype); // Cast to int for error messages
nvinfer1::DataType ttype;
TF_CHECK_OK(ConvertDType(weights.type_, &ttype));
- CHECK_EQ_TYPE(ttype, dtype); // Cast to int for error message
// Check scale mode
auto dims_w = weights.shape_;
auto dims_t = tensor->getDimensions();
- // Default to channel-wise
+ // default to element-wise
auto scale_mode = nvinfer1::ScaleMode::kELEMENTWISE;
+ // TODO(jie): maybe use a permuatation instead to support more cases;
+ bool permutation_flag = false;
+
if (weights.count() == 1) {
VLOG(2) << "UNIFORM";
scale_mode = nvinfer1::ScaleMode::kUNIFORM;
} else {
- // No broadcasting on Batch dimension;
- assert(dims_w.d[0] == 1);
-
- // Broadcasting on Channel dimension only allowed in kUNIFORM
- assert(dims_w.d[1] == dims_t.d[0]);
- assert(dims_w.nbDims == dims_t.nbDims);
-
- // Default is element;
- for (int i = 2; i < dims_w.nbDims; i++) {
- if (dims_w.d[i] != dims_t.d[i - 1]) {
- scale_mode = nvinfer1::ScaleMode::kCHANNEL;
- break;
+ // no broadcasting on Batch dimension;
+ VLOG(2) << "WEIGHTS DIM: " << dims_w.nbDims
+ << " tensor DIM: " << dims_t.nbDims;
+ if (dims_w.nbDims == dims_t.nbDims + 1) {
+ if (dims_w.d[0] == 1) {
+ for (int i = 1; i < dims_w.nbDims; i++) {
+ dims_w.d[i - 1] = dims_w.d[i];
+ }
+ dims_w.nbDims--;
+ } else {
+ return tensorflow::errors::InvalidArgument(
+ "Binary op cannot operate on batch, " + node_def.name());
}
}
- if (scale_mode == nvinfer1::ScaleMode::kELEMENTWISE) {
+
+ if (dims_w.nbDims == dims_t.nbDims && dims_w.d[0] == dims_t.d[0]) {
scale_mode = nvinfer1::ScaleMode::kELEMENTWISE;
- for (int i = 2; i < dims_w.nbDims; i++) {
- if (dims_w.d[i] != 1)
- return tensorflow::errors::InvalidArgument(
- "Weight shape not compatible at, " + node_def.name());
+ // default is element;
+ for (int i = 1; i < dims_w.nbDims; i++) {
+ if (dims_w.d[i] != dims_t.d[i]) {
+ // if dimension does not match, switch back to channel;
+ VLOG(2) << "channel";
+ scale_mode = nvinfer1::ScaleMode::kCHANNEL;
+ break;
+ }
}
+ // if channel as candidate, validate it
+ if (scale_mode == nvinfer1::ScaleMode::kCHANNEL) {
+ for (int i = 1; i < dims_w.nbDims; i++) {
+ if (dims_w.d[i] != 1)
+ return tensorflow::errors::InvalidArgument(
+ "Weight shape not compatible at, " + node_def.name());
+ }
+ } else {
+ VLOG(2) << "elementwise";
+ }
+ } else if (dims_w.nbDims == 1 &&
+ dims_w.d[0] == dims_t.d[dims_t.nbDims - 1]) {
+ // channel wise and broadcast required;
+ permutation_flag = true;
+ scale_mode = nvinfer1::ScaleMode::kCHANNEL;
+ } else {
+ return tensorflow::errors::InvalidArgument(
+ "Weight shape not compatible at, " + node_def.name());
}
}
- // Prepare weights
+ // transpose last dimension
+ std::vector<int> permutation(dims_t.nbDims + 1);
+ if (permutation_flag) {
+ if (scale_mode == nvinfer1::ScaleMode::kCHANNEL && dims_t.nbDims > 1) {
+ // we swap the last dimension into channel for trt.
+ // because of tensorflow default broadcasting rules.
+ for (int i = 0; i < static_cast<int>(permutation.size()); i++) {
+ permutation[i] = i;
+ }
+ permutation[1] = dims_t.nbDims;
+ permutation[dims_t.nbDims] = 1;
+ tensor = ctx.TransposeTensor(const_cast<nvinfer1::ITensor*>(tensor),
+ permutation);
+ } else {
+ return tensorflow::errors::InvalidArgument(
+ "Transpose cannot be applied, " + node_def.name());
+ }
+ }
+
+ // prepare weights
TRT_ShapedWeights shift_weights(weights.type_);
TRT_ShapedWeights scale_weights(weights.type_);
TRT_ShapedWeights power_weights(weights.type_);
@@ -779,88 +983,26 @@
scale_weights, power_weights);
nvinfer1::ITensor* output_tensor = layer->getOutput(0);
-
- // Pass the output
- outputs->push_back(TRT_TensorOrWeights(output_tensor));
- return tensorflow::Status::OK();
-}
-
-tensorflow::Status BinaryTensorOpTensor(
- Converter& ctx, const tensorflow::NodeDef& node_def,
- const nvinfer1::ITensor* tensor_l, const nvinfer1::ITensor* tensor_r,
- std::vector<TRT_TensorOrWeights>* outputs) {
- static const std::unordered_map<string, nvinfer1::ElementWiseOperation> ops{
- {"Add", nvinfer1::ElementWiseOperation::kSUM},
- {"Mul", nvinfer1::ElementWiseOperation::kPROD},
- // {"max", nvinfer1::ElementWiseOperation::kMAX},
- // {"min", nvinfer1::ElementWiseOperation::kMIN},
- {"Sub", nvinfer1::ElementWiseOperation::kSUB},
- {"Div", nvinfer1::ElementWiseOperation::kDIV},
- };
-
- // FIXME assume type matches input weights
- // Get trt type & shape
- TFAttrs attrs(node_def);
- // Maybe this part has to be moved into the block of rsqrt later
- nvinfer1::DataType dtype = attrs.get<nvinfer1::DataType>("T");
-
- // Check type consistency
- CHECK_EQ_TYPE(tensor_l->getType(), dtype);
- CHECK_EQ_TYPE(tensor_r->getType(), dtype);
- auto op_pair = ops.find(node_def.op());
- if (op_pair == ops.end())
- return tensorflow::errors::Unimplemented("binary op: " + node_def.op() +
- " not supported at: " +
- node_def.name());
-
- nvinfer1::IElementWiseLayer* layer = ctx.network()->addElementWise(
- *const_cast<nvinfer1::ITensor*>(tensor_l),
- *const_cast<nvinfer1::ITensor*>(tensor_r), op_pair->second);
-
- nvinfer1::ITensor* output_tensor = layer->getOutput(0);
-
- // Pass the output
- outputs->push_back(TRT_TensorOrWeights(output_tensor));
- return tensorflow::Status::OK();
-}
-
-tensorflow::Status ConvertPlaceholder(
- Converter& ctx, const tensorflow::NodeDef& node_def,
- std::vector<TRT_TensorOrWeights> const& inputs,
- std::vector<TRT_TensorOrWeights>* outputs) {
- VLOG(2) << "Placeholder should have been replace already";
- return tensorflow::errors::Unimplemented(", cannot convert Placeholder op");
- // OK this make sense since we are supposed to replace it with input
- TFAttrs attrs(node_def);
- nvinfer1::DataType dtype = attrs.get<nvinfer1::DataType>("dtype");
- nvinfer1::Dims dims = attrs.get<nvinfer1::Dims>("shape");
-
- dims.nbDims--;
- for (int i = 0; i < dims.nbDims; i++) dims.d[i] = dims.d[i + 1];
-
- nvinfer1::ITensor* output =
- ctx.network()->addInput(node_def.name().c_str(), dtype, dims);
- if (!output) {
- return tensorflow::errors::InvalidArgument("Failed to create Input layer");
+ // transpose back dimension
+ if (permutation_flag) {
+ output_tensor = ctx.TransposeTensor(output_tensor, permutation);
}
- outputs->push_back(TRT_TensorOrWeights(output));
+
+ // Pass the output
+ outputs->push_back(TRT_TensorOrWeights(output_tensor));
return tensorflow::Status::OK();
}
-tensorflow::Status ConvertConv2D(Converter& ctx,
- const tensorflow::NodeDef& node_def,
- const std::vector<TRT_TensorOrWeights>& inputs,
- std::vector<TRT_TensorOrWeights>* outputs) {
- nvinfer1::ITensor const* tensor = inputs.at(0).tensor();
- // TODO(jie): handle NHWC/NCHW transpose;
- TRT_ShapedWeights weights_rsck = inputs.at(1).weights();
- TRT_ShapedWeights weights = ctx.get_temp_weights_like(weights_rsck);
- ReorderRSCKToKCRS(weights_rsck, &weights);
- TRT_ShapedWeights biases(weights.type_);
- int noutput = weights.shape_.d[0];
- nvinfer1::DimsHW kernel_size;
- kernel_size.h() = weights.shape_.d[2];
- kernel_size.w() = weights.shape_.d[3];
+enum class ConvolutionType { DEFAULT, DEPTHWISE_CONV };
+
+tensorflow::Status ConvertConv2DHelper(
+ Converter& ctx, const tensorflow::NodeDef& node_def,
+ const std::vector<TRT_TensorOrWeights>& inputs,
+ std::vector<TRT_TensorOrWeights>* outputs,
+ int group // group ==0 specifies depthwise conv
+) {
+ const nvinfer1::ITensor* tensor = inputs.at(0).tensor();
+
TFAttrs attrs(node_def);
int h_index = 2;
@@ -874,11 +1016,31 @@
// TODO(jie): transpose it
}
+ // tensor after transpose (NCHW)
+ auto tensor_dim = tensor->getDimensions();
+
+ int num_groups = group;
+ if (num_groups == 0) // depthwise convolution
+ num_groups = tensor_dim.d[0];
+ VLOG(2) << "groups count: " << num_groups;
+
+ TRT_ShapedWeights weights_rsck = inputs.at(1).weights();
+ TRT_ShapedWeights weights = ctx.get_temp_weights_like(weights_rsck);
+ ReorderRSCKToKCRS(weights_rsck, &weights, num_groups);
+ TRT_ShapedWeights biases(weights.type_);
+ int noutput = weights.shape_.d[0] * num_groups;
+ nvinfer1::DimsHW kernel_size;
+ kernel_size.h() = weights.shape_.d[2];
+ kernel_size.w() = weights.shape_.d[3];
+ VLOG(2) << "kernel size: " << kernel_size.h() << ", " << kernel_size.w();
+
// TODO(jie): stride. (NHWC/NCHW)
auto tf_stride = attrs.get<std::vector<int>>("strides");
+ VLOG(2) << "h_INDEX" << h_index << ", w_index " << w_index;
+ VLOG(2) << "stride!!!: " << tf_stride[0] << tf_stride[1] << tf_stride[2]
+ << tf_stride[3];
nvinfer1::DimsHW stride(tf_stride[h_index], tf_stride[w_index]);
- auto tensor_dim = tensor->getDimensions();
std::vector<std::pair<int, int>> padding;
// TODO(jie): padding.
if (attrs.get<string>("padding") == "SAME") {
@@ -919,10 +1081,11 @@
layer->setStride(stride);
layer->setPadding({padding[0].first, padding[1].first});
layer->setName(node_def.name().c_str());
+ layer->setNbGroups(num_groups);
nvinfer1::ITensor* output_tensor = layer->getOutput(0);
auto dim_after = output_tensor->getDimensions();
- VLOG(2) << "TENSOR out: " << dim_after.d[0] << ", " << dim_after.d[1]
+ VLOG(2) << "TENSOR out: " << dim_after.d[0] << ", " << dim_after.d[1] << ", "
<< dim_after.d[2] << ", " << dim_after.d[3];
if (data_format == "NHWC") {
@@ -935,11 +1098,101 @@
return tensorflow::Status::OK();
}
+tensorflow::Status ConvertConv2DHelper(
+ Converter& ctx, const tensorflow::NodeDef& node_def,
+ const std::vector<TRT_TensorOrWeights>& inputs,
+ std::vector<TRT_TensorOrWeights>* outputs, ConvolutionType type) {
+ switch (type) {
+ case ConvolutionType::DEFAULT:
+ return ConvertConv2DHelper(ctx, node_def, inputs, outputs, 1);
+ case ConvolutionType::DEPTHWISE_CONV:
+ return ConvertConv2DHelper(ctx, node_def, inputs, outputs, 0);
+ }
+ return tensorflow::errors::Unimplemented("unsupported convolution type at, " +
+ node_def.name());
+}
+
+tensorflow::Status BinaryTensorOpTensor(
+ Converter& ctx, const tensorflow::NodeDef& node_def,
+ const nvinfer1::ITensor* tensor_l, const nvinfer1::ITensor* tensor_r,
+ std::vector<TRT_TensorOrWeights>* outputs) {
+ static const std::unordered_map<string, nvinfer1::ElementWiseOperation> ops{
+ {"Add", nvinfer1::ElementWiseOperation::kSUM},
+ {"Mul", nvinfer1::ElementWiseOperation::kPROD},
+ {"Sub", nvinfer1::ElementWiseOperation::kSUB},
+ {"Div", nvinfer1::ElementWiseOperation::kDIV},
+ };
+
+ // FIXME assume type matches input weights
+ // get trt type & shape
+ TFAttrs attrs(node_def);
+ // maybe this part has to be moved into the block of rsqrt later
+ nvinfer1::DataType dtype = attrs.get<nvinfer1::DataType>("T");
+
+ // check type consistency
+ CHECK_EQ_TYPE(tensor_l->getType(), dtype);
+ CHECK_EQ_TYPE(tensor_r->getType(), dtype);
+ auto op_pair = ops.find(node_def.op());
+ if (op_pair == ops.end())
+ return tensorflow::errors::Unimplemented("binary op: " + node_def.op() +
+ " not supported at: " +
+ node_def.name());
+
+ nvinfer1::IElementWiseLayer* layer = ctx.network()->addElementWise(
+ *const_cast<nvinfer1::ITensor*>(tensor_l),
+ *const_cast<nvinfer1::ITensor*>(tensor_r), op_pair->second);
+
+ nvinfer1::ITensor* output_tensor = layer->getOutput(0);
+
+ // pass the output
+ outputs->push_back(TRT_TensorOrWeights(output_tensor));
+ return tensorflow::Status::OK();
+}
+
+tensorflow::Status ConvertPlaceholder(
+ Converter& ctx, const tensorflow::NodeDef& node_def,
+ const std::vector<TRT_TensorOrWeights>& inputs,
+ std::vector<TRT_TensorOrWeights>* outputs) {
+ VLOG(2) << "Placeholder should have been replace already";
+ return tensorflow::errors::Unimplemented("cannot convert Placeholder op");
+ // OK this make sense since we are supposed to replace it with input
+ TFAttrs attrs(node_def);
+ nvinfer1::DataType dtype = attrs.get<nvinfer1::DataType>("dtype");
+ nvinfer1::Dims dims = attrs.get<nvinfer1::Dims>("shape");
+
+ dims.nbDims--;
+ for (int i = 0; i < dims.nbDims; i++) dims.d[i] = dims.d[i + 1];
+
+ nvinfer1::ITensor* output =
+ ctx.network()->addInput(node_def.name().c_str(), dtype, dims);
+ if (!output) {
+ return tensorflow::errors::InvalidArgument("Failed to create Input layer");
+ }
+ outputs->push_back(TRT_TensorOrWeights(output));
+ return tensorflow::Status::OK();
+}
+
+tensorflow::Status ConvertConv2D(Converter& ctx,
+ const tensorflow::NodeDef& node_def,
+ const std::vector<TRT_TensorOrWeights>& inputs,
+ std::vector<TRT_TensorOrWeights>* outputs) {
+ return ConvertConv2DHelper(ctx, node_def, inputs, outputs,
+ ConvolutionType::DEFAULT);
+}
+
+tensorflow::Status ConvertConv2DDepthwise(
+ Converter& ctx, const tensorflow::NodeDef& node_def,
+ const std::vector<TRT_TensorOrWeights>& inputs,
+ std::vector<TRT_TensorOrWeights>* outputs) {
+ return ConvertConv2DHelper(ctx, node_def, inputs, outputs,
+ ConvolutionType::DEPTHWISE_CONV);
+}
+
tensorflow::Status ConvertPool(Converter& ctx,
const tensorflow::NodeDef& node_def,
- std::vector<TRT_TensorOrWeights> const& inputs,
+ const std::vector<TRT_TensorOrWeights>& inputs,
std::vector<TRT_TensorOrWeights>* outputs) {
- nvinfer1::ITensor const* tensor = inputs.at(0).tensor();
+ const nvinfer1::ITensor* tensor = inputs.at(0).tensor();
TFAttrs attrs(node_def);
int h_index = 2;
@@ -957,6 +1210,8 @@
// TODO(jie): support other pooling type
if (node_def.op() == "MaxPool")
type = nvinfer1::PoolingType::kMAX;
+ else if (node_def.op() == "AvgPool")
+ type = nvinfer1::PoolingType::kAVERAGE;
else
return tensorflow::errors::Unimplemented("Only supports Max pool");
@@ -1019,9 +1274,9 @@
tensorflow::Status ConvertActivation(
Converter& ctx, const tensorflow::NodeDef& node_def,
- std::vector<TRT_TensorOrWeights> const& inputs,
+ const std::vector<TRT_TensorOrWeights>& inputs,
std::vector<TRT_TensorOrWeights>* outputs) {
- nvinfer1::ITensor const* tensor = inputs.at(0).tensor();
+ const nvinfer1::ITensor* tensor = inputs.at(0).tensor();
nvinfer1::IActivationLayer* layer = ctx.network()->addActivation(
*const_cast<nvinfer1::ITensor*>(tensor), nvinfer1::ActivationType::kRELU);
nvinfer1::ITensor* output_tensor = layer->getOutput(0);
@@ -1031,14 +1286,14 @@
tensorflow::Status ConvertScale(Converter& ctx,
const tensorflow::NodeDef& node_def,
- std::vector<TRT_TensorOrWeights> const& inputs,
+ const std::vector<TRT_TensorOrWeights>& inputs,
std::vector<TRT_TensorOrWeights>* outputs) {
if (inputs.size() != 2 || !inputs.at(0).is_tensor() ||
!inputs.at(1).is_weights())
return tensorflow::errors::Unimplemented(
"Only supports tensor op weight for now, at " + node_def.name());
// Implement tensor binaryOp weight [channel wise] for now;
- nvinfer1::ITensor const* tensor = inputs.at(0).tensor();
+ const nvinfer1::ITensor* tensor = inputs.at(0).tensor();
// TODO(jie): handle NHWC/NCHW transpose;
TRT_ShapedWeights weights = inputs.at(1).weights();
@@ -1055,9 +1310,26 @@
} else {
VLOG(2) << "NCHW !!!!";
}
- nvinfer1::IScaleLayer* layer = ctx.network()->addScale(
- *const_cast<nvinfer1::ITensor*>(tensor), nvinfer1::ScaleMode::kCHANNEL,
- weights, empty_weights, empty_weights);
+
+ auto dims = tensor->getDimensions();
+ VLOG(2) << "tensor dimensions: " << dims.nbDims;
+ for (int i = 0; i < dims.nbDims; i++) {
+ VLOG(2) << "i: " << dims.d[i];
+ }
+ dims = weights.shape_;
+ VLOG(2) << "tensor dimensions: " << dims.nbDims;
+ for (int i = 0; i < dims.nbDims; i++) {
+ VLOG(2) << "i: " << dims.d[i];
+ }
+
+ nvinfer1::ScaleMode mode = nvinfer1::ScaleMode::kCHANNEL;
+ if (weights.shape_.d[0] == 1) {
+ mode = nvinfer1::ScaleMode::kUNIFORM;
+ }
+
+ nvinfer1::IScaleLayer* layer =
+ ctx.network()->addScale(*const_cast<nvinfer1::ITensor*>(tensor), mode,
+ weights, empty_weights, empty_weights);
nvinfer1::ITensor* output_tensor = layer->getOutput(0);
if (data_format == "NHWC") {
@@ -1072,7 +1344,7 @@
tensorflow::Status ConvertConst(Converter& ctx,
const tensorflow::NodeDef& node_def,
- std::vector<TRT_TensorOrWeights> const& inputs,
+ const std::vector<TRT_TensorOrWeights>& inputs,
std::vector<TRT_TensorOrWeights>* outputs) {
const auto& weights_tensor = node_def.attr().at("value").tensor();
@@ -1091,20 +1363,144 @@
VLOG(2) << "SCALAR!!!" << node_def.name();
nvinfer1::Dims scalar_shape;
if (tensor.dims() > 0) {
- VLOG(2) << "Dimensions: " << tensor.dims();
- weights = TRT_ShapedWeights(dtype, weights_tensor.float_val().data(),
- GetTensorShape(tensor));
+ VLOG(2) << "dimensions: " << tensor.dims();
+ VLOG(2) << "size: " << weights_tensor.float_val_size();
+ scalar_shape = GetTensorShape(tensor);
+ for (int i = 0; i < scalar_shape.nbDims; i++)
+ VLOG(2) << scalar_shape.d[i];
+ if (GetShapeSize(scalar_shape) != weights_tensor.float_val_size()) {
+ if (weights_tensor.float_val_size() == 1 ||
+ scalar_shape.d[0] == weights_tensor.float_val_size()) {
+ scalar_shape.nbDims = 1;
+ // no dimension provided. flatten it
+ scalar_shape.d[0] = weights_tensor.float_val_size();
+ scalar_shape.type[0] = nvinfer1::DimensionType::kSPATIAL;
+ } else {
+ LOG(FATAL) << "Broadcast on weights only supports kCHANNEL and"
+ << " kUNIFORM, at: " << node_def.name();
+ }
+ }
} else {
VLOG(2) << "Dimensions: " << tensor.dims();
scalar_shape.nbDims = 1;
- scalar_shape.d[0] = 1;
+ // no dimension provided. flatten it
+ scalar_shape.d[0] = weights_tensor.float_val_size();
scalar_shape.type[0] = nvinfer1::DimensionType::kSPATIAL;
for (int i = 1; i < nvinfer1::Dims::MAX_DIMS; i++) {
scalar_shape.d[i] = 0;
scalar_shape.type[i] = nvinfer1::DimensionType::kSPATIAL;
}
- weights = TRT_ShapedWeights(dtype, weights_tensor.float_val().data(),
- scalar_shape);
+ }
+ if (ctx.isFP16()) {
+ auto dtype_new = tensorflow::DataType::DT_HALF;
+ size_t len_data = tensorflow::DataTypeSize(dtype_new);
+ for (int i = 0; i < scalar_shape.nbDims; i++)
+ len_data *= scalar_shape.d[i];
+ ctx.weight_store()->store_.push_back(std::vector<uint8_t>(len_data));
+ void* dst = static_cast<void*>(&(ctx.weight_store()->store_.back()[0]));
+ tensorflow::Tensor temp_tensor(tensorflow::DT_HALF, tensor.shape());
+ auto half_tensor = temp_tensor.flat<Eigen::half>();
+ Eigen::DefaultDevice defd;
+ half_tensor.device(defd) =
+ tensor.flat<float>().template cast<Eigen::half>();
+ memcpy(dst, half_tensor.data(), len_data); // store into weight store
+ weights = TRT_ShapedWeights(dtype_new, dst, scalar_shape);
+ } else {
+ size_t len_data = tensorflow::DataTypeSize(dtype);
+ for (int i = 0; i < scalar_shape.nbDims; i++)
+ len_data *= scalar_shape.d[i];
+ ctx.weight_store()->store_.push_back(std::vector<uint8_t>(len_data));
+ void* dst = static_cast<void*>(&(ctx.weight_store()->store_.back()[0]));
+ std::vector<float> tensor_data(
+ weights_tensor.float_val().begin(),
+ weights_tensor.float_val()
+ .end()); // make a local copy first to flatten
+ memcpy(dst, tensor_data.data(), len_data); // store into weight store
+ weights = TRT_ShapedWeights(dtype, dst, scalar_shape);
+ }
+ } else if (!weights_tensor.int_val().empty()) {
+ VLOG(2) << "int!!!" << node_def.name();
+ nvinfer1::Dims scalar_shape;
+ if (tensor.dims() > 0) {
+ VLOG(2) << "dimensions: " << tensor.dims();
+ scalar_shape = GetTensorShape(tensor);
+ if (GetShapeSize(scalar_shape) != weights_tensor.int_val_size()) {
+ if (weights_tensor.int_val_size() == 1 ||
+ scalar_shape.d[0] == weights_tensor.int_val_size()) {
+ scalar_shape.nbDims = 1;
+ // no dimension provided. flatten it
+ scalar_shape.d[0] = weights_tensor.int_val_size();
+ scalar_shape.type[0] = nvinfer1::DimensionType::kSPATIAL;
+ } else {
+ LOG(FATAL) << "Broadcast on weights only supports kCHANNEL and"
+ << " kUNIFORM, at: " << node_def.name();
+ }
+ }
+ } else {
+ VLOG(2) << "dimensions: " << tensor.dims();
+ scalar_shape.nbDims = 1;
+ // no dimension provided. flatten it
+ scalar_shape.d[0] = weights_tensor.int_val_size();
+ scalar_shape.type[0] = nvinfer1::DimensionType::kSPATIAL;
+ for (int i = 1; i < nvinfer1::Dims::MAX_DIMS; i++) {
+ scalar_shape.d[i] = 0;
+ scalar_shape.type[i] = nvinfer1::DimensionType::kSPATIAL;
+ }
+ }
+ if (ctx.isFP16()) {
+ auto dtype_new = tensorflow::DataType::DT_HALF;
+ size_t len_data = tensorflow::DataTypeSize(dtype_new);
+ for (int i = 0; i < scalar_shape.nbDims; i++)
+ len_data *= scalar_shape.d[i];
+ ctx.weight_store()->store_.push_back(std::vector<uint8_t>(len_data));
+ void* dst = static_cast<void*>(&(ctx.weight_store()->store_.back()[0]));
+ tensorflow::Tensor temp_tensor(tensorflow::DT_HALF, tensor.shape());
+ TTypes<Eigen::half>::Flat half_tensor = temp_tensor.flat<Eigen::half>();
+ Eigen::DefaultDevice defd;
+ switch (dtype) {
+ case (tensorflow::DT_INT32): {
+ half_tensor.device(defd) =
+ tensor.flat<int32>().template cast<Eigen::half>();
+ break;
+ }
+ case (tensorflow::DT_INT16): {
+ half_tensor.device(defd) =
+ tensor.flat<int16>().template cast<Eigen::half>();
+ break;
+ }
+ case (tensorflow::DT_INT8): {
+ half_tensor.device(defd) =
+ tensor.flat<int8>().template cast<Eigen::half>();
+ break;
+ }
+ case (tensorflow::DT_UINT8): {
+ half_tensor.device(defd) =
+ tensor.flat<uint8>().template cast<Eigen::half>();
+ break;
+ }
+ default:
+ return tensorflow::errors::InvalidArgument(
+ "Datatype " + tensorflow::DataTypeString(dtype) +
+ " for FP16 conversion");
+ break;
+ };
+ memcpy(dst, half_tensor.data(), len_data); // store into weight store
+ weights = TRT_ShapedWeights(dtype_new, dst, scalar_shape);
+ } else {
+ size_t len_data = tensorflow::DataTypeSize(dtype);
+ for (int i = 0; i < scalar_shape.nbDims; i++)
+ len_data *= scalar_shape.d[i];
+ size_t len_tensor = weights_tensor.int_val_size() * sizeof(int32);
+ len_data = std::max(len_data, len_tensor);
+ ctx.weight_store()->store_.push_back(std::vector<uint8_t>(len_data));
+ void* dst = static_cast<void*>(&(ctx.weight_store()->store_.back()[0]));
+ std::vector<int32> tensor_data(
+ weights_tensor.int_val().begin(),
+ weights_tensor.int_val()
+ .end()); // make a local copy first to flatten
+ // doesn't have to be contigous
+ memcpy(dst, tensor_data.data(), len_tensor); // store into weight store
+ weights = TRT_ShapedWeights(dtype, dst, scalar_shape);
}
} else if (!weights_tensor.tensor_content().empty()) {
VLOG(2) << "TENSOR!!!" << node_def.name();
@@ -1130,7 +1526,7 @@
tensorflow::Status ConvertIdentity(
Converter& ctx, const tensorflow::NodeDef& node_def,
- std::vector<TRT_TensorOrWeights> const& inputs,
+ const std::vector<TRT_TensorOrWeights>& inputs,
std::vector<TRT_TensorOrWeights>* outputs) {
outputs->push_back(inputs.at(0));
return tensorflow::Status::OK();
@@ -1138,7 +1534,7 @@
tensorflow::Status ConvertBinary(Converter& ctx,
const tensorflow::NodeDef& node_def,
- std::vector<TRT_TensorOrWeights> const& inputs,
+ const std::vector<TRT_TensorOrWeights>& inputs,
std::vector<TRT_TensorOrWeights>* outputs) {
if (inputs.size() != 2)
return tensorflow::errors::FailedPrecondition(
@@ -1165,7 +1561,7 @@
tensorflow::Status ConvertUnary(Converter& ctx,
const tensorflow::NodeDef& node_def,
- std::vector<TRT_TensorOrWeights> const& inputs,
+ const std::vector<TRT_TensorOrWeights>& inputs,
std::vector<TRT_TensorOrWeights>* outputs) {
if (inputs.size() != 1)
return tensorflow::errors::FailedPrecondition(
@@ -1183,7 +1579,7 @@
tensorflow::Status ConvertReduce(Converter& ctx,
const tensorflow::NodeDef& node_def,
- std::vector<TRT_TensorOrWeights> const& inputs,
+ const std::vector<TRT_TensorOrWeights>& inputs,
std::vector<TRT_TensorOrWeights>* outputs) {
if (inputs.size() != 2 || !inputs.at(0).is_tensor() ||
!inputs.at(1).is_weights())
@@ -1191,7 +1587,7 @@
"Input expects tensor and weights, at" + node_def.name());
// Implement tensor binaryOp weight [channel wise] for now;
- nvinfer1::ITensor const* tensor = inputs.at(0).tensor();
+ const nvinfer1::ITensor* tensor = inputs.at(0).tensor();
auto dims = tensor->getDimensions();
// Restore implicit batch dimension
int nb_dims = dims.nbDims + 1;
@@ -1229,6 +1625,7 @@
return tensorflow::errors::InvalidArgument("TRT cannot reduce at 0, at" +
node_def.name());
if (index_list_data[i] == 1) permuted_index = 1;
+
idx_set.emplace(index_list_data[i]);
}
@@ -1236,7 +1633,7 @@
nvinfer1::DimsHW pool_kernel;
if (permuted_index == 1) {
for (int i = 2; i < nb_dims; i++) {
- if (idx_set.count(i)) {
+ if (idx_set.count(i) == 0) {
permuted_index = i;
break;
}
@@ -1271,12 +1668,13 @@
output_tensor = ctx.TransposeTensor(
const_cast<nvinfer1::ITensor*>(output_tensor), permutation_order);
}
+ outputs->push_back(TRT_TensorOrWeights(output_tensor));
return tensorflow::Status::OK();
}
tensorflow::Status ConvertPad(Converter& ctx,
const tensorflow::NodeDef& node_def,
- std::vector<TRT_TensorOrWeights> const& inputs,
+ const std::vector<TRT_TensorOrWeights>& inputs,
std::vector<TRT_TensorOrWeights>* outputs) {
if (inputs.size() != 2 || !inputs.at(0).is_tensor() ||
!inputs.at(1).is_weights())
@@ -1284,7 +1682,7 @@
"Input expects tensor and weights, at" + node_def.name());
// Implement tensor binaryOp weight [channel wise] for now;
- nvinfer1::ITensor const* tensor = inputs.at(0).tensor();
+ const nvinfer1::ITensor* tensor = inputs.at(0).tensor();
auto dims = tensor->getDimensions();
// Restore implicit batch dimension
int nb_dims = dims.nbDims + 1;
@@ -1371,19 +1769,287 @@
return tensorflow::Status::OK();
}
+tensorflow::Status ConvertConcat(Converter& ctx,
+ const tensorflow::NodeDef& node_def,
+ const std::vector<TRT_TensorOrWeights>& inputs,
+ std::vector<TRT_TensorOrWeights>* outputs) {
+ // not including the last input (axis) here
+ int input_size = static_cast<int>(inputs.size()) - 1;
+
+ if (!inputs.at(0).is_tensor())
+ return tensorflow::errors::InvalidArgument(
+ "Concat in TRT support only Tensor input, at " + node_def.name());
+
+ // We are retrieving the axis
+ TRT_ShapedWeights axis = inputs.at(input_size).weights();
+
+ TFAttrs attrs(node_def);
+ // auto attr_size = attrs.at("N")->i();
+ // auto data_type = attrs.get<nvinfer1::DataType>("T");
+ auto index_type = attrs.get<tensorflow::DataType>("Tidx");
+
+ // TODO(jie): handle data type
+ // Only expect to handle INT32 as index attributes for now
+ if (index_type != tensorflow::DataType::DT_INT32)
+ return tensorflow::errors::Unimplemented(
+ "Tidx supports only DT_INT32, at " + node_def.name());
+
+ int index = *(static_cast<int*>(const_cast<void*>(axis.GetValues())));
+
+ // TODO(jie): early termination with no-op (attr_size==1)
+
+ auto dim = inputs.at(0).tensor()->getDimensions();
+ // dimension check
+ if (index > dim.nbDims + 1)
+ return tensorflow::errors::InvalidArgument(
+ "Concatenate on axis out of dimension range, at " + node_def.name());
+
+ if (index == 0)
+ return tensorflow::errors::InvalidArgument(
+ "Concatenate on batch dimension not supported, at " + node_def.name());
+
+ // incase we need permutation;
+ std::vector<int> permutation_order(dim.nbDims + 1);
+
+ for (int i = 0; i < dim.nbDims + 1; i++) permutation_order[i] = i;
+
+ if (index != 1) {
+ permutation_order[1] = index - 1;
+ permutation_order[index - 1] = 1;
+ }
+
+ std::vector<nvinfer1::ITensor const*> inputs_vec;
+ // Shap chack (all input tensor should have same shape)
+ // starting from 0 since we are probably also doing transpose here;
+ for (int i = 0; i < input_size; i++) {
+ auto tensor_i = inputs.at(i).tensor();
+ auto dim_i = tensor_i->getDimensions();
+ if (dim_i.nbDims != dim.nbDims)
+ return tensorflow::errors::InvalidArgument(
+ "Concatenate receives inputs with inconsistent dimensions, at " +
+ node_def.name());
+
+ for (int j = 0; j < dim.nbDims; j++) {
+ // check dimension consistency on non-concatenate axis
+ if (j != index - 1 && dim_i.d[j] != dim.d[j])
+ return tensorflow::errors::InvalidArgument(
+ "Concatenate receives inputs with inconsistent shape, at" +
+ node_def.name());
+ }
+
+ // TRT does concatenation only on channel!
+ if (index != 1)
+ tensor_i = ctx.TransposeTensor(const_cast<nvinfer1::ITensor*>(tensor_i),
+ permutation_order);
+
+ inputs_vec.push_back(tensor_i);
+ }
+
+ // nvinfer1::ITensor const* tensor = inputs.at(0).tensor();
+ nvinfer1::IConcatenationLayer* layer = ctx.network()->addConcatenation(
+ const_cast<nvinfer1::ITensor* const*>(inputs_vec.data()),
+ inputs_vec.size());
+ nvinfer1::ITensor* output_tensor = layer->getOutput(0);
+
+ if (index != 1) {
+ output_tensor = ctx.TransposeTensor(output_tensor, permutation_order);
+ }
+ outputs->push_back(TRT_TensorOrWeights(output_tensor));
+ return tensorflow::Status::OK();
+}
+
+tensorflow::Status ConvertFusedBatchNorm(
+ Converter& ctx, const tensorflow::NodeDef& node_def,
+ const std::vector<TRT_TensorOrWeights>& inputs,
+ std::vector<TRT_TensorOrWeights>* outputs) {
+ TFAttrs attrs(node_def);
+ float epsilon = attrs.get<float>("epsilon");
+ auto data_format = attrs.get<string>("data_format");
+ if (data_format != "NCHW") {
+ return tensorflow::errors::Unimplemented(
+ "only data_format=NCHW is supported, at " + node_def.name());
+ }
+ bool is_training = attrs.get<bool>("is_training");
+ if (is_training) {
+ return tensorflow::errors::Unimplemented(
+ "only is_training=false is supported, at " + node_def.name());
+ }
+ nvinfer1::ITensor const* tensor = inputs.at(0).tensor();
+ TRT_ShapedWeights scale_weights = inputs.at(1).weights();
+ TRT_ShapedWeights offset_weights = inputs.at(2).weights();
+ TRT_ShapedWeights mean_weights = inputs.at(3).weights();
+ TRT_ShapedWeights variance_weights = inputs.at(4).weights();
+ TRT_ShapedWeights dummy_power_weights(scale_weights.type_);
+ TRT_ShapedWeights combined_scale_weights =
+ ctx.get_temp_weights_like(scale_weights);
+ TRT_ShapedWeights combined_offset_weights =
+ ctx.get_temp_weights_like(offset_weights);
+ size_t nweight = scale_weights.count();
+ if ((scale_weights.type_ == offset_weights.type_) &&
+ (mean_weights.type_ == variance_weights.type_) &&
+ (scale_weights.type_ == variance_weights.type_)) {
+ if ((scale_weights.type_ != tensorflow::DataType::DT_FLOAT) &&
+ (scale_weights.type_ != tensorflow::DataType::DT_HALF)) {
+ return tensorflow::errors::Unimplemented(
+ "only float32 or float16 weight data type is supported, for node " +
+ node_def.name() + " got " +
+ tensorflow::DataTypeString(scale_weights.type_));
+ }
+ if (scale_weights.type_ == tensorflow::DT_FLOAT) {
+ for (size_t i = 0; i < nweight; ++i) {
+ float scale = (static_cast<float const*>(scale_weights.GetValues()))[i];
+ float offset =
+ (static_cast<float const*>(offset_weights.GetValues()))[i];
+ float mean = (static_cast<float const*>(mean_weights.GetValues()))[i];
+ float variance =
+ (static_cast<float const*>(variance_weights.GetValues()))[i];
+ float& combined_scale_ref = const_cast<float*>(
+ static_cast<float const*>(combined_scale_weights.GetValues()))[i];
+ float& combined_offset_ref = const_cast<float*>(
+ static_cast<float const*>(combined_offset_weights.GetValues()))[i];
+ combined_scale_ref = scale / sqrtf(variance + epsilon);
+ combined_offset_ref = offset - mean * combined_scale_ref;
+ }
+ } else {
+ const Eigen::half* scale_vals =
+ (static_cast<Eigen::half const*>(scale_weights.GetValues()));
+ const Eigen::half* off_vals =
+ (static_cast<Eigen::half const*>(offset_weights.GetValues()));
+ const Eigen::half* mean_vals =
+ (static_cast<Eigen::half const*>(mean_weights.GetValues()));
+ const Eigen::half* variance_vals =
+ (static_cast<Eigen::half const*>(variance_weights.GetValues()));
+ Eigen::half* comb_scale_vals = const_cast<Eigen::half*>(
+ static_cast<Eigen::half const*>(combined_scale_weights.GetValues()));
+ Eigen::half* comb_off_vals = const_cast<Eigen::half*>(
+ static_cast<Eigen::half const*>(combined_offset_weights.GetValues()));
+ for (size_t i = 0; i < nweight; ++i) {
+ float scale(scale_vals[i]);
+ float offset(off_vals[i]);
+ float mean(mean_vals[i]);
+ float variance(variance_vals[i]);
+ float combined_scale_ref = scale / sqrtf(variance + epsilon);
+ comb_scale_vals[i] = Eigen::half(combined_scale_ref);
+ float combined_offset_ref = offset - mean * combined_scale_ref;
+ comb_off_vals[i] = Eigen::half(combined_offset_ref);
+ }
+ }
+ }
+ nvinfer1::IScaleLayer* layer = ctx.network()->addScale(
+ *const_cast<nvinfer1::ITensor*>(tensor), nvinfer1::ScaleMode::kCHANNEL,
+ combined_offset_weights.GetWeightsForTRT(),
+ combined_scale_weights.GetWeightsForTRT(),
+ dummy_power_weights.GetWeightsForTRT());
+ nvinfer1::ITensor* output_tensor = layer->getOutput(0);
+ outputs->push_back(TRT_TensorOrWeights(output_tensor));
+ return tensorflow::Status::OK();
+}
+
+tensorflow::Status ConvertMatMul(Converter& ctx,
+ const tensorflow::NodeDef& node_def,
+ const std::vector<TRT_TensorOrWeights>& inputs,
+ std::vector<TRT_TensorOrWeights>* outputs) {
+ const nvinfer1::ITensor* tensor = inputs.at(0).tensor();
+
+ // TODO(jie): transpose!
+ TFAttrs attrs(node_def);
+
+ TRT_ShapedWeights weights_ck = inputs.at(1).weights();
+ TRT_ShapedWeights weights = ctx.get_temp_weights_like(weights_ck);
+ ReorderCKtoKC(weights_ck, &weights);
+ TRT_ShapedWeights biases(weights.type_);
+
+ int noutput = weights.shape_.d[0];
+
+ nvinfer1::IFullyConnectedLayer* layer = ctx.network()->addFullyConnected(
+ *const_cast<nvinfer1::ITensor*>(tensor), noutput, weights, biases);
+
+ nvinfer1::ITensor* output_tensor = layer->getOutput(0);
+ outputs->push_back(TRT_TensorOrWeights(output_tensor));
+ return tensorflow::Status::OK();
+}
+
+tensorflow::Status ConvertReshape(
+ Converter& ctx, const tensorflow::NodeDef& node_def,
+ const std::vector<TRT_TensorOrWeights>& inputs,
+ std::vector<TRT_TensorOrWeights>* outputs) {
+ if (inputs.size() != 2 || !inputs.at(0).is_tensor() ||
+ !inputs.at(1).is_weights())
+ return tensorflow::errors::InvalidArgument(
+ "Input expects tensor and weights, at" + node_def.name());
+
+ // implement tensor binaryOp weight [channel wise] for now;
+ const nvinfer1::ITensor* tensor = inputs.at(0).tensor();
+ auto dims = tensor->getDimensions();
+ // restore implicit batch dimension
+
+ TRT_ShapedWeights shape = inputs.at(1).weights();
+
+ TFAttrs attrs(node_def);
+
+ auto padding_type = attrs.get<tensorflow::DataType>("Tshape");
+
+ if (shape.shape_.nbDims != 1)
+ return tensorflow::errors::InvalidArgument(
+ "reshape new shape is not 1 dimensional, at " + node_def.name());
+
+ // Only expect to handle INT32 as attributes for now
+ if (padding_type != tensorflow::DataType::DT_INT32)
+ return tensorflow::errors::Unimplemented(
+ "reshape new shape supports only DT_INT32, at " + node_def.name());
+
+ auto shape_data = static_cast<int*>(const_cast<void*>(shape.GetValues()));
+
+ if (shape_data[0] != -1)
+ return tensorflow::errors::InvalidArgument(
+ "reshape new shape first dimension is not -1, at " + node_def.name());
+
+ auto shape_num_dims = shape.shape_.d[0];
+ VLOG(2) << "shape dimensions: " << shape_num_dims;
+ int volume_w = 1;
+ for (int i = 1; i < shape.shape_.d[0]; i++) volume_w *= shape_data[i];
+
+ int volume_t = 1;
+ for (int i = 0; i < dims.nbDims; i++) volume_t *= dims.d[i];
+
+ VLOG(2) << "volume: " << volume_t << " volume weights: " << volume_w;
+ if (volume_w != volume_t)
+ return tensorflow::errors::InvalidArgument(
+ "volume does not agree between tensor and new shape, at " +
+ node_def.name());
+
+ nvinfer1::IShuffleLayer* layer =
+ ctx.network()->addShuffle(*const_cast<nvinfer1::ITensor*>(tensor));
+
+ nvinfer1::Dims reshape_dims;
+ VLOG(2) << "new dimension: " << shape_num_dims - 1;
+ reshape_dims.nbDims = shape_num_dims - 1;
+ for (int32_t i = 0; i < reshape_dims.nbDims; ++i) {
+ reshape_dims.d[i] = shape_data[i + 1];
+ }
+ layer->setReshapeDimensions(reshape_dims);
+ VLOG(2) << "new dimension: " << shape_num_dims - 1;
+
+ nvinfer1::ITensor* output_tensor = layer->getOutput(0);
+ auto dims_output = output_tensor->getDimensions();
+ VLOG(2) << "output tensor dimension:" << dims_output.nbDims;
+ outputs->push_back(TRT_TensorOrWeights(output_tensor));
+ return tensorflow::Status::OK();
+}
+
void Converter::register_op_converters() {
// vgg_16 slim implementation
op_registry_["Placeholder"] = ConvertPlaceholder;
op_registry_["Conv2D"] = ConvertConv2D;
+ op_registry_["DepthwiseConv2dNative"] = ConvertConv2DDepthwise;
op_registry_["Relu"] = ConvertActivation;
op_registry_["MaxPool"] = ConvertPool;
+ op_registry_["AvgPool"] = ConvertPool;
// This could be really handled as ConvertBinary
op_registry_["BiasAdd"] = ConvertScale;
op_registry_["Const"] = ConvertConst;
- // op_registry_["MatMul"] = ConvertFullyConnected; // Not used in vgg
// TODO(ben,jie): this is a temp hack.
op_registry_["Identity"] = ConvertIdentity; // Identity should be removed
- // op_registry_["AvgPool"] = ConvertPool;
// resnet_50_v1 slim implementation
op_registry_["Add"] = ConvertBinary;
@@ -1393,26 +2059,364 @@
op_registry_["Mean"] = ConvertReduce;
op_registry_["Pad"] = ConvertPad;
// TODO(ben,jie): Add more ops
+
+ op_registry_["ConcatV2"] = ConvertConcat;
+ op_registry_["MatMul"] = ConvertMatMul;
+ op_registry_["Reshape"] = ConvertReshape;
+ op_registry_["FusedBatchNorm"] = ConvertFusedBatchNorm;
+ op_registry_["FusedBatchNormV2"] = ConvertFusedBatchNorm;
}
} // namespace
+tensorflow::Status GetTensorRTGraph(tensorrt::convert::SubGraphParams& s) {
+ return tensorflow::errors::Unimplemented("Not implemented yet");
+}
+tensorflow::Status ConvertCalibrationNodeToEngineNode(
+ tensorflow::Graph& graph, tensorflow::Node* c_node) {
+ const auto ndef = c_node->def();
-tensorflow::Status ConvertSubGraphToTensorRTNodeDef(
- const tensorflow::Graph& graph, const std::set<int>& subgraph_node_ids,
- const std::vector<std::pair<int, int>>& input_inds,
- const std::vector<std::pair<int, int>>& output_inds, size_t max_batch_size,
- size_t max_workspace_size_bytes,
- const tensorflow::grappler::GraphProperties& graph_properties,
- tensorflow::NodeDef* trt_node) {
+ TFAttrs attrs(ndef);
+ std::vector<string> segment_nodes(
+ attrs.get<std::vector<string>>("segment_nodes"));
+ std::vector<string> output_nodes(
+ attrs.get<std::vector<string>>("segment_output_names"));
+ std::vector<string> input_names(
+ attrs.get<std::vector<string>>("input_names"));
+ string res_name = attrs.get<string>("resource_name");
+ VLOG(1) << "Node name " << c_node->name() << " res_name " << res_name;
+ string engine_name = "my_trt_op";
+ {
+ const auto node_id = tensorflow::str_util::Split(res_name, "_");
+ engine_name += node_id.back();
+ }
+ std::map<string, tensorflow::Node*> node_maps;
+
+ for (auto n : graph.op_nodes()) {
+ node_maps.insert({n->name(), n});
+ }
+ VLOG(1) << "Output Nodes:";
+ std::vector<tensorflow::DataType> out_types;
+ std::vector<const tensorflow::Edge*> out_edges;
+ for (auto& i : output_nodes) {
+ auto node_port = tensorflow::str_util::Split(i, ":");
+ VLOG(1) << " " << i << " in graph " << node_maps.count(i);
+ auto out_node_name = node_port.at(0);
+ if (node_port.size() > 1) {
+ VLOG(1) << "Multi port output" << node_port.at(0) << " "
+ << node_port.at(1) << " size=" << node_port.size();
+ }
+ auto node_it = node_maps.find(out_node_name);
+ if (node_it != node_maps.end()) {
+ tensorflow::Node* out_node = node_it->second;
+ int port = 0;
+ if (node_port.size() == 2) {
+ port = std::strtoul(node_port.at(1).c_str(), nullptr, 10);
+ out_types.push_back(out_node->output_type(port));
+ } else {
+ out_types.push_back(out_node->output_type(0));
+ }
+ for (auto out_edge : out_node->out_edges()) {
+ if (out_edge->src_output() == port) {
+ out_edges.push_back(out_edge);
+ break;
+ }
+ }
+ } else {
+ LOG(WARNING) << " couldn't find output node " << out_node_name;
+ }
+ }
+ VLOG(1) << "Input Nodes:";
+ for (auto& i : input_names) {
+ VLOG(1) << " " << i << " in graph " << node_maps.count(i);
+ }
+ auto trt_rm = tensorflow::tensorrt::TRTResourceManager::instance();
+ auto resmgr = trt_rm->getManager("TRTCalibOps");
+ tensorflow::tensorrt::TRTCalibrationResource* calib_res = nullptr;
+ auto status = resmgr->Lookup(res_name, res_name, &calib_res);
+ if (!status.ok() || !calib_res->calibrator_) {
+ return tensorflow::errors::FailedPrecondition(
+ "You must run calibration"
+ " and inference conversion in the same proces");
+ }
+
+ calib_res->calibrator_->setDone();
+ calib_res->thr_->join();
+ delete calib_res->thr_;
+ if (!calib_res->engine_) {
+ LOG(FATAL) << "Calibration failed!, engine is nullptr. Did you run "
+ "calibration graph?";
+ }
+ auto weight_rmgr = trt_rm->getManager("WeightStore");
+ TF_CHECK_OK(weight_rmgr->Delete<tensorflow::tensorrt::TRTWeightStore>(
+ res_name, res_name));
+ auto engine_plan = calib_res->engine_->serialize();
+ calib_res->engine_->destroy();
+ calib_res->network_->destroy();
+ calib_res->builder_->destroy();
+ calib_res->thr_ = nullptr;
+ calib_res->engine_ = nullptr;
+ calib_res->builder_ = nullptr;
+ tensorflow::NodeDefBuilder op_builder(engine_name, "TRTEngineOp");
+ std::vector<tensorflow::NodeDefBuilder::NodeOut> income_edges;
+ for (const auto in_edge : c_node->in_edges()) {
+ auto src = in_edge->src();
+ int dest_port = in_edge->dst_input();
+ income_edges.emplace_back(src->name(), in_edge->src_output(),
+ c_node->input_type(dest_port));
+ }
+ tensorflow::gtl::ArraySlice<tensorflow::NodeDefBuilder::NodeOut> input_list(
+ income_edges);
+ op_builder.Input(input_list);
+ tensorflow::NodeDef engine_node;
+ const char* engine_plan_data = static_cast<const char*>(engine_plan->data());
+ string engine_plan_string(engine_plan_data,
+ engine_plan_data + engine_plan->size());
+ status = op_builder.Attr("serialized_engine", engine_plan_string)
+ .Attr("input_nodes", input_names)
+ .Attr("output_nodes", output_nodes)
+ .Attr("OutT", out_types)
+ .Finalize(&engine_node);
+ if (!status.ok()) {
+ LOG(ERROR) << "Engine Node creation failed";
+ return status;
+ }
+ auto trt_engine_node = graph.AddNode(engine_node, &status);
+ TF_CHECK_OK(status);
+ for (size_t i = 0; i < out_edges.size(); i++) {
+ VLOG(1) << "Connecting trt_engine_node output " << i << " with "
+ << out_edges.at(i)->dst()->name() << " port "
+ << out_edges.at(i)->dst_input();
+ TF_RETURN_IF_ERROR(graph.UpdateEdge(trt_engine_node, i,
+ out_edges.at(i)->dst(),
+ out_edges.at(i)->dst_input()));
+ }
+ VLOG(1) << "Segment nodes:";
+ for (auto& i : segment_nodes) {
+ VLOG(1) << " " << i << " in graph " << node_maps.count(i);
+ auto it = node_maps.find(i);
+ if (it != node_maps.end()) {
+ graph.RemoveNode(it->second);
+ }
+ }
+ graph.RemoveNode(c_node);
+ return tensorflow::Status::OK();
+}
+
+tensorflow::Status InjectCalibrationNode(tensorrt::convert::SubGraphParams& s) {
// Visit nodes in reverse topological order and construct the TRT network.
// Toposort
std::vector<tensorflow::Node*> order_vec;
- tensorflow::GetPostOrder(graph, &order_vec);
+ tensorflow::GetPostOrder(s.graph, &order_vec);
// Select just the subgraph
std::list<tensorflow::Node*> order;
for (tensorflow::Node* node : order_vec) {
- if (subgraph_node_ids.count(node->id())) {
+ if (s.subgraph_node_ids.count(node->id())) {
+ order.push_front(node); // we want topological order to contstruct the
+ // network layer by layer
+ }
+ }
+ // topological order is needed to build TRT network
+ static int static_id = 0;
+ string subgraph_name_scope;
+ if (!order.empty()) {
+ subgraph_name_scope = order.front()->name();
+ }
+ for (const tensorflow::Node* node : order) {
+ subgraph_name_scope = GetCommonNameScope(subgraph_name_scope, node->name());
+ }
+ // TODO(sami,ben,jie): proper naming!
+ string calib_op_name =
+ StrCat(subgraph_name_scope, "my_trt_calib_op_", static_id);
+ string engine_name = StrCat(subgraph_name_scope, "my_trt_op", static_id);
+ static_id++;
+ auto trt_rmgr = tensorflow::tensorrt::TRTResourceManager::instance();
+ auto op_rmgr = trt_rmgr->getManager("TRTCalibOps");
+ auto op_res = new tensorflow::tensorrt::TRTCalibrationResource();
+ TF_CHECK_OK(op_rmgr->Create(calib_op_name, calib_op_name, op_res));
+ op_res->logger_ = new tensorflow::tensorrt::Logger();
+ op_res->builder_ = nvinfer1::createInferBuilder(*(op_res->logger_));
+
+ if (!op_res->builder_) {
+ return tensorflow::errors::Internal(
+ "failed to create TensorRT builder object");
+ }
+
+ op_res->network_ = op_res->builder_->createNetwork();
+ if (!op_res->network_) {
+ return tensorflow::errors::Internal(
+ "failed to create TensorRT network object");
+ }
+
+ // Build the network
+ auto weight_rmgr = trt_rmgr->getManager("WeightStore");
+ auto ws = new tensorflow::tensorrt::TRTWeightStore();
+ TF_CHECK_OK(weight_rmgr->Create(calib_op_name, calib_op_name, ws));
+ Converter converter(op_res->network_, ws, s.precision_mode == FP16MODE);
+ std::vector<string> input_names;
+ std::vector<tensorflow::DataType> input_dtypes;
+ for (const std::pair<int, int>& input : s.input_inds) {
+ VLOG(2) << "parsing input. Node id= " << input.first;
+ int node_id = input.first;
+ int output_idx = input.second;
+ tensorflow::Node* node = s.graph.FindNodeId(node_id);
+ auto node_name = node->name();
+ input_names.push_back(node_name); // insert original node name without port
+ // TODO(jie): alternative :)
+ if (!s.graph_properties.HasOutputProperties(node_name))
+ return tensorflow::errors::Internal("failed to find input node: " +
+ node_name);
+
+ auto op_info_vec = s.graph_properties.GetOutputProperties(node_name);
+ if (static_cast<int>(op_info_vec.size()) < output_idx)
+ return tensorflow::errors::Internal(
+ "accessing output index of: ", output_idx, ", at node: ", node_name,
+ "with output entry from shape_map: ", op_info_vec.size());
+
+ auto op_info = op_info_vec.at(output_idx);
+
+ tensorflow::DataType tf_dtype = op_info.dtype();
+ input_dtypes.push_back(tf_dtype);
+
+ nvinfer1::DataType dtype(nvinfer1::DataType::kFLOAT);
+ TF_CHECK_OK(ConvertDType(tf_dtype, &dtype));
+
+ VLOG(2) << "accessing output index of: " << output_idx
+ << ", at node: " << node_name
+ << "with output entry from shape_map: " << op_info_vec.size();
+
+ // TODO(ben,jie): update TRT input format/dimension
+ nvinfer1::DimsCHW input_dim_psuedo_chw;
+ for (int i = 0; i < 3; i++) input_dim_psuedo_chw.d[i] = 1;
+
+ for (int i = 1; i < op_info.shape().dim_size(); i++) {
+ VLOG(2) << "dimension: " << i
+ << " , size: " << op_info.shape().dim(i).size();
+ input_dim_psuedo_chw.d[i - 1] = op_info.shape().dim(i).size();
+ }
+
+ // TODO(ben,jie): proper way to restore input tensor name?
+ auto input_tensor_name = node_name;
+ if (output_idx != 0) input_tensor_name = StrCat(node_name, ":", output_idx);
+
+ nvinfer1::ITensor* input_tensor = converter.network()->addInput(
+ input_tensor_name.c_str(), dtype, input_dim_psuedo_chw);
+
+ if (!input_tensor)
+ return tensorflow::errors::InvalidArgument(
+ "Failed to create Input layer");
+ VLOG(2) << "input tensor name :" << input_tensor_name;
+
+ if (!converter.insert_input_tensor(input_tensor_name, input_tensor))
+ return tensorflow::errors::AlreadyExists(
+ "output tensor already exists for op: " + input_tensor_name);
+ }
+
+ VLOG(2) << "finished sorting";
+
+ for (const tensorflow::Node* node : order) {
+ const tensorflow::NodeDef& node_def = node->def();
+ VLOG(2) << "converting node: " << node_def.name() << " , " << node_def.op();
+ TF_RETURN_IF_ERROR(converter.convert_node(node_def));
+ }
+
+ VLOG(2) << "finished conversion";
+
+ // Gather output metadata
+ std::vector<string> output_names;
+ std::vector<tensorflow::DataType> output_dtypes;
+ int trt_engine_op_output_idx = 0;
+ for (const std::pair<int, int>& output : s.output_inds) {
+ int node_id = output.first;
+ int output_idx = output.second;
+ tensorflow::Node* node = s.graph.FindNodeId(node_id);
+ string op_name = node->name();
+ string tensor_name = op_name;
+
+ s.output_edge_map->insert(
+ {trt_engine_op_output_idx == 0
+ ? engine_name
+ : StrCat(engine_name, ":", trt_engine_op_output_idx),
+ {output_idx, tensor_name}});
+ trt_engine_op_output_idx++;
+ if (output_idx != 0) {
+ tensor_name = StrCat(tensor_name, ":", output_idx);
+ }
+ VLOG(1) << "output tensor name: " << tensor_name;
+ output_names.push_back(tensor_name);
+ auto tensor_or_weights = converter.get_tensor(tensor_name);
+ if (!tensor_or_weights.is_tensor()) {
+ return tensorflow::errors::InvalidArgument(
+ "Output node is weights not tensor");
+ }
+ nvinfer1::ITensor* tensor = tensor_or_weights.tensor();
+ if (!tensor) {
+ return tensorflow::errors::NotFound("Output tensor not found: " +
+ tensor_name);
+ }
+ converter.network()->markOutput(*tensor);
+ tensorflow::DataType tf_dtype = node->output_type(output_idx);
+ output_dtypes.push_back(tf_dtype);
+ nvinfer1::DataType trt_dtype = nvinfer1::DataType::kFLOAT;
+ TF_RETURN_IF_ERROR(ConvertDType(tf_dtype, &trt_dtype));
+ tensor->setType(trt_dtype);
+ }
+
+ VLOG(2) << "finished output";
+
+ // Build the engine
+ op_res->builder_->setMaxBatchSize(s.max_batch_size);
+ op_res->builder_->setMaxWorkspaceSize(s.max_workspace_size_bytes);
+
+ // Build the TRT op
+ // TODO(sami,ben,jie): proper naming!
+ tensorflow::NodeDefBuilder op_builder(calib_op_name, "TRTCalibOp");
+ std::vector<tensorflow::NodeDefBuilder::NodeOut> income_edges;
+ for (size_t i = 0; i < input_names.size(); ++i) {
+ int output_idx = s.input_inds.at(i).second;
+ // we wired up the input here already, it is redundant to do it again in
+ // ConvertSubGraphToTensorRT(convert_graph.cc)
+ auto incoming_edge = tensorflow::NodeDefBuilder::NodeOut(
+ input_names.at(i), output_idx, input_dtypes.at(i));
+ VLOG(1) << calib_op_name << " input " << i << " = " << input_names.at(i)
+ << ":" << output_idx
+ << " dType= " << tensorflow::DataTypeString(input_dtypes.at(i));
+ income_edges.push_back(incoming_edge);
+ }
+ tensorflow::gtl::ArraySlice<tensorflow::NodeDefBuilder::NodeOut> input_list(
+ income_edges);
+ op_builder.Input(input_list);
+ std::vector<string> segment_names;
+ segment_names.reserve(s.subgraph_node_ids.size());
+ for (int i : s.subgraph_node_ids) {
+ auto node = s.graph.FindNodeId(i);
+ segment_names.push_back(node->name());
+ }
+ LOG(INFO) << "finished op preparation";
+
+ auto status = op_builder.Attr("segment_nodes", segment_names)
+ .Attr("input_names", input_names)
+ .Attr("segment_output_names", output_names)
+ .Attr("resource_name", calib_op_name)
+ .Finalize(s.trt_node);
+
+ LOG(INFO) << status.ToString();
+ LOG(INFO) << "finished op building";
+
+ return tensorflow::Status::OK();
+}
+
+tensorflow::Status ConvertSubGraphToTensorRTNodeDef(
+ tensorrt::convert::SubGraphParams& s) {
+ // Visit nodes in reverse topological order and construct the TRT network.
+
+ // Toposort
+ std::vector<tensorflow::Node*> order_vec;
+ tensorflow::GetPostOrder(s.graph, &order_vec);
+ // Select just the subgraph
+ std::list<tensorflow::Node*> order;
+ for (tensorflow::Node* node : order_vec) {
+ if (s.subgraph_node_ids.count(node->id())) {
// We want topological order to contstruct the
// network layer by layer
order.push_front(node);
@@ -1434,46 +2438,86 @@
"Failed to create TensorRT network object");
}
+ string subgraph_name_scope;
+ if (!order.empty()) {
+ subgraph_name_scope = order.front()->name();
+ }
+ for (const tensorflow::Node* node : order) {
+ subgraph_name_scope = GetCommonNameScope(subgraph_name_scope, node->name());
+ }
+ static int static_id = 0;
+ // TODO(sami,ben,jie): proper naming!
+ string engine_name = StrCat(subgraph_name_scope, "my_trt_op");
+ engine_name = StrCat(engine_name, static_id++);
+ auto trt_rmgr = tensorflow::tensorrt::TRTResourceManager::instance();
+ auto weight_rmgr = trt_rmgr->getManager("WeightStore");
+ auto ws = new tensorflow::tensorrt::TRTWeightStore();
+ TF_CHECK_OK(weight_rmgr->Create(engine_name, engine_name, ws));
+
// Build the network
- Converter converter(trt_network.get());
+ Converter converter(trt_network.get(), ws, s.precision_mode == FP16MODE);
std::vector<string> input_names;
std::vector<tensorflow::DataType> input_dtypes;
- for (std::pair<int, int> const& input : input_inds) {
+ for (const std::pair<int, int>& input : s.input_inds) {
+ VLOG(2) << "parsing input!!!!!";
int node_id = input.first;
int output_idx = input.second;
- tensorflow::Node* node = graph.FindNodeId(node_id);
+ tensorflow::Node* node = s.graph.FindNodeId(node_id);
auto node_name = node->name();
- input_names.push_back(node_name); // Insert original node name without port
- // TODO(jie): alternative :)
- if (!graph_properties.HasOutputProperties(node_name))
- return tensorflow::errors::Internal("Failed to find input node: " +
- node_name);
+ // input_names should use the node name in the graph
+ // here it should be the input tensor name -> matching the binding
+ // insert original node name without port
+ auto tensor_name = node_name;
+ if (output_idx != 0) {
+ tensor_name = StrCat(tensor_name, ":", output_idx);
+ }
- auto op_info_vec = graph_properties.GetOutputProperties(node_name);
- if (static_cast<int>(op_info_vec.size()) < output_idx)
+ VLOG(2) << "input name: " << node_name << " tensor_name: " << tensor_name
+ << " idx: " << output_idx;
+
+ auto shape_inference_node_name = node_name;
+ auto shape_inference_output_idx = output_idx;
+ // rewire the shape inference to original node in the graph
+ if (s.output_edge_map->count(tensor_name)) {
+ shape_inference_node_name = s.output_edge_map->at(tensor_name).second;
+ shape_inference_output_idx = s.output_edge_map->at(tensor_name).first;
+ }
+ if (shape_inference_output_idx < 0) continue;
+ VLOG(2) << "shapeinference name: " << shape_inference_node_name
+ << " idx: " << shape_inference_output_idx;
+
+ if (!s.graph_properties.HasOutputProperties(shape_inference_node_name))
+ return tensorflow::errors::Internal("failed to find input node: " +
+ shape_inference_node_name);
+
+ auto op_info_vec =
+ s.graph_properties.GetOutputProperties(shape_inference_node_name);
+ if (static_cast<int>(op_info_vec.size()) <= shape_inference_output_idx)
return tensorflow::errors::Internal(
- "Accessing output index of: " + std::to_string(output_idx) +
- ", at node: " + node_name + " with output entry from shape_map: " +
- std::to_string(op_info_vec.size()));
+ "accessing output index of: ", shape_inference_output_idx,
+ ", at node: ", shape_inference_node_name,
+ " with output entry from shape_map: ", op_info_vec.size());
- auto op_info = op_info_vec.at(output_idx);
-
+ auto op_info = op_info_vec.at(shape_inference_output_idx);
tensorflow::DataType tf_dtype = op_info.dtype();
input_dtypes.push_back(tf_dtype);
nvinfer1::DataType dtype(nvinfer1::DataType::kFLOAT);
TF_CHECK_OK(ConvertDType(tf_dtype, &dtype));
- VLOG(2) << "Accessing output index of: " << std::to_string(output_idx)
+ VLOG(2) << "Accessing output index of: " << output_idx
<< ", at node: " << node_name
- << " with output entry from shape_map: "
- << std::to_string(op_info_vec.size());
-
+ << " with output entry from shape_map: " << op_info_vec.size();
// TODO(ben,jie): update TRT input format/dimension
nvinfer1::DimsCHW input_dim_psuedo_chw;
for (int i = 0; i < 3; i++) input_dim_psuedo_chw.d[i] = 1;
+ // TODO(jie): TRT 3.x only support 4 dimensional input tensor.
+ // update the code once TRT 4.0 comes out.
+ if (op_info.shape().dim_size() != 4)
+ return tensorflow::errors::Unimplemented("require 4 dimensional input");
+
for (int i = 1; i < op_info.shape().dim_size(); i++) {
VLOG(2) << "dimension: " << i
<< " , size: " << op_info.shape().dim(i).size();
@@ -1482,9 +2526,11 @@
// TODO(ben,jie): proper way to restore input tensor name?
auto input_tensor_name = node_name;
- if (output_idx != 0)
- input_tensor_name = node_name + ":" + std::to_string(output_idx);
+ if (output_idx != 0) {
+ input_tensor_name = StrCat(node_name, ":", output_idx);
+ }
+ input_names.push_back(input_tensor_name);
nvinfer1::ITensor* input_tensor = converter.network()->addInput(
input_tensor_name.c_str(), dtype, input_dim_psuedo_chw);
@@ -1511,14 +2557,22 @@
// Gather output metadata
std::vector<string> output_names;
std::vector<tensorflow::DataType> output_dtypes;
- for (std::pair<int, int> const& output : output_inds) {
+ int trt_engine_op_output_idx = 0;
+ for (const std::pair<int, int>& output : s.output_inds) {
int node_id = output.first;
int output_idx = output.second;
- tensorflow::Node* node = graph.FindNodeId(node_id);
+ tensorflow::Node* node = s.graph.FindNodeId(node_id);
string op_name = node->name();
string tensor_name = op_name;
+
+ s.output_edge_map->insert(
+ {trt_engine_op_output_idx == 0
+ ? engine_name
+ : StrCat(engine_name, ":", trt_engine_op_output_idx),
+ {output_idx, tensor_name}});
+ trt_engine_op_output_idx++;
if (output_idx != 0)
- tensor_name = tensor_name + ":" + std::to_string(output_idx);
+ tensorflow::strings::StrAppend(&tensor_name, ":", output_idx);
VLOG(2) << "Output tensor name: " << tensor_name;
output_names.push_back(tensor_name);
auto tensor_or_weights = converter.get_tensor(tensor_name);
@@ -1540,19 +2594,25 @@
}
VLOG(2) << "Finished output";
- // TODO(jie): static_id is not thread safe.
- static int static_id = 0;
// Build the engine
- trt_builder->setMaxBatchSize(max_batch_size);
- trt_builder->setMaxWorkspaceSize(max_workspace_size_bytes);
- VLOG(0) << "Starting build engine " << static_id;
- // TODO(ben,jie): half2 and int8 mode support
+ trt_builder->setMaxBatchSize(s.max_batch_size);
+ trt_builder->setMaxWorkspaceSize(s.max_workspace_size_bytes);
+ VLOG(0) << "Max batch size= " << s.max_batch_size
+ << " max workspace size= " << s.max_workspace_size_bytes;
+ if (s.precision_mode == FP16MODE) {
+ trt_builder->setHalf2Mode(true);
+ VLOG(0) << "Using FP16 precision mode";
+ }
+ LOG(INFO) << "starting build engine";
string engine_plan_string;
{
auto trt_engine =
infer_object(trt_builder->buildCudaEngine(*converter.network()));
VLOG(0) << "Built network";
+ if (trt_engine.get() == nullptr) {
+ return tensorflow::errors::Internal("Engine building failure");
+ }
auto engine_plan = infer_object(trt_engine->serialize());
VLOG(0) << "Serialized engine";
const char* engine_plan_data =
@@ -1560,18 +2620,19 @@
engine_plan_string =
string(engine_plan_data, engine_plan_data + engine_plan->size());
}
-
- VLOG(0) << "Finished engine";
+ TF_RETURN_IF_ERROR(weight_rmgr->Delete<tensorflow::tensorrt::TRTWeightStore>(
+ engine_name, engine_name));
+ LOG(INFO) << "finished engine " << engine_name;
// Build the TRT op
- // TODO(sami,ben,jie): proper naming!
- tensorflow::NodeDefBuilder op_builder(
- tensorflow::strings::StrCat("my_trt_op", static_id++), "TRTEngineOp");
+ tensorflow::NodeDefBuilder op_builder(engine_name, "TRTEngineOp");
std::vector<tensorflow::NodeDefBuilder::NodeOut> income_edges;
+ VLOG(2) << "input edge size: " << input_names.size();
for (size_t i = 0; i < input_names.size(); ++i) {
- int output_idx = input_inds.at(i).second;
- // We wired up the input here already, it is redundant to do it again in
- // ConvertSubGraphToTensorRT(convert_graph.cc)
+ VLOG(2) << "input edges: " << i << " " << input_names.at(i);
+ int output_idx = s.input_inds.at(i).second;
+ // we wired up the input here already, it is redundant to do it again in
+ // ConvertSubGraphToTensorRT(convert_graph.cc)
auto incoming_edge = tensorflow::NodeDefBuilder::NodeOut(
input_names.at(i), output_idx, input_dtypes.at(i));
income_edges.push_back(incoming_edge);
@@ -1586,7 +2647,7 @@
.Attr("input_nodes", input_names)
.Attr("output_nodes", output_names)
.Attr("OutT", output_dtypes)
- .Finalize(trt_node);
+ .Finalize(s.trt_node);
VLOG(0) << status.ToString() << " finished op building";
diff --git a/tensorflow/contrib/tensorrt/convert/convert_nodes.h b/tensorflow/contrib/tensorrt/convert/convert_nodes.h
index 2e7fd19..954a1e7 100644
--- a/tensorflow/contrib/tensorrt/convert/convert_nodes.h
+++ b/tensorflow/contrib/tensorrt/convert/convert_nodes.h
@@ -17,6 +17,8 @@
#define TENSORFLOW_CONTRIB_TENSORRT_CONVERT_CONVERT_NODES_H_
#include <set>
+#include <string>
+#include <unordered_map>
#include <utility>
#include <vector>
@@ -32,16 +34,49 @@
namespace tensorrt {
namespace convert {
-tensorflow::Status ConvertSubGraphToTensorRTNodeDef(
- const tensorflow::Graph& graph, const std::set<int>& subgraph_node_ids,
- const std::vector<std::pair<int, int>>&
- input_inds, // {node_id, output_idx}
- const std::vector<std::pair<int, int>>&
- output_inds, // {node_id, output_idx}
- size_t max_batch_size, size_t max_workspace_size_bytes,
- const tensorflow::grappler::GraphProperties& graph_prop,
- tensorflow::NodeDef* trt_node);
+const int FP32MODE = 0;
+const int FP16MODE = 1;
+const int INT8MODE = 2;
+struct SubGraphParams {
+ SubGraphParams(
+ tensorflow::Graph& inp_graph,
+ const std::set<int>& subgraph_node_id_numbers,
+ const std::vector<std::pair<int, int>>& input_indices,
+ const std::vector<std::pair<int, int>>& output_indices,
+ size_t max_supported_batch_size, size_t max_consumed_workspace_size_bytes,
+ const tensorflow::grappler::GraphProperties& current_graph_properties,
+ std::unordered_map<string, std::pair<int, string>>* output_edges,
+ tensorflow::NodeDef* constructed_trt_node,
+ int engine_precision_mode = FP32MODE)
+ : graph(inp_graph),
+ subgraph_node_ids(subgraph_node_id_numbers),
+ input_inds(input_indices),
+ output_inds(output_indices),
+ max_batch_size(max_supported_batch_size),
+ max_workspace_size_bytes(max_consumed_workspace_size_bytes),
+ graph_properties(current_graph_properties),
+ output_edge_map(output_edges),
+ trt_node(constructed_trt_node),
+ precision_mode(engine_precision_mode) {}
+
+ tensorflow::Graph& graph;
+ const std::set<int>& subgraph_node_ids;
+ const std::vector<std::pair<int, int>>& input_inds; // {node_id, output_idx}
+ const std::vector<std::pair<int, int>>& output_inds; // {node_id, output_idx}
+ size_t max_batch_size;
+ size_t max_workspace_size_bytes;
+ const tensorflow::grappler::GraphProperties& graph_properties;
+ std::unordered_map<string, std::pair<int, string>>* output_edge_map;
+ tensorflow::NodeDef* trt_node;
+ const int precision_mode;
+};
+
+// TODO(sami): Replace references with const reference or pointers
+tensorflow::Status ConvertSubGraphToTensorRTNodeDef(SubGraphParams& params);
+tensorflow::Status InjectCalibrationNode(SubGraphParams& params);
+tensorflow::Status ConvertCalibrationNodeToEngineNode(tensorflow::Graph& graph,
+ tensorflow::Node* c_node);
} // namespace convert
} // namespace tensorrt
} // namespace tensorflow
diff --git a/tensorflow/contrib/tensorrt/kernels/trt_calib_op.cc b/tensorflow/contrib/tensorrt/kernels/trt_calib_op.cc
index 1dcb87e..aea44fd 100644
--- a/tensorflow/contrib/tensorrt/kernels/trt_calib_op.cc
+++ b/tensorflow/contrib/tensorrt/kernels/trt_calib_op.cc
@@ -21,10 +21,11 @@
#include "tensorflow/core/framework/tensor_shape.h"
#include "tensorflow/core/framework/tensor_types.h"
#include "tensorflow/core/framework/types.h"
+#include "tensorflow/core/platform/stream_executor.h"
#if GOOGLE_CUDA
#if GOOGLE_TENSORRT
-#include "cuda_runtime_api.h"
+#include "cuda/include/cuda_runtime_api.h"
#include "tensorrt/include/NvInfer.h"
namespace tensorflow {
@@ -113,7 +114,13 @@
ctx->set_output(i, t);
}
VLOG(2) << "Filled map for sending";
- calib_res->calibrator_->setBatch(input_data);
+ // copied from cuda_kernel_helper since it seems only valid in *.cu.cc files
+ const cudaStream_t* stream = CHECK_NOTNULL(
+ reinterpret_cast<const cudaStream_t*>(ctx->op_device_context()
+ ->stream()
+ ->implementation()
+ ->CudaStreamMemberHack()));
+ calib_res->calibrator_->setBatch(input_data, *stream);
VLOG(2) << "Passed calibration data";
// TODO(aaroey): make sure we wait for the completion of calibration on the
// last batch in future PR.
diff --git a/tensorflow/contrib/tensorrt/kernels/trt_engine_op.cc b/tensorflow/contrib/tensorrt/kernels/trt_engine_op.cc
index 8efdf63..b32371b 100644
--- a/tensorflow/contrib/tensorrt/kernels/trt_engine_op.cc
+++ b/tensorflow/contrib/tensorrt/kernels/trt_engine_op.cc
@@ -24,8 +24,12 @@
#include "cuda/include/cuda_runtime_api.h"
namespace tensorflow {
-namespace tensorrt {
static ::tensorflow::tensorrt::Logger logger;
+namespace gpu = ::perftools::gputools;
+using IRuntime = nvinfer1::IRuntime;
+using Dims = nvinfer1::Dims;
+
+namespace tensorrt {
TRTEngineOp::TRTEngineOp(OpKernelConstruction* context) : OpKernel(context) {
// read serialized_engine
@@ -40,10 +44,21 @@
// TODO(samikama) runtime should be taken from a resourcemanager as well.
// Only engine should be in the op and context and runtime should be taken
// from resourcemanager
- nvinfer1::IRuntime* infer = nvinfer1::createInferRuntime(logger);
+ // TODO(jie): cudaSetDevice make sure trt engine is allocated on the same
+ // gpu where the input/output is also located.
+ int gpu_id = context->device()->tensorflow_gpu_device_info()->gpu_id;
+ cudaSetDevice(gpu_id);
+ int device;
+ cudaGetDevice(&device);
+ if (gpu_id != device) LOG(FATAL) << "set device failed!";
+
+ // TODO(samikama) runtime should be taken from a resourcemanager as well.
+ // Only engine should be in the op and context and runtime should be taken
+ // from resourcemanager
+
+ IRuntime* infer = nvinfer1::createInferRuntime(logger);
trt_engine_ptr_.reset(infer->deserializeCudaEngine(
serialized_engine.c_str(), serialized_engine.size(), nullptr));
-
trt_execution_context_ptr_.reset(trt_engine_ptr_->createExecutionContext());
// Runtime is safe to delete after engine creation
infer->destroy();
@@ -55,7 +70,6 @@
size_t binding_index;
int num_batch = 0;
- bool valid = true;
for (int i = 0; i < context->num_inputs(); i++) {
// Grab the input tensor
binding_index = trt_engine_ptr_->getBindingIndex(input_nodes_[i].c_str());
@@ -64,8 +78,12 @@
const TensorShape& input_shape = input_tensor.shape();
if (i == 0) {
num_batch = input_shape.dim_size(0);
+ if (num_batch > trt_engine_ptr_->getMaxBatchSize()) {
+ LOG(FATAL) << "input tensor batch larger than max_batch_size: "
+ << trt_engine_ptr_->getMaxBatchSize();
+ }
} else if (num_batch != input_shape.dim_size(0)) {
- valid = false;
+ LOG(FATAL) << "input data inconsistent batch size";
break;
}
switch (trt_engine_ptr_->getBindingDataType(binding_index)) {
@@ -81,9 +99,6 @@
}
}
- // Might want a different way to inform the user of batch size inconsistency
- if (!valid) LOG(WARNING) << "input data inconsistent batch size";
-
for (int i = 0; i < static_cast<int>(output_nodes_.size()); i++) {
// This is bad that we have to reallocate output buffer every run.
// Create an output tensor
@@ -126,9 +141,11 @@
->implementation()
->CudaStreamMemberHack()));
- // execution handled by TF since we are getting stream from TF.
- // it is safe for CPU pointer array (buffers) to go out of scope after enqueue
- trt_execution_context_ptr_->enqueue(num_batch, &buffers[0], *stream, nullptr);
+ // TODO(jie): trt enqueue does not return error
+ auto ret = trt_execution_context_ptr_->enqueue(num_batch, &buffers[0],
+ *stream, nullptr);
+ VLOG(2) << "enqueue returns: " << ret;
+ // sync should be done by TF.
}
REGISTER_KERNEL_BUILDER(Name("TRTEngineOp").Device(DEVICE_GPU), TRTEngineOp);
diff --git a/tensorflow/contrib/tensorrt/log/trt_logger.cc b/tensorflow/contrib/tensorrt/log/trt_logger.cc
index 7add8cb..dda0dc9 100644
--- a/tensorflow/contrib/tensorrt/log/trt_logger.cc
+++ b/tensorflow/contrib/tensorrt/log/trt_logger.cc
@@ -27,19 +27,19 @@
// Suppress info-level messages
switch (severity) {
case Severity::kINFO: { // Mark TRT info messages as debug!
- VLOG(2) << msg;
+ VLOG(2) << name_ << " " << msg;
break;
}
case Severity::kWARNING: {
- LOG(WARNING) << msg;
+ LOG(WARNING) << name_ << " " << msg;
break;
}
case Severity::kERROR: {
- LOG(ERROR) << msg;
+ LOG(ERROR) << name_ << " " << msg;
break;
}
case Severity::kINTERNAL_ERROR: {
- LOG(FATAL) << msg;
+ LOG(FATAL) << name_ << " " << msg;
break;
}
// This is useless for now. But would catch it in future if enum changes. It
diff --git a/tensorflow/contrib/tensorrt/log/trt_logger.h b/tensorflow/contrib/tensorrt/log/trt_logger.h
index d71f66b..7f3544f 100644
--- a/tensorflow/contrib/tensorrt/log/trt_logger.h
+++ b/tensorflow/contrib/tensorrt/log/trt_logger.h
@@ -27,9 +27,11 @@
// Logger for GIE info/warning/errors
class Logger : public nvinfer1::ILogger {
- private:
+ public:
+ Logger(string name = "DefaultLogger") : name_(name){};
void log(nvinfer1::ILogger::Severity severity, const char* msg) override;
+ private:
string name_;
};
diff --git a/tensorflow/contrib/tensorrt/python/__init__.py b/tensorflow/contrib/tensorrt/python/__init__.py
index 7e050a7..0b2321b 100644
--- a/tensorflow/contrib/tensorrt/python/__init__.py
+++ b/tensorflow/contrib/tensorrt/python/__init__.py
@@ -20,5 +20,6 @@
# pylint: disable=unused-import,line-too-long
from tensorflow.contrib.tensorrt.python.ops import trt_engine_op
+from tensorflow.contrib.tensorrt.python.trt_convert import calib_graph_to_infer_graph
from tensorflow.contrib.tensorrt.python.trt_convert import create_inference_graph
# pylint: enable=unused-import,line-too-long
diff --git a/tensorflow/contrib/tensorrt/python/trt_convert.py b/tensorflow/contrib/tensorrt/python/trt_convert.py
index 9454862..666220d 100644
--- a/tensorflow/contrib/tensorrt/python/trt_convert.py
+++ b/tensorflow/contrib/tensorrt/python/trt_convert.py
@@ -20,11 +20,17 @@
# pylint: disable=unused-import,line-too-long
import six as _six
+from tensorflow.contrib.tensorrt.wrap_conversion import calib_convert
from tensorflow.contrib.tensorrt.wrap_conversion import trt_convert
from tensorflow.core.framework import graph_pb2
+from tensorflow.core.protobuf import rewriter_config_pb2
from tensorflow.python.framework import errors
from tensorflow.python.framework import errors_impl as _impl
+from tensorflow.python.framework import meta_graph
from tensorflow.python.framework import ops
+from tensorflow.python.grappler import tf_optimizer
+from tensorflow.python.util import compat
+# pylint: enable=unused-import,line-too-long
# TODO(skama): get outputs from session when implemented as c++
@@ -32,22 +38,33 @@
def create_inference_graph(input_graph_def,
outputs,
max_batch_size=1,
- max_workspace_size_bytes=2 << 20):
+ max_workspace_size_bytes=2 << 20,
+ precision_mode="FP32",
+ minimum_segment_size=3):
"""Python wrapper for the TRT transormation.
-
Args:
input_graph_def: GraphDef object containing a model to be transformed.
- outputs: List of tensors or node names for the model outputs.
+ outputs: list of tensors or node names for the model outputs.
max_batch_size: max size for the input batch
max_workspace_size_bytes: parameter to control memory allocation (in Bytes)
+ precision_mode: one of 'FP32', 'FP16' and 'INT8'
+ minimum_segment_size: the minimum number of nodes required for a subgraph to
+ be replaced by TRTEngineOp.
Returns:
New GraphDef with TRTEngineOps placed in graph replacing subgraphs.
Raises:
+ ValueError: if the provided precision mode is invalid.
RuntimeError: if the returned status message is malformed.
"""
+ supported_precision_modes = {"FP32": 0, "FP16": 1, "INT8": 2}
+ if precision_mode.upper() not in supported_precision_modes:
+ raise ValueError(("precision mode '{}' is not supported."
+ "It should be one of {}").format(
+ precision_mode, "{'FP32', 'FP16', 'INT8'}"))
+ mode = supported_precision_modes[precision_mode.upper()]
def py2bytes(inp):
return inp
@@ -83,7 +100,7 @@
# pair or strings where first one is encoded status and the second
# one is the transformed graphs protobuf string.
out = trt_convert(input_graph_def_str, out_names, max_batch_size,
- max_workspace_size_bytes)
+ max_workspace_size_bytes, mode, minimum_segment_size)
status = to_string(out[0])
output_graph_def_string = out[1]
del input_graph_def_str # Save some memory
@@ -101,3 +118,46 @@
output_graph_def.ParseFromString(output_graph_def_string)
del output_graph_def_string # Save some memory
return output_graph_def
+
+
+def calib_graph_to_infer_graph(calibration_graph_def):
+ """Convert an existing calibration graph to inference graph.
+
+ Args:
+ calibration_graph_def: the calibration GraphDef object with calibration data
+ Returns:
+ New GraphDef with TRTEngineOps placed in graph replacing calibration nodes.
+ Raises:
+ RuntimeError: if the returned status message is malformed.
+ """
+
+ def py2string(inp):
+ return inp
+
+ def py3string(inp):
+ return inp.decode("utf-8")
+
+ if _six.PY2:
+ to_string = py2string
+ else:
+ to_string = py3string
+
+ graph_str = calibration_graph_def.SerializeToString()
+ out = calib_convert(graph_str)
+ status = to_string(out[0])
+ output_graph_def_string = out[1]
+ del graph_str # Save some memory
+ if len(status) < 2:
+ raise _impl.UnknownError(None, None, status)
+ if status[:2] != "OK":
+ msg = status.split(";")
+ if len(msg) == 1:
+ raise RuntimeError("Status message is malformed {}".format(status))
+ # pylint: disable=protected-access
+ raise _impl._make_specific_exception(None, None, ";".join(msg[1:]),
+ int(msg[0]))
+ # pylint: enable=protected-access
+ output_graph_def = graph_pb2.GraphDef()
+ output_graph_def.ParseFromString(output_graph_def_string)
+ del output_graph_def_string # Save some memory
+ return output_graph_def
diff --git a/tensorflow/contrib/tensorrt/resources/trt_int8_calibrator.cc b/tensorflow/contrib/tensorrt/resources/trt_int8_calibrator.cc
index 3d5cc76..74df759 100644
--- a/tensorflow/contrib/tensorrt/resources/trt_int8_calibrator.cc
+++ b/tensorflow/contrib/tensorrt/resources/trt_int8_calibrator.cc
@@ -23,7 +23,7 @@
#if GOOGLE_CUDA
#if GOOGLE_TENSORRT
-#include "cuda_runtime_api.h"
+#include "cuda/include/cuda_runtime_api.h"
namespace tensorflow {
namespace tensorrt {
@@ -38,22 +38,18 @@
done_(false),
dev_buffers_(dev_buffers),
calib_running_(false),
+ batch_is_set_(false),
engine_name_(engine_name) {}
-bool TRTInt8Calibrator::setBatch(
- const std::unordered_map<string, void*>& data) {
- // TODO(aaroey): make sure that in future PR:
- // 1. the mutex_lock is outside of the loop
- // 2. wait() is used instead of wait_for()
- // 3. done_ is to be protected by the mutex
- // 4. the first batch is not missed
- if (done_) return false;
- while (calib_running_.load(
- std::memory_order_acquire)) { // wait while calibration is running
- tensorflow::mutex_lock l(cond_mtx_);
- cond_.wait_for(l, std::chrono::milliseconds(50));
- if (done_) return false;
+bool TRTInt8Calibrator::setBatch(const std::unordered_map<string, void*>& data,
+ const cudaStream_t stream) {
+ tensorflow::mutex_lock lock(cond_mtx_);
+ while ((calib_running_ || batch_is_set_) &&
+ !done_) { // wait while calibration is running
+ cond_.wait(lock);
}
+ if (done_) return false;
+ CHECK(!calib_running_ && !batch_is_set_);
VLOG(1) << "Set Batch Waiting finished";
for (const auto it : data) {
auto devptr = dev_buffers_.find(it.first);
@@ -65,27 +61,32 @@
// TODO(aaroey): we should not use sync copy on default stream. Make sure
// stream->ThenMemcpy() is used in future PRs.
- auto status =
- cudaMemcpy(d.first, it.second, d.second, cudaMemcpyDeviceToDevice);
+ // TODO(sami,aaroey): Need to figureout a way to ensure synchronization
+ // between stream, perhaps using a tensor?
+ auto status = cudaMemcpyAsync(d.first, it.second, d.second,
+ cudaMemcpyDeviceToDevice, stream);
if (status != cudaSuccess) {
LOG(FATAL) << "cudaMemcpy " << engine_name_ << " for '" << it.first
<< "' failed with " << status;
}
}
- calib_running_.store(true, std::memory_order_release); // release builder
+
+ // TODO(Sami, aaorey): Find an alternative way!
+ cudaStreamSynchronize(
+ stream); // we have to wait for the stream before returning!
+ batch_is_set_ = true;
cond_.notify_all();
return true;
}
bool TRTInt8Calibrator::getBatch(void** bindings, const char** names,
int num_bindings) {
- calib_running_.store(false, std::memory_order_release); // wait for new batch
+ tensorflow::mutex_lock lock(cond_mtx_);
+ calib_running_ = false;
cond_.notify_all();
- while (!calib_running_.load(
- std::memory_order_acquire)) { // wait until new batch arrives
- tensorflow::mutex_lock l(cond_mtx_);
- cond_.wait_for(l, std::chrono::milliseconds(50));
- if (done_) return false;
+ while ((!batch_is_set_ && !done_)) { // wait until new batch arrives
+ cond_.wait(lock);
+
}
if (done_) {
return false;
@@ -100,6 +101,8 @@
bindings[i] = it->second.first;
}
+ batch_is_set_ = false;
+ calib_running_ = true;
return true;
}
@@ -107,6 +110,12 @@
return nullptr;
}
+void TRTInt8Calibrator::setDone() {
+ tensorflow::mutex_lock lock(cond_mtx_);
+ done_ = true;
+ cond_.notify_all();
+}
+
void TRTInt8Calibrator::writeCalibrationCache(const void* ptr,
std::size_t length) {}
TRTInt8Calibrator::~TRTInt8Calibrator() {
@@ -115,5 +124,6 @@
} // namespace tensorrt
} // namespace tensorflow
+
#endif
#endif
diff --git a/tensorflow/contrib/tensorrt/resources/trt_int8_calibrator.h b/tensorflow/contrib/tensorrt/resources/trt_int8_calibrator.h
index 8830f7e..d77aa2c 100644
--- a/tensorflow/contrib/tensorrt/resources/trt_int8_calibrator.h
+++ b/tensorflow/contrib/tensorrt/resources/trt_int8_calibrator.h
@@ -24,7 +24,10 @@
#if GOOGLE_CUDA
#if GOOGLE_TENSORRT
+
+#include "cuda/include/cuda_runtime_api.h"
#include "tensorrt/include/NvInfer.h"
+
namespace tensorflow {
namespace tensorrt {
// This class provides a 1 element queue to match TFs push model to
@@ -39,8 +42,9 @@
int getBatchSize() const override;
bool getBatch(void* bindings[], const char* names[],
int num_bindings) override;
- bool setBatch(const std::unordered_map<string, void*>& data);
- void setDone() { done_ = true; }
+ bool setBatch(const std::unordered_map<string, void*>& data,
+ const cudaStream_t stream);
+ void setDone();
const void* readCalibrationCache(std::size_t& length) override;
void writeCalibrationCache(const void* ptr, std::size_t length) override;
~TRTInt8Calibrator();
@@ -55,11 +59,14 @@
const std::unordered_map<string, std::pair<void*, size_t>>
dev_buffers_; // map to keep tensorrt input buffers and sizes keyed with
// buffer names
- std::atomic_bool calib_running_;
+ bool calib_running_;
+ bool batch_is_set_;
string engine_name_;
};
+
} // namespace tensorrt
} // namespace tensorflow
+
+#endif
+#endif
#endif // TENSORFLOW_CONTRIB_TENSORRT_RESOURCES_TRT_INT8_CALIBRATOR_H_
-#endif
-#endif
diff --git a/tensorflow/contrib/tensorrt/test/test_tftrt.py b/tensorflow/contrib/tensorrt/test/test_tftrt.py
index c78f6f2..0b661bd 100644
--- a/tensorflow/contrib/tensorrt/test/test_tftrt.py
+++ b/tensorflow/contrib/tensorrt/test/test_tftrt.py
@@ -60,6 +60,7 @@
def run_graph(gdef, dumm_inp):
+ """Run given graphdef once."""
gpu_options = cpb2.GPUOptions(per_process_gpu_memory_fraction=0.50)
ops.reset_default_graph()
g = ops.Graph()
@@ -74,15 +75,65 @@
return val
+# Use real data that is representatitive of the inference dataset
+# for calibration. For this test script it is random data.
+def run_calibration(gdef, dumm_inp):
+ """Run given calibration graph multiple times."""
+ gpu_options = cpb2.GPUOptions(per_process_gpu_memory_fraction=0.50)
+ ops.reset_default_graph()
+ g = ops.Graph()
+ with g.as_default():
+ inp, out = importer.import_graph_def(
+ graph_def=gdef, return_elements=["input", "output"])
+ inp = inp.outputs[0]
+ out = out.outputs[0]
+ with csess.Session(
+ config=cpb2.ConfigProto(gpu_options=gpu_options), graph=g) as sess:
+ # run over real calibration data here, we are mimicking a calibration set of
+ # 30 different batches. Use as much calibration data as you want
+ for _ in range(30):
+ val = sess.run(out, {inp: dumm_inp})
+ return val
+
+
if "__main__" in __name__:
inp_dims = (100, 24, 24, 2)
dummy_input = np.random.random_sample(inp_dims)
- gdef = get_simple_graph_def()
+ orig_graph = get_simple_graph_def() # use a frozen graph for inference
# Get optimized graph
- trt_graph = trt.create_inference_graph(gdef, ["output"], inp_dims[0])
- o1 = run_graph(gdef, dummy_input)
+ trt_graph = trt.create_inference_graph(
+ input_graph_def=orig_graph,
+ outputs=["output"],
+ max_batch_size=inp_dims[0],
+ max_workspace_size_bytes=1 << 25,
+ precision_mode="FP32", # TRT Engine precision "FP32","FP16" or "INT8"
+ minimum_segment_size=2 # minimum number of nodes in an engine
+ )
+ o1 = run_graph(orig_graph, dummy_input)
o2 = run_graph(trt_graph, dummy_input)
o3 = run_graph(trt_graph, dummy_input)
assert np.array_equal(o1, o2)
assert np.array_equal(o3, o2) # sanity check
+ fp16_graph = trt.create_inference_graph(
+ input_graph_def=orig_graph,
+ outputs=["output"],
+ max_batch_size=inp_dims[0],
+ max_workspace_size_bytes=1 << 25,
+ precision_mode="FP16", # TRT Engine precision "FP32","FP16" or "INT8"
+ minimum_segment_size=2 # minimum number of nodes in an engine
+ )
+ int8_calib_gdef = trt.create_inference_graph(
+ input_graph_def=orig_graph,
+ outputs=["output"],
+ max_batch_size=inp_dims[0],
+ max_workspace_size_bytes=1 << 25,
+ precision_mode="INT8", # TRT Engine precision "FP32","FP16" or "INT8"
+ minimum_segment_size=2 # minimum number of nodes in an engine
+ )
+ o4 = run_graph(fp16_graph, dummy_input)
+ _ = run_calibration(int8_calib_gdef, dummy_input)
+ int8_graph = trt.calib_graph_to_infer_graph(int8_calib_gdef)
+ o5 = run_graph(int8_graph, dummy_input)
+ assert np.allclose(o1, o4)
+ assert np.allclose(o1, o5)
print("Pass")
diff --git a/tensorflow/contrib/tensorrt/trt_conversion.i b/tensorflow/contrib/tensorrt/trt_conversion.i
index d679945..46480e9 100644
--- a/tensorflow/contrib/tensorrt/trt_conversion.i
+++ b/tensorflow/contrib/tensorrt/trt_conversion.i
@@ -64,13 +64,17 @@
%ignoreall
%unignore tensorflow;
%unignore trt_convert;
+%unignore calib_convert;
%{
+
std::pair<string, string> trt_convert(
string graph_def_string, // The serialized GraphDef string.
std::vector<string> output_names,
size_t max_batch_size,
- size_t max_workspace_size_bytes
+ size_t max_workspace_size_bytes,
+ int precision_mode,
+ int minimum_segment_size
// Unfortunately we can't use TF_Status here since it
// is in c/c_api and brings in a lot of other libraries
// which in turn declare ops. These ops are included
@@ -90,16 +94,64 @@
return std::pair<string, string>{out_status, ""};
}
+ if(precision_mode < 0 || precision_mode > 2){
+ out_status = "InvalidArgument;Invalid precision_mode";
+ return std::pair<string, string>{out_status, ""};
+ }
if (!output_names.size()) {
out_status = "InvalidArgument;Size of the output_names vector is 0";
return std::pair<string, string>{out_status, ""};
- // return "";
}
tensorflow::GraphDef outGraph;
tensorflow::Status conversion_status =
tensorflow::tensorrt::convert::ConvertGraphDefToTensorRT(
graph_def, output_names, max_batch_size, max_workspace_size_bytes,
- &outGraph);
+ &outGraph, precision_mode, minimum_segment_size);
+ if (!conversion_status.ok()) {
+ auto retCode = (int)conversion_status.code();
+ char buff[2000];
+ snprintf(buff, 2000, "%d;%s", retCode,
+ conversion_status.error_message().c_str());
+ out_status = buff;
+ return std::pair<string, string>{out_status, ""};
+ }
+ string result;
+ if (!outGraph.SerializeToString(&result)) {
+ out_status = "InvalidArgument;Couldn't serialize output as a GraphDef";
+ return std::pair<string, string>{out_status, ""};
+ }
+ out_status = "OK;All good!";
+ return std::pair<string, string>{out_status, result};
+#else
+ // Returns FAILED_PRECONDITION.
+ return std::pair<string, string>{"9;TensorRT is not enabled!", ""};
+#endif // GOOGLE_CUDA && GOOGLE_TENSORRT
+}
+
+std::pair<string, string> calib_convert(string graph_def_string // const tensorflow::GraphDef&
+ // unfortunately we can't use TF_Status here since it
+ // is in c/c_api and brings in a lot of other libraries
+ // which in turn declare ops. These ops are included
+ // statically in our library and cause an abort when
+ // module is loaded due to double registration
+ // until Tensorflow properly exposes these headers
+ // we have to work around this by returning a string
+ // and converting it to exception on python side.
+ //,TF_Status* out_status) {
+) {
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
+ string out_status;
+
+ tensorflow::GraphDef graph_def;
+ if (!graph_def.ParseFromString(graph_def_string)) {
+ out_status = "InvalidArgument;Couldn't interpret input as a GraphDef";
+ return std::pair<string, string>{out_status, ""};
+ }
+
+ tensorflow::GraphDef outGraph;
+ tensorflow::Status conversion_status =
+ tensorflow::tensorrt::convert::ConvertCalibGraphToInferGraph(graph_def,
+ &outGraph);
if (!conversion_status.ok()) {
auto retCode = (int)conversion_status.code();
char buff[2000];
@@ -122,10 +174,13 @@
}
%}
+std::pair<string, string> calib_convert(string graph_def_string);
+
std::pair<string, string> trt_convert(string graph_def_string,
std::vector<string> output_names,
size_t max_batch_size,
- size_t max_workspace_size_bytes);
+ size_t max_workspace_size_bytes,
+ int precision_mode, int minimum_segment_size);
%unignoreall
diff --git a/tensorflow/contrib/timeseries/examples/BUILD b/tensorflow/contrib/timeseries/examples/BUILD
index bb86ecb..70bf67c 100644
--- a/tensorflow/contrib/timeseries/examples/BUILD
+++ b/tensorflow/contrib/timeseries/examples/BUILD
@@ -25,7 +25,10 @@
srcs = ["predict_test.py"],
data = ["data/period_trend.csv"],
srcs_version = "PY2AND3",
- tags = ["notsan"], # b/67513579
+ tags = [
+ "no_windows", # TODO: needs investigation on Windows
+ "notsan", # b/67513579
+ ],
deps = [
":predict",
"//tensorflow/python:client_testlib",
diff --git a/tensorflow/contrib/timeseries/python/timeseries/BUILD b/tensorflow/contrib/timeseries/python/timeseries/BUILD
index ed3ed4c..64f5cd8 100644
--- a/tensorflow/contrib/timeseries/python/timeseries/BUILD
+++ b/tensorflow/contrib/timeseries/python/timeseries/BUILD
@@ -156,9 +156,7 @@
"head_test.py",
],
srcs_version = "PY2AND3",
- tags = [
- "no_pip_gpu", # b/63391119
- ],
+ tags = ["no_pip_gpu"], # b/63391119
deps = [
":feature_keys",
":head",
@@ -427,6 +425,7 @@
srcs_version = "PY2AND3",
tags = [
"no_pip_gpu", # b/63391119
+ "no_windows", # TODO: needs investigation on Windows
],
deps = [
":feature_keys",
diff --git a/tensorflow/contrib/timeseries/python/timeseries/state_space_models/BUILD b/tensorflow/contrib/timeseries/python/timeseries/state_space_models/BUILD
index c86d06e..07df7bc 100644
--- a/tensorflow/contrib/timeseries/python/timeseries/state_space_models/BUILD
+++ b/tensorflow/contrib/timeseries/python/timeseries/state_space_models/BUILD
@@ -40,6 +40,7 @@
timeout = "long", # Moderate but for asan
srcs = ["state_space_model_test.py"],
srcs_version = "PY2AND3",
+ tags = ["no_windows"], # TODO: needs investigation on Windows
deps = [
":state_space_model",
"//tensorflow/contrib/layers:layers_py",
diff --git a/tensorflow/contrib/tpu/BUILD b/tensorflow/contrib/tpu/BUILD
index ed930e4..f9d433a 100644
--- a/tensorflow/contrib/tpu/BUILD
+++ b/tensorflow/contrib/tpu/BUILD
@@ -225,6 +225,7 @@
"//tensorflow/python:framework",
"//tensorflow/python:layers",
],
+ tags = ["no_windows"], # TODO: needs investigation on Windows
)
tf_py_test(
diff --git a/tensorflow/contrib/util/loader.py b/tensorflow/contrib/util/loader.py
index f4283cd..dca01d2 100644
--- a/tensorflow/contrib/util/loader.py
+++ b/tensorflow/contrib/util/loader.py
@@ -42,9 +42,10 @@
plugin.
"""
if os.name == 'nt':
- # To avoid makeing every user_ops aware of windows, re-write
- # the file extension from .so to .dll.
- path = re.sub(r'\.so$', '.dll', path)
+ # To avoid making every user_ops aware of windows, re-write
+ # the file extension from .so to .dll if .so file doesn't exist.
+ if not os.path.exists(path):
+ path = re.sub(r'\.so$', '.dll', path)
# Currently we have only some user_ops as dlls on windows - don't try
# to load them if the dll is not found.
diff --git a/tensorflow/core/BUILD b/tensorflow/core/BUILD
index 352e183..5c9fd2f 100644
--- a/tensorflow/core/BUILD
+++ b/tensorflow/core/BUILD
@@ -3327,6 +3327,10 @@
size = "small",
srcs = ["common_runtime/function_test.cc"],
linkstatic = tf_kernel_tests_linkstatic(),
+ tags = [
+ "manual",
+ "no_oss",
+ ],
deps = [
":core",
":core_cpu",
diff --git a/tensorflow/core/api_def/base_api/api_def_SlideDataset.pbtxt b/tensorflow/core/api_def/base_api/api_def_SlideDataset.pbtxt
new file mode 100644
index 0000000..9fabe78
--- /dev/null
+++ b/tensorflow/core/api_def/base_api/api_def_SlideDataset.pbtxt
@@ -0,0 +1,18 @@
+op {
+ graph_op_name: "SlideDataset"
+ in_arg {
+ name: "window_size"
+ description: <<END
+A scalar representing the number of elements in the
+sliding window.
+END
+ }
+ in_arg {
+ name: "stride"
+ description: <<END
+A scalar representing the steps moving the sliding window
+forward in one iteration. It must be in `[1, window_size)`.
+END
+ }
+ summary: "Creates a dataset that passes a sliding window over `input_dataset`."
+}
diff --git a/tensorflow/core/distributed_runtime/tensor_coding.cc b/tensorflow/core/distributed_runtime/tensor_coding.cc
index 34a4013..fe2d1a1 100644
--- a/tensorflow/core/distributed_runtime/tensor_coding.cc
+++ b/tensorflow/core/distributed_runtime/tensor_coding.cc
@@ -81,7 +81,7 @@
Status TensorResponse::ParseFrom(Source* source) {
if (!on_host_) {
protobuf::io::CodedInputStream input(source->contents());
- input.SetTotalBytesLimit(INT_MAX); // Unlimited
+ input.SetTotalBytesLimit(INT_MAX, INT_MAX); // Unlimited
// Pre-parse into local storage, then delegate to device.
if (!meta_.ParseFromCodedStream(&input) || !input.ConsumedEntireMessage()) {
@@ -217,7 +217,7 @@
bool TensorResponse::ParseFast(Source* source) {
protobuf::io::CodedInputStream input(source->contents());
- input.SetTotalBytesLimit(INT_MAX); // Unlimited
+ input.SetTotalBytesLimit(INT_MAX, INT_MAX); // Unlimited
while (true) {
auto p = input.ReadTagWithCutoff(127);
int tag = GetTagFieldNumber(p.first);
diff --git a/tensorflow/core/framework/dataset.h b/tensorflow/core/framework/dataset.h
index beaf0ad..cfe23d1 100644
--- a/tensorflow/core/framework/dataset.h
+++ b/tensorflow/core/framework/dataset.h
@@ -474,11 +474,11 @@
}
// Key for storing the Dataset graph in the serialized format.
- static const char kDatasetGraphKey[];
+ TF_EXPORT static const char kDatasetGraphKey[];
// Key for storing the output node of the Dataset graph in the serialized
// format.
- static const char kDatasetGraphOutputNodeKey[];
+ TF_EXPORT static const char kDatasetGraphOutputNodeKey[];
private:
Status Serialize(OpKernelContext* ctx, string* serialized_graph_def,
diff --git a/tensorflow/core/graph/mkl_layout_pass.cc b/tensorflow/core/graph/mkl_layout_pass.cc
index 02038c5..1507b6e 100644
--- a/tensorflow/core/graph/mkl_layout_pass.cc
+++ b/tensorflow/core/graph/mkl_layout_pass.cc
@@ -2492,10 +2492,10 @@
mkl_op_registry::GetMklOpName(csinfo_.identity),
CopyAttrsDataType, AlwaysRewrite});
rinfo_.push_back({csinfo_.lrn, mkl_op_registry::GetMklOpName(csinfo_.lrn),
- CopyAttrsLRN, AlwaysRewrite});
+ CopyAttrsLRN, LrnRewrite});
rinfo_.push_back({csinfo_.lrn_grad,
mkl_op_registry::GetMklOpName(csinfo_.lrn_grad),
- CopyAttrsLRN, AlwaysRewrite});
+ CopyAttrsLRN, LrnRewrite});
rinfo_.push_back({csinfo_.max_pool,
mkl_op_registry::GetMklOpName(csinfo_.max_pool),
CopyAttrsPooling, NonDepthBatchWisePoolRewrite});
@@ -2865,6 +2865,28 @@
return false;
}
+ // If the depth_radius of LRN is not 2, then MKL DNN takes unoptimized
+ // path. The unoptimized path is slow. Thus we dont rewrite the node
+ // and use default Eigen. But for depth_radius=2, MKL DNN optimized
+ // path is taken, i.e., eigen node is rewritten by MKl DNN node.
+ static bool LrnRewrite(const Node* n) {
+ CHECK_NOTNULL(n);
+
+ int depth_radius;
+ CHECK_EQ(GetNodeAttr(n->def(), "depth_radius", &depth_radius).ok(), true);
+
+ // if the depth_radius of LRN is not 2, don't rewrite the node by MKL DNN
+ // and use eigen node instead
+ if (depth_radius == 2) {
+ return true;
+ }
+ VLOG(1) << "LrnRewrite: The model sets depth_radius as not 2 which"
+ << "case is not optimized by Intel MKL, thus using Eigen op"
+ << "for LRN " ;
+
+ return false;
+ }
+
static bool AddNRewrite(const Node* n) {
CHECK_NOTNULL(n);
@@ -3528,11 +3550,13 @@
string data_format;
string padding;
std::vector<int32> strides;
+ std::vector<int32> dilations;
bool use_cudnn_on_gpu;
// Get all attributes from old node.
TF_CHECK_OK(GetNodeAttr(orig_node->def(), "T", &T));
TF_CHECK_OK(GetNodeAttr(orig_node->def(), "strides", &strides));
+ TF_CHECK_OK(GetNodeAttr(orig_node->def(), "dilations", &dilations));
TF_CHECK_OK(GetNodeAttr(orig_node->def(), "padding", &padding));
TF_CHECK_OK(GetNodeAttr(orig_node->def(), "data_format", &data_format));
TF_CHECK_OK(
@@ -3541,6 +3565,7 @@
// Add attributes to new node.
nb->Attr("T", T);
nb->Attr("strides", strides);
+ nb->Attr("dilations", dilations);
nb->Attr("padding", padding);
nb->Attr("data_format", data_format);
nb->Attr("use_cudnn_on_gpu", use_cudnn_on_gpu);
@@ -3778,12 +3803,14 @@
DataType T_pred, T_succ;
string padding;
std::vector<int32> strides;
+ std::vector<int32> dilations;
string data_format_pred, data_format_succ;
bool use_cudnn_on_gnu;
TF_CHECK_OK(GetNodeAttr(pred->def(), "T", &T_pred));
TF_CHECK_OK(GetNodeAttr(succ->def(), "T", &T_succ));
TF_CHECK_OK(GetNodeAttr(pred->def(), "padding", &padding));
TF_CHECK_OK(GetNodeAttr(pred->def(), "strides", &strides));
+ TF_CHECK_OK(GetNodeAttr(pred->def(), "dilations", &dilations));
TF_CHECK_OK(GetNodeAttr(pred->def(), "data_format", &data_format_pred));
TF_CHECK_OK(GetNodeAttr(succ->def(), "data_format", &data_format_succ));
TF_CHECK_OK(GetNodeAttr(pred->def(), "use_cudnn_on_gpu", &use_cudnn_on_gnu));
diff --git a/tensorflow/core/grappler/optimizers/loop_optimizer.cc b/tensorflow/core/grappler/optimizers/loop_optimizer.cc
index 2446535..8f13c4a 100644
--- a/tensorflow/core/grappler/optimizers/loop_optimizer.cc
+++ b/tensorflow/core/grappler/optimizers/loop_optimizer.cc
@@ -25,6 +25,7 @@
#include "tensorflow/core/framework/attr_value.pb.h"
#include "tensorflow/core/framework/node_def.pb.h"
#include "tensorflow/core/framework/op.h"
+#include "tensorflow/core/framework/tensor_shape.pb.h"
#include "tensorflow/core/framework/types.h"
#include "tensorflow/core/grappler/costs/graph_properties.h"
#include "tensorflow/core/grappler/grappler_item.h"
@@ -408,7 +409,7 @@
frame_children_[frame_ids[0]].insert(frame_ids[1]);
frame_parent_[frame_ids.back()] = frame_ids[frame_ids.size() - 2];
}
- if (!frame_ids.empty()) {
+ if (frame_ids.size() >= 1) {
frame_children_.insert(std::make_pair(frame_ids.back(), empty_set_));
if (node->op() == "LoopCond") {
if (loop_cond_.count(frame_ids.back())) {
@@ -427,7 +428,7 @@
}
for (auto it = frame_children_.begin(); it != frame_children_.end(); ++it) {
- if (it->second.empty()) {
+ if (it->second.size() == 0) {
worklist.push_back(it->first);
}
}
@@ -440,7 +441,7 @@
if (parent_it != frame_parent_.end()) {
int parent_id = parent_it->second;
frame_children_[parent_id].erase(frame_id);
- if (frame_children_[parent_id].empty()) {
+ if (frame_children_[parent_id].size() == 0) {
worklist.push_back(parent_id);
}
}
@@ -464,7 +465,6 @@
Status LoopOptimizer::Optimize(Cluster* cluster, const GrapplerItem& item,
GraphDef* optimized_graph) {
TF_RETURN_IF_ERROR(RemoveStackOps(item.graph, optimized_graph));
-
optimized_graph_ = optimized_graph;
// Set up helper data structures.
diff --git a/tensorflow/core/kernels/BUILD b/tensorflow/core/kernels/BUILD
index 48d5955..2e39f25 100644
--- a/tensorflow/core/kernels/BUILD
+++ b/tensorflow/core/kernels/BUILD
@@ -5157,7 +5157,6 @@
],
hdrs = [
"meta_support.h",
- "quantization_utils.h",
"reference_gemm.h",
],
deps = [
@@ -5232,6 +5231,7 @@
name = "quantization_utils_test",
srcs = ["quantization_utils_test.cc"],
deps = [
+ ":quantization_utils",
":quantized_ops",
"//tensorflow/core:array_ops_op_lib",
"//tensorflow/core:core_cpu",
@@ -5294,6 +5294,7 @@
deps = [
":ops_testutil",
":ops_util",
+ ":quantization_utils",
":quantized_ops",
"//tensorflow/core:array_ops_op_lib",
"//tensorflow/core:framework",
@@ -5355,6 +5356,7 @@
":math",
":ops_testutil",
":ops_util",
+ ":quantization_utils",
":quantized_ops",
"//tensorflow/cc:cc_ops",
"//tensorflow/cc:client_session",
@@ -5377,6 +5379,7 @@
deps = [
":ops_testutil",
":ops_util",
+ ":quantization_utils",
":quantized_ops",
"//tensorflow/cc:cc_ops",
"//tensorflow/cc:client_session",
@@ -5441,6 +5444,7 @@
deps = [
":ops_testutil",
":ops_util",
+ ":quantization_utils",
":quantized_ops",
"//tensorflow/core:array_ops_op_lib",
"//tensorflow/core:framework",
@@ -5461,6 +5465,7 @@
deps = [
":ops_testutil",
":ops_util",
+ ":quantization_utils",
":quantized_ops",
"//tensorflow/core:array_ops_op_lib",
"//tensorflow/core:framework",
@@ -5500,6 +5505,7 @@
deps = [
":ops_testutil",
":ops_util",
+ ":quantization_utils",
":quantized_ops",
"//tensorflow/core:array_ops_op_lib",
"//tensorflow/core:framework",
@@ -5556,6 +5562,7 @@
":math",
":ops_testutil",
":ops_util",
+ ":quantization_utils",
":quantized_ops",
"//tensorflow/cc:cc_ops",
"//tensorflow/cc:client_session",
@@ -5578,6 +5585,7 @@
deps = [
":ops_testutil",
":ops_util",
+ ":quantization_utils",
":quantized_ops",
"//tensorflow/core:array_ops_op_lib",
"//tensorflow/core:framework",
@@ -5614,6 +5622,7 @@
deps = [
":ops_testutil",
":ops_util",
+ ":quantization_utils",
":quantized_ops",
"//tensorflow/core:array_ops_op_lib",
"//tensorflow/core:core_cpu",
@@ -5635,6 +5644,7 @@
deps = [
":batch_norm_op",
":ops_testutil",
+ ":quantization_utils",
":quantized_ops",
"//tensorflow/core:array_ops_op_lib",
"//tensorflow/core:core_cpu_internal",
diff --git a/tensorflow/core/kernels/data/BUILD b/tensorflow/core/kernels/data/BUILD
index 484d4f8..01754ec 100644
--- a/tensorflow/core/kernels/data/BUILD
+++ b/tensorflow/core/kernels/data/BUILD
@@ -114,6 +114,19 @@
)
tf_kernel_library(
+ name = "slide_dataset_op",
+ srcs = ["slide_dataset_op.cc"],
+ deps = [
+ ":dataset",
+ "//tensorflow/core:dataset_ops_op_lib",
+ "//tensorflow/core:framework",
+ "//tensorflow/core:lib",
+ "//tensorflow/core:lib_internal",
+ "//tensorflow/core/kernels:batch_util",
+ ],
+)
+
+tf_kernel_library(
name = "padded_batch_dataset_op",
srcs = ["padded_batch_dataset_op.cc"],
deps = [
@@ -538,6 +551,7 @@
":scan_dataset_op",
":shuffle_dataset_op",
":skip_dataset_op",
+ ":slide_dataset_op",
":sparse_tensor_slice_dataset_op",
":sql_dataset_ops",
":stats_aggregator_ops",
diff --git a/tensorflow/core/kernels/data/slide_dataset_op.cc b/tensorflow/core/kernels/data/slide_dataset_op.cc
new file mode 100644
index 0000000..4f3537b
--- /dev/null
+++ b/tensorflow/core/kernels/data/slide_dataset_op.cc
@@ -0,0 +1,252 @@
+/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#include "tensorflow/core/framework/partial_tensor_shape.h"
+#include "tensorflow/core/framework/tensor.h"
+#include "tensorflow/core/kernels/batch_util.h"
+#include "tensorflow/core/kernels/data/dataset.h"
+
+namespace tensorflow {
+
+namespace {
+
+// See documentation in ../ops/dataset_ops.cc for a high-level
+// description of the following op.
+
+class SlideDatasetOp : public UnaryDatasetOpKernel {
+ public:
+ explicit SlideDatasetOp(OpKernelConstruction* ctx)
+ : UnaryDatasetOpKernel(ctx) {}
+
+ void MakeDataset(OpKernelContext* ctx, DatasetBase* input,
+ DatasetBase** output) override {
+ int64 window_size = 0;
+ int64 stride = 1;
+ OP_REQUIRES_OK(ctx,
+ ParseScalarArgument<int64>(ctx, "window_size", &window_size));
+ OP_REQUIRES_OK(ctx,
+ ParseScalarArgument<int64>(ctx, "stride", &stride));
+ OP_REQUIRES(
+ ctx, window_size > 0,
+ errors::InvalidArgument("Window size must be greater than zero."));
+ OP_REQUIRES(
+ ctx, stride > 0 && stride < window_size,
+ errors::InvalidArgument("Stride must be in [1, window_size)."));
+
+ *output = new Dataset(ctx, window_size, stride, input);
+ }
+
+ private:
+ class Dataset : public GraphDatasetBase {
+ public:
+ Dataset(OpKernelContext* ctx, int64 window_size, int64 stride, const DatasetBase* input)
+ : GraphDatasetBase(ctx), window_size_(window_size), stride_(stride), input_(input) {
+ input_->Ref();
+
+ const auto& input_shapes = input_->output_shapes();
+ output_shapes_.reserve(input_shapes.size());
+ for (const auto& input_shape : input_shapes) {
+ output_shapes_.emplace_back(
+ PartialTensorShape({-1}).Concatenate(input_shape));
+ }
+ }
+
+ ~Dataset() override { input_->Unref(); }
+
+ std::unique_ptr<IteratorBase> MakeIterator(
+ const string& prefix) const override {
+ return std::unique_ptr<IteratorBase>(new Iterator(
+ Iterator::Params{this, strings::StrCat(prefix, "::Slide")}));
+ }
+
+ const DataTypeVector& output_dtypes() const override {
+ return input_->output_dtypes();
+ }
+
+ const std::vector<PartialTensorShape>& output_shapes() const override {
+ return output_shapes_;
+ }
+
+ string DebugString() override {
+ return strings::StrCat("SlideDatasetOp(", window_size_, ", ", stride_, ")::Dataset");
+ }
+
+ protected:
+ Status AsGraphDefInternal(OpKernelContext* ctx, DatasetGraphDefBuilder* b,
+ Node** output) const override {
+ Node* input_graph_node = nullptr;
+ TF_RETURN_IF_ERROR(b->AddParentDataset(ctx, input_, &input_graph_node));
+ Node* window_size = nullptr;
+ Node* stride = nullptr;
+ TF_RETURN_IF_ERROR(b->AddScalar(window_size_, &window_size));
+ TF_RETURN_IF_ERROR(b->AddScalar(stride_, &stride));
+ TF_RETURN_IF_ERROR(
+ b->AddDataset(this, {input_graph_node, window_size, stride}, output));
+ return Status::OK();
+ }
+
+ private:
+
+ class Iterator : public DatasetIterator<Dataset> {
+ public:
+ explicit Iterator(const Params& params)
+ : DatasetIterator<Dataset>(params),
+ input_impl_(params.dataset->input_->MakeIterator(params.prefix)) {}
+
+ Status GetNextInternal(IteratorContext* ctx,
+ std::vector<Tensor>* out_tensors,
+ bool* end_of_sequence) override {
+ const int64 window_size = dataset()->window_size_;
+ const int64 stride = dataset()->stride_;
+ std::vector<std::vector<Tensor>> batch_elements;
+ {
+ mutex_lock l(mu_);
+ if (!input_impl_) {
+ *end_of_sequence = true;
+ return Status::OK();
+ }
+ batch_elements.reserve(window_size);
+ const bool first_call = cache_.empty();
+ if (first_call) {
+ cache_.reserve(window_size);
+ } else {
+ // Reuse cache in the previous iteration.
+ cache_.swap(batch_elements);
+ }
+ // Fill up with new elements.
+ *end_of_sequence = false;
+ for (size_t i = batch_elements.size(); i < window_size && !*end_of_sequence;
+ ++i) {
+ std::vector<Tensor> batch_element_tuple;
+ TF_RETURN_IF_ERROR(input_impl_->GetNext(ctx, &batch_element_tuple,
+ end_of_sequence));
+ if (!*end_of_sequence) {
+ batch_elements.push_back(std::move(batch_element_tuple));
+ } else {
+ input_impl_.reset();
+ }
+ }
+ // Drop the final smaller blocks.
+ if (batch_elements.size() < window_size) {
+ DCHECK(*end_of_sequence);
+ return Status::OK();
+ }
+ // Cache the data used for the next iteration.
+ for (size_t i = stride; i < window_size; ++i) {
+ cache_.emplace_back(batch_elements[i]);
+ }
+ }
+
+ // Construct output tensors.
+ // Those codes below are copied from batch_dataset_op.cc.
+ const size_t num_tuple_components = batch_elements[0].size();
+ const int64 num_batch_elements = batch_elements.size();
+ for (size_t component_index = 0; component_index < num_tuple_components;
+ ++component_index) {
+ const Tensor& first_element = batch_elements[0][component_index];
+ TensorShape batch_component_shape({num_batch_elements});
+ batch_component_shape.AppendShape(first_element.shape());
+ Tensor batch_component(cpu_allocator(), first_element.dtype(),
+ batch_component_shape);
+ // Build the output tuple component by copying one slice
+ // from each input element in the batch.
+ for (size_t i = 0; i < num_batch_elements; ++i) {
+ if (batch_elements[i][component_index].shape() !=
+ first_element.shape()) {
+ return errors::InvalidArgument(
+ "Cannot batch tensors with different shapes in component ",
+ component_index, ". First element had shape ",
+ first_element.shape().DebugString(), " and element ", i,
+ " had shape ",
+ batch_elements[i][component_index].shape().DebugString(),
+ ".");
+ }
+ TF_RETURN_IF_ERROR(batch_util::CopyElementToSlice(
+ std::move(batch_elements[i][component_index]), &batch_component,
+ i));
+ }
+ out_tensors->emplace_back(std::move(batch_component));
+ }
+ *end_of_sequence = false;
+ return Status::OK();
+ }
+
+ protected:
+ Status SaveInternal(IteratorStateWriter* writer) override {
+ mutex_lock l(mu_);
+ if (!input_impl_) {
+ TF_RETURN_IF_ERROR(
+ writer->WriteScalar(full_name("input_impl_empty"), ""));
+ } else {
+ TF_RETURN_IF_ERROR(SaveParent(writer, input_impl_));
+ }
+ // Save cache.
+ TF_RETURN_IF_ERROR(
+ writer->WriteScalar(strings::StrCat("cache_size"), cache_.size()));
+ for (int64 i = 0; i < cache_.size(); i++) {
+ TF_RETURN_IF_ERROR(writer->WriteScalar(
+ strings::StrCat("cache[", i, "]_size"), cache_[i].size()));
+ for (int64 j = 0; j < cache_[i].size(); j++) {
+ TF_RETURN_IF_ERROR(writer->WriteTensor(
+ strings::StrCat("cache[", i, "][", j, "]"), cache_[i][j]));
+ }
+ }
+ return Status::OK();
+ }
+
+ Status RestoreInternal(IteratorContext* ctx,
+ IteratorStateReader* reader) override {
+ mutex_lock l(mu_);
+ if (!reader->Contains(full_name("input_impl_empty"))) {
+ TF_RETURN_IF_ERROR(RestoreParent(ctx, reader, input_impl_));
+ } else {
+ input_impl_.reset();
+ }
+ // Restore cache.
+ int64 cache_size;
+ TF_RETURN_IF_ERROR(
+ reader->ReadScalar(strings::StrCat("cache_size"), &cache_size));
+ cache_.resize(cache_size);
+ for (int64 i = 0; i < cache_size; i++) {
+ int64 vector_size;
+ TF_RETURN_IF_ERROR(reader->ReadScalar(
+ strings::StrCat("cache[", i, "]_size"), &vector_size));
+ cache_[i].resize(vector_size);
+ for (int64 j = 0; j < vector_size; j++) {
+ TF_RETURN_IF_ERROR(reader->ReadTensor(
+ strings::StrCat("cache[", i, "][", j, "]"), &cache_[i][j]));
+ }
+ }
+ return Status::OK();
+ }
+
+ private:
+ mutex mu_;
+ std::vector<std::vector<Tensor>> cache_ GUARDED_BY(mu_);
+ std::unique_ptr<IteratorBase> input_impl_ GUARDED_BY(mu_);
+ };
+
+ const int64 window_size_;
+ const int64 stride_;
+ const DatasetBase* const input_;
+ std::vector<PartialTensorShape> output_shapes_;
+ };
+};
+
+REGISTER_KERNEL_BUILDER(Name("SlideDataset").Device(DEVICE_CPU),
+ SlideDatasetOp);
+
+} // namespace
+
+} // namespace tensorflow
diff --git a/tensorflow/core/kernels/depthtospace_op.cc b/tensorflow/core/kernels/depthtospace_op.cc
index 39aa3e9..b74a09e 100644
--- a/tensorflow/core/kernels/depthtospace_op.cc
+++ b/tensorflow/core/kernels/depthtospace_op.cc
@@ -188,6 +188,9 @@
Name("DepthToSpace").Device(DEVICE_GPU).TypeConstraint<float>("T"),
DepthToSpaceOp<GPUDevice, float>);
REGISTER_KERNEL_BUILDER(
+ Name("DepthToSpace").Device(DEVICE_GPU).TypeConstraint<Eigen::half>("T"),
+ DepthToSpaceOp<GPUDevice, Eigen::half>);
+REGISTER_KERNEL_BUILDER(
Name("DepthToSpace").Device(DEVICE_GPU).TypeConstraint<qint8>("T"),
DepthToSpaceOp<GPUDevice, qint8>);
#endif // GOOGLE_CUDA
diff --git a/tensorflow/core/kernels/depthtospace_op_gpu.cu.cc b/tensorflow/core/kernels/depthtospace_op_gpu.cu.cc
index 184c703..0656081 100644
--- a/tensorflow/core/kernels/depthtospace_op_gpu.cu.cc
+++ b/tensorflow/core/kernels/depthtospace_op_gpu.cu.cc
@@ -238,6 +238,12 @@
template struct functor::DepthToSpaceOpFunctor<GPUDevice, float, FORMAT_NCHW>;
template struct functor::DepthToSpaceOpFunctor<GPUDevice, float, FORMAT_NHWC>;
+// Instantiate the GPU implementations for Eigen::half.
+template struct functor::DepthToSpaceOpFunctor<GPUDevice, Eigen::half,
+ FORMAT_NCHW>;
+template struct functor::DepthToSpaceOpFunctor<GPUDevice, Eigen::half,
+ FORMAT_NHWC>;
+
// NCHW_VECT_C with 4 x qint8 can be treated as NCHW int32.
template struct functor::DepthToSpaceOpFunctor<GPUDevice, int32, FORMAT_NCHW>;
diff --git a/tensorflow/core/kernels/hexagon/BUILD b/tensorflow/core/kernels/hexagon/BUILD
index 108d59d..7688305 100644
--- a/tensorflow/core/kernels/hexagon/BUILD
+++ b/tensorflow/core/kernels/hexagon/BUILD
@@ -45,6 +45,7 @@
"//tensorflow/core:test_main",
"//tensorflow/core:testlib",
"//tensorflow/core/kernels:cwise_op",
+ "//tensorflow/core/kernels:quantization_utils",
"//tensorflow/core/kernels:quantized_ops",
"//tensorflow/core/kernels:reduction_ops",
"//tensorflow/core/kernels:remote_fused_graph_execute_utils",
diff --git a/tensorflow/core/kernels/mkl_conv_grad_filter_ops.cc b/tensorflow/core/kernels/mkl_conv_grad_filter_ops.cc
index 1401bc6..e070656 100644
--- a/tensorflow/core/kernels/mkl_conv_grad_filter_ops.cc
+++ b/tensorflow/core/kernels/mkl_conv_grad_filter_ops.cc
@@ -444,6 +444,7 @@
~MklConv2DCustomBackpropFilterOp() {}
private:
+ const int kDilationH = 0, kDilationW = 1;
void ValidateMklShapes(const MklDnnShape& input_mkl_shape,
const MklDnnShape& filter_mkl_shape,
const MklDnnShape& obp_mkl_shape) {
@@ -492,7 +493,9 @@
const convolution_forward::primitive_desc& conv_fwd_pd,
MklDnnData<T>* input, MklDnnData<T>* filter,
MklDnnData<T>* outbackprop, MklDnnData<T>* output,
- Tensor** output_tensor, const memory::dims& strides,
+ Tensor** output_tensor,
+ const memory::dims& strides,
+ const memory::dims& dilations,
const memory::dims& padding_l,
const memory::dims& padding_r, padding_kind padding,
const memory::dims& bwd_output_dims,
@@ -518,31 +521,32 @@
bias_grad->SetOpMemDesc(bias_grad_dims, memory::format::x);
}
- // Create convolution backward weights primitive.
- auto bwd_desc =
- (biasEnabled && (bias_grad != nullptr))
- ? convolution_backward_weights::desc(
- convolution_direct, input->GetOpMemDesc(),
- output->GetOpMemDesc(), bias_grad->GetOpMemDesc(),
- outbackprop->GetOpMemDesc(), strides, padding_l, padding_r,
- padding)
- : convolution_backward_weights::desc(
- convolution_direct, input->GetOpMemDesc(),
- output->GetOpMemDesc(), outbackprop->GetOpMemDesc(), strides,
- padding_l, padding_r, padding);
-
- auto bwd_pd = convolution_backward_weights::primitive_desc(
- bwd_desc, cpu_engine, conv_fwd_pd);
-
- // Allocate output tensor.
- AllocateOutputTensor(context, bwd_pd, bwd_output_dims, bwd_output_format,
- output_tensor);
-
- CHECK_NOTNULL(*output_tensor);
- // Set buffer handle using allocated output tensor.
- output->SetUsrMemDataHandle(*output_tensor);
-
if (biasEnabled && (bias_grad != nullptr)) {
+ // Create convolution backward weights with bias primitive.
+ // Use dilated convolution in case dilate rates are greater than zero.
+ auto bwd_desc = (dilations[kDilationH] > 0 || dilations[kDilationW] > 0) ?
+ convolution_backward_weights::desc(convolution_direct,
+ input->GetOpMemDesc(), output->GetOpMemDesc(),
+ bias_grad->GetOpMemDesc(),
+ outbackprop->GetOpMemDesc(), strides,
+ dilations, padding_l, padding_r, padding) :
+ convolution_backward_weights::desc(convolution_direct,
+ input->GetOpMemDesc(), output->GetOpMemDesc(),
+ bias_grad->GetOpMemDesc(),
+ outbackprop->GetOpMemDesc(),
+ strides, padding_l, padding_r, padding);
+ auto bwd_pd = convolution_backward_weights::primitive_desc(bwd_desc,
+ cpu_engine,
+ conv_fwd_pd);
+
+ // Allocate output tensor.
+ AllocateOutputTensor(context, bwd_pd, bwd_output_dims,
+ bwd_output_format, output_tensor);
+
+ CHECK_NOTNULL(*output_tensor);
+ // Set buffer handle using allocated output tensor.
+ output->SetUsrMemDataHandle(*output_tensor);
+
// Allocate bias_grad tensor
TensorShape bias_grad_shape({depth});
Tensor* bias_grad_tensor = nullptr;
@@ -553,11 +557,32 @@
memory::desc({bias_grad_dims}, MklDnnType<T>(), memory::format::x);
bias_grad->SetUsrMem(bias_grad_md, bias_grad_tensor);
bias_grad->SetUsrMemDataHandle(bias_grad_tensor);
- }
- if (biasEnabled && (bias_grad != nullptr)) {
- PrepareAndExecutePrimitive(bwd_pd, input, outbackprop, output, bias_grad);
+ PrepareAndExecutePrimitive(bwd_pd, input, outbackprop, output,
+ bias_grad);
} else {
+ // Create convolution backward weights primitive.
+ // Use dilated convolution in case dilate rates are greater than zero.
+ auto bwd_desc = (dilations[kDilationH] > 0 || dilations[kDilationW] > 0) ?
+ convolution_backward_weights::desc(convolution_direct,
+ input->GetOpMemDesc(), output->GetOpMemDesc(),
+ outbackprop->GetOpMemDesc(), strides,
+ dilations, padding_l, padding_r, padding) :
+ convolution_backward_weights::desc(convolution_direct,
+ input->GetOpMemDesc(), output->GetOpMemDesc(),
+ outbackprop->GetOpMemDesc(),
+ strides, padding_l, padding_r, padding);
+ auto bwd_pd = convolution_backward_weights::primitive_desc(bwd_desc,
+ cpu_engine,
+ conv_fwd_pd);
+
+ // Allocate output tensor.
+ AllocateOutputTensor(context, bwd_pd, bwd_output_dims,
+ bwd_output_format, output_tensor);
+
+ CHECK_NOTNULL(*output_tensor);
+ // Set buffer handle using allocated output tensor.
+ output->SetUsrMemDataHandle(*output_tensor);
PrepareAndExecutePrimitive(bwd_pd, input, outbackprop, output);
}
}
diff --git a/tensorflow/core/kernels/mkl_conv_grad_input_ops.cc b/tensorflow/core/kernels/mkl_conv_grad_input_ops.cc
index eeed009..d203c04 100644
--- a/tensorflow/core/kernels/mkl_conv_grad_input_ops.cc
+++ b/tensorflow/core/kernels/mkl_conv_grad_input_ops.cc
@@ -369,6 +369,7 @@
private:
const int kInputIndex_Filter = 1, kInputIndex_InputSizes = 0,
kInputIndex_OutBackProp = 2;
+ const int kDilationH = 0, kDilationW = 1;
void ValidateMklShapes(const MklDnnShape& input_mkl_shape,
const MklDnnShape& filter_mkl_shape,
const MklDnnShape& obp_mkl_shape) {
@@ -419,7 +420,9 @@
const convolution_forward::primitive_desc& conv_fwd_pd,
MklDnnData<T>* input, MklDnnData<T>* filter,
MklDnnData<T>* outbackprop, MklDnnData<T>* output,
- Tensor** output_tensor, const memory::dims& strides,
+ Tensor** output_tensor,
+ const memory::dims& strides,
+ const memory::dims& dilations,
const memory::dims& padding_l,
const memory::dims& padding_r, padding_kind padding,
const memory::dims& bwd_output_dims,
@@ -432,9 +435,16 @@
CHECK_NOTNULL(output_tensor);
// Create convolution backward data primitive.
- auto bwd_desc = convolution_backward_data::desc(
- convolution_direct, output->GetOpMemDesc(), filter->GetOpMemDesc(),
- outbackprop->GetOpMemDesc(), strides, padding_l, padding_r, padding);
+ // Use dilated convolution in case dilate rates are greater than zero.
+ auto bwd_desc = (dilations[kDilationH] > 0 || dilations[kDilationW] > 0) ?
+ convolution_backward_data::desc(convolution_direct,
+ output->GetOpMemDesc(), filter->GetOpMemDesc(),
+ outbackprop->GetOpMemDesc(), strides,
+ dilations, padding_l, padding_r, padding):
+ convolution_backward_data::desc(convolution_direct,
+ output->GetOpMemDesc(), filter->GetOpMemDesc(),
+ outbackprop->GetOpMemDesc(),
+ strides, padding_l, padding_r, padding);
auto bwd_pd = convolution_backward_data::primitive_desc(
bwd_desc, cpu_engine, conv_fwd_pd);
diff --git a/tensorflow/core/kernels/mkl_conv_ops.cc b/tensorflow/core/kernels/mkl_conv_ops.cc
index 1440da8..f0818eb 100644
--- a/tensorflow/core/kernels/mkl_conv_ops.cc
+++ b/tensorflow/core/kernels/mkl_conv_ops.cc
@@ -493,6 +493,7 @@
~MklConv2DOp() {}
explicit MklConv2DOp(OpKernelConstruction* context) : OpKernel(context) {
+ OP_REQUIRES_OK(context, context->GetAttr("dilations", &dilations_));
OP_REQUIRES_OK(context, context->GetAttr("strides", &strides_));
string data_format;
OP_REQUIRES_OK(context, context->GetAttr("data_format", &data_format));
@@ -509,6 +510,20 @@
errors::InvalidArgument("Current implementation does not yet support "
"strides in the batch and depth dimensions."));
OP_REQUIRES_OK(context, context->GetAttr("padding", &padding_));
+ OP_REQUIRES(context, dilations_.size() == 4,
+ errors::InvalidArgument("Sliding window dilations field must "
+ "specify 4 dimensions"));
+ const int64 dilation_n = GetTensorDim(dilations_, data_format_, 'N');
+ const int64 dilation_c = GetTensorDim(dilations_, data_format_, 'C');
+ const int64 dilation_h = GetTensorDim(dilations_, data_format_, 'H');
+ const int64 dilation_w = GetTensorDim(dilations_, data_format_, 'W');
+ OP_REQUIRES(context, dilation_n == 1 && dilation_c == 1,
+ errors::InvalidArgument(
+ "Current implementation does not yet support "
+ "dilations in the batch and depth dimensions."));
+ OP_REQUIRES(
+ context, dilation_h > 0 && dilation_w > 0,
+ errors::InvalidArgument("Dilated rates should be larger than 0."));
}
void Compute(OpKernelContext* context) override {
@@ -530,17 +545,19 @@
MklDnnData<T> filter(&cpu_engine);
MklDnnData<T> output(&cpu_engine);
- memory::dims src_dims, filter_dims, padding_l, padding_r, strides;
+ memory::dims src_dims, filter_dims, padding_l, padding_r,
+ dilations, strides;
memory::dims output_dims_tf_order, output_dims_mkl_order;
// Get shapes of input tensors in MKL-DNN order
- MklDnnConvUtil conv_utl(context, strides_, padding_, data_format_);
+ MklDnnConvUtil conv_utl(context, strides_, padding_, data_format_,
+ dilations_);
auto src_tf_shape = GetTfShape(context, kInputIndex_Src);
auto filter_tf_shape = GetTfShape(context, kInputIndex_Filter);
conv_utl.GetConvFwdSizesInMklOrder(
src_tf_shape, filter_tf_shape, &src_dims, &filter_dims, &strides,
- &output_dims_tf_order, &output_dims_mkl_order, &padding_l,
- &padding_r);
+ &dilations, &output_dims_tf_order, &output_dims_mkl_order,
+ &padding_l, &padding_r);
if (!context->status().ok()) return;
// Check for corner case - if there is nothing to compute, return.
@@ -553,6 +570,7 @@
// Need semantics for Null MKL tensor
MklDnnShape output_mkl_shape;
output_mkl_shape.SetMklTensor(false);
+
AllocateOutputSetMklShape(context, kOutputIndex_Dst, &output_tensor,
src_tf_shape, output_mkl_shape);
@@ -596,55 +614,79 @@
filter.SetOpMemDesc(filter_dims, memory::format::any);
output.SetOpMemDesc(output_dims_mkl_order, memory::format::any);
- // If bias is enabled, then do the same steps as above for bias.
+ // MKLDNN dilation starts from 0.
+ dilations[kDilationH] -= 1;
+ dilations[kDilationW] -= 1;
+
if (biasEnabled) {
- MklDnnData<T> bias(&cpu_engine);
- memory::dims bias_size;
- conv_utl.GetBiasSizeInMklOrder(kInputIndex_Bias, &bias_size);
- const Tensor& bias_tensor = MklGetInput(context, kInputIndex_Bias);
- bias.SetUsrMem(bias_size, memory::format::x, &bias_tensor);
- bias.SetOpMemDesc(bias_size, memory::format::any);
+ // Create convolution primitive with Bias.
+ MklDnnData<T> bias(&cpu_engine);
+ memory::dims bias_size;
+ conv_utl.GetBiasSizeInMklOrder(kInputIndex_Bias, &bias_size);
+ const Tensor& bias_tensor = MklGetInput(context, kInputIndex_Bias);
+ bias.SetUsrMem(bias_size, memory::format::x, &bias_tensor);
+ bias.SetOpMemDesc(bias_size, memory::format::any);
- // Create convolution primitive with Bias.
- auto conv_desc = convolution_forward::desc(
- prop_kind::forward, convolution_direct, src.GetOpMemDesc(),
- filter.GetOpMemDesc(), bias.GetOpMemDesc(), output.GetOpMemDesc(),
- strides, padding_l, padding_r, TFPaddingToMklDnnPadding(padding_));
+ // Create convolution primitive with Bias.
+ // Use MKLDNN dilated convolution in case of dilated rate (>0).
+ auto conv_desc = (dilations[kDilationH] > 0 ||
+ dilations[kDilationW] > 0) ?
+ convolution_forward::desc(prop_kind::forward,
+ convolution_direct, src.GetOpMemDesc(),
+ filter.GetOpMemDesc(), bias.GetOpMemDesc(),
+ output.GetOpMemDesc(), strides, dilations,
+ padding_l, padding_r,
+ TFPaddingToMklDnnPadding(padding_)):
+ convolution_forward::desc(prop_kind::forward,
+ convolution_direct, src.GetOpMemDesc(),
+ filter.GetOpMemDesc(), bias.GetOpMemDesc(),
+ output.GetOpMemDesc(), strides,
+ padding_l, padding_r,
+ TFPaddingToMklDnnPadding(padding_));
- auto conv_prim_desc =
- convolution_forward::primitive_desc(conv_desc, cpu_engine);
- AllocateOutputTensor(context, conv_prim_desc, output_dims_mkl_order,
- tf_fmt, &output_tensor);
- // Set data handle for output.
- output.SetUsrMemDataHandle(output_tensor);
+ auto conv_prim_desc = convolution_forward::primitive_desc(conv_desc,
+ cpu_engine);
+ AllocateOutputTensor(context, conv_prim_desc,
+ output_dims_mkl_order, tf_fmt, &output_tensor);
+ // Set data handle for output.
+ output.SetUsrMemDataHandle(output_tensor);
- Tensor* filter_out_tensor = nullptr;
- AllocateFilterOutputTensor(context, conv_prim_desc,
- TFShapeToMklDnnDims(filter_tf_shape),
- &filter_out_tensor);
+ Tensor* filter_out_tensor = nullptr;
+ AllocateFilterOutputTensor(context, conv_prim_desc,
+ TFShapeToMklDnnDims(filter_tf_shape),
+ &filter_out_tensor);
- PrepareAndExecuteNet(conv_prim_desc, &src, &filter, &bias, &output,
- filter_out_tensor);
+ PrepareAndExecuteNet(conv_prim_desc, &src, &filter, &bias, &output,
+ filter_out_tensor);
} else {
- // Create convolution primitive without Bias.
- auto conv_desc = convolution_forward::desc(
- prop_kind::forward, convolution_direct, src.GetOpMemDesc(),
- filter.GetOpMemDesc(), output.GetOpMemDesc(), strides, padding_l,
- padding_r, TFPaddingToMklDnnPadding(padding_));
+ // Create convolution primitive without Bias.
+ // Use MKLDNN dilated convolution in case of dilated rate (>0).
+ auto conv_desc = (dilations[kDilationH] > 0 ||
+ dilations[kDilationW] > 0) ?
+ convolution_forward::desc(prop_kind::forward,
+ convolution_direct, src.GetOpMemDesc(),
+ filter.GetOpMemDesc(), output.GetOpMemDesc(),
+ strides, dilations, padding_l, padding_r,
+ TFPaddingToMklDnnPadding(padding_)):
+ convolution_forward::desc(prop_kind::forward,
+ convolution_direct, src.GetOpMemDesc(),
+ filter.GetOpMemDesc(), output.GetOpMemDesc(),
+ strides, padding_l, padding_r,
+ TFPaddingToMklDnnPadding(padding_));
- auto conv_prim_desc =
- convolution_forward::primitive_desc(conv_desc, cpu_engine);
- AllocateOutputTensor(context, conv_prim_desc, output_dims_mkl_order,
- tf_fmt, &output_tensor);
- // Set data handle for output.
- output.SetUsrMemDataHandle(output_tensor);
+ auto conv_prim_desc = convolution_forward::primitive_desc(conv_desc,
+ cpu_engine);
+ AllocateOutputTensor(context, conv_prim_desc, output_dims_mkl_order,
+ tf_fmt, &output_tensor);
+ // Set data handle for output.
+ output.SetUsrMemDataHandle(output_tensor);
- Tensor* filter_out_tensor = nullptr;
- AllocateFilterOutputTensor(context, conv_prim_desc,
- TFShapeToMklDnnDims(filter_tf_shape),
- &filter_out_tensor);
- PrepareAndExecuteNet(conv_prim_desc, &src, &filter, nullptr, &output,
- filter_out_tensor);
+ Tensor* filter_out_tensor = nullptr;
+ AllocateFilterOutputTensor(context, conv_prim_desc,
+ TFShapeToMklDnnDims(filter_tf_shape),
+ &filter_out_tensor);
+ PrepareAndExecuteNet(conv_prim_desc, &src, &filter,
+ nullptr, &output, filter_out_tensor);
}
} catch (mkldnn::error& e) {
string error_msg = "Status: " + std::to_string(e.status) +
@@ -658,10 +700,12 @@
private:
std::vector<int32> strides_;
+ std::vector<int32> dilations_;
Padding padding_;
TensorFormat data_format_;
const int kInputIndex_Src = 0, kInputIndex_Filter = 1, kInputIndex_Bias = 2;
const int kOutputIndex_Dst = 0, kOutputIndex_Filter = 1;
+ const int kDilationH = 0, kDilationW = 1;
// Allocate output tensor.
void AllocateOutputTensor(
diff --git a/tensorflow/core/kernels/mkl_conv_ops.h b/tensorflow/core/kernels/mkl_conv_ops.h
index 9dd8822..7ca10db 100644
--- a/tensorflow/core/kernels/mkl_conv_ops.h
+++ b/tensorflow/core/kernels/mkl_conv_ops.h
@@ -58,13 +58,16 @@
protected:
OpKernelContext* context_; // We don't own this.
std::vector<int32> strides_;
+ std::vector<int32> dilations_;
Padding padding_;
TensorFormat data_format_;
public:
MklDnnConvUtil(OpKernelContext* context, const std::vector<int32>& strides,
- Padding pad, TensorFormat fm)
- : context_(context), strides_(strides), padding_(pad), data_format_(fm) {}
+ Padding pad, TensorFormat fm,
+ const std::vector<int32>& dilations) :
+ context_(context), strides_(strides), padding_(pad),
+ data_format_(fm), dilations_(dilations) {}
virtual ~MklDnnConvUtil() { context_ = nullptr; }
@@ -78,6 +81,16 @@
*strides = {stride_rows, stride_cols};
}
+ // Calculate Convolution dilations
+ virtual inline void GetDilationsInMklOrder(memory::dims *dilations) {
+ // For now we take the dilation from the second and third dimensions only
+ // (we do not support dilation on the batch or depth dimension).
+ CHECK_NOTNULL(dilations);
+ int dilations_rows = GetTensorDim(dilations_, data_format_, 'H');
+ int dilations_cols = GetTensorDim(dilations_, data_format_, 'W');
+ *dilations = {dilations_rows, dilations_cols};
+ }
+
// Calculate Convolution input size in MKL-DNN order. MKL-DNN
// requires input in NCHW format. Function does not return anything.
// But errors arising from sanity checks are returned in context's
@@ -213,7 +226,8 @@
// TODO(nhasabni): Add similar function for input and filter in MklShape.
virtual inline void GetOutputAndPadSizeInMklOrder(
const TensorShape& input_shape, const TensorShape& filter_shape,
- const memory::dims& strides, memory::dims* output_dims_tf_order,
+ const memory::dims& strides, const memory::dims& dilations,
+ memory::dims* output_dims_tf_order,
memory::dims* output_dims_mkl_order, memory::dims* pad_l,
memory::dims* pad_r) {
CHECK_NOTNULL(output_dims_tf_order);
@@ -232,6 +246,8 @@
// Stride is vector of 2 elements: {s_r, s_c}
int stride_rows = strides[0];
int stride_cols = strides[1];
+ int dilation_rows = dilations[0];
+ int dilation_cols = dilations[1];
// Output batch is same as input batch.
int out_batch = GetTensorDim(input_shape, data_format_, 'N');
@@ -241,11 +257,13 @@
int64 out_rows = 0, out_cols = 0;
int64 pad_top = 0, pad_bottom = 0, pad_left, pad_right;
- OP_REQUIRES_OK(context_, GetWindowedOutputSizeVerbose(
- input_rows, filter_rows, stride_rows, padding_,
+ OP_REQUIRES_OK(context_,
+ GetWindowedOutputSizeVerboseV2(input_rows, filter_rows,
+ dilation_rows, stride_rows, padding_,
&out_rows, &pad_top, &pad_bottom));
- OP_REQUIRES_OK(context_, GetWindowedOutputSizeVerbose(
- input_cols, filter_cols, stride_cols, padding_,
+ OP_REQUIRES_OK(context_,
+ GetWindowedOutputSizeVerboseV2(input_cols, filter_cols,
+ dilation_cols, stride_cols, padding_,
&out_cols, &pad_left, &pad_right));
// Tensorflow output is in data_format order. (NHWC or NCHW)
@@ -271,7 +289,8 @@
//
// Function does not return anything, but sets error in context status.
inline void GetOutputAndPadSizeInMklOrder(
- size_t src_index, size_t filter_index, const memory::dims& strides,
+ size_t src_index, size_t filter_index,
+ const memory::dims& strides, const memory::dims& dilations,
memory::dims* output_dims_tf_order, memory::dims* output_dims_mkl_order,
memory::dims* pad_l, memory::dims* pad_r) {
CHECK_NOTNULL(output_dims_tf_order);
@@ -286,9 +305,9 @@
errors::InvalidArgument("input must be 4-dimensional",
input_tf_shape.DebugString()));
- GetOutputAndPadSizeInMklOrder(input_tf_shape, filter_tf_shape, strides,
- output_dims_tf_order, output_dims_mkl_order,
- pad_l, pad_r);
+ GetOutputAndPadSizeInMklOrder(input_tf_shape, filter_tf_shape,
+ strides, dilations, output_dims_tf_order,
+ output_dims_mkl_order, pad_l, pad_r);
}
// Wrapper function to calculate input, filter, and output sizes of
@@ -300,12 +319,14 @@
inline void GetConvFwdSizesInMklOrder(
const TensorShape& input_shape, const TensorShape& filter_shape,
memory::dims* input_dims, memory::dims* filter_dims,
- memory::dims* strides, memory::dims* output_dims_tf_order,
+ memory::dims* strides, memory::dims *dilations,
+ memory::dims* output_dims_tf_order,
memory::dims* output_dims_mkl_order, memory::dims* pad_l,
memory::dims* pad_r) {
CHECK_NOTNULL(input_dims);
CHECK_NOTNULL(filter_dims);
CHECK_NOTNULL(strides);
+ CHECK_NOTNULL(dilations);
CHECK_NOTNULL(output_dims_tf_order);
CHECK_NOTNULL(output_dims_mkl_order);
CHECK_NOTNULL(pad_l);
@@ -316,7 +337,9 @@
GetFilterSizeInMklOrder(input_shape, filter_shape, filter_dims);
if (!context_->status().ok()) return;
GetStridesInMklOrder(strides);
- GetOutputAndPadSizeInMklOrder(input_shape, filter_shape, *strides,
+ GetDilationsInMklOrder(dilations);
+ GetOutputAndPadSizeInMklOrder(input_shape, filter_shape,
+ *strides, *dilations,
output_dims_tf_order, output_dims_mkl_order,
pad_l, pad_r);
if (!context_->status().ok()) return;
@@ -344,7 +367,21 @@
context, (stride_n == 1 && stride_c == 1),
errors::InvalidArgument("Current implementation does not yet support "
"strides in the batch and depth dimensions."));
-
+ OP_REQUIRES_OK(context, context->GetAttr("dilations", &dilations_));
+ OP_REQUIRES(context, dilations_.size() == 4,
+ errors::InvalidArgument("Sliding window dilations field must "
+ "specify 4 dimensions"));
+ int dilation_n = GetTensorDim(dilations_, data_format_, 'N');
+ int dilation_c = GetTensorDim(dilations_, data_format_, 'C');
+ int dilation_h = GetTensorDim(dilations_, data_format_, 'H');
+ int dilation_w = GetTensorDim(dilations_, data_format_, 'W');
+ OP_REQUIRES(context, (dilation_n == 1 && dilation_c == 1),
+ errors::InvalidArgument(
+ "Current implementation does not yet support "
+ "dilations in the batch and depth dimensions."));
+ OP_REQUIRES(
+ context, dilation_h > 0 && dilation_w > 0,
+ errors::InvalidArgument("Dilated rates should be larger than 0."));
OP_REQUIRES_OK(context, context->GetAttr("padding", &padding_));
}
@@ -406,15 +443,16 @@
// By default, all dims are in MKL order. Only dims in TF order
// are those with prefix tf_order.
memory::dims outbprop_dims, fwd_input_dims, fwd_filter_dims;
- memory::dims padding_l, padding_r, strides, fwd_output_dims;
+ memory::dims padding_l, padding_r, dilations, strides, fwd_output_dims;
memory::dims fwd_output_dims_tf_order;
// Get forward convolution parameters.
- MklDnnConvUtil conv_utl(context, strides_, padding_, data_format_);
+ MklDnnConvUtil conv_utl(context, strides_, padding_, data_format_,
+ dilations_);
conv_utl.GetConvFwdSizesInMklOrder(
input_tf_shape, filter_tf_shape, &fwd_input_dims, &fwd_filter_dims,
- &strides, &fwd_output_dims_tf_order, &fwd_output_dims, &padding_l,
- &padding_r);
+ &strides, &dilations, &fwd_output_dims_tf_order, &fwd_output_dims,
+ &padding_l, &padding_r);
if (!context->status().ok()) return;
// Create Convolution forward descriptor since Convolution backward
@@ -437,10 +475,21 @@
memory::format::hwio);
// Tensorflow Output of Conv2D is in data_format order.
auto fwd_out_md = memory::desc(fwd_output_dims, MklDnnType<T>(), tf_fmt);
- auto fwd_desc = convolution_forward::desc(
- prop_kind::forward, convolution_direct, fwd_input_md, fwd_filter_md,
- fwd_out_md, strides, padding_l, padding_r,
- TFPaddingToMklDnnPadding(padding_));
+
+ const int kDilationH = 0, kDilationW = 1;
+ dilations[kDilationH] -= 1;
+ dilations[kDilationW] -= 1;
+ auto fwd_desc = (dilations[kDilationH] > 0 || dilations[kDilationW] > 0)?
+ convolution_forward::desc(prop_kind::forward,
+ convolution_direct, fwd_input_md,
+ fwd_filter_md, fwd_out_md,
+ strides, dilations, padding_l, padding_r,
+ TFPaddingToMklDnnPadding(padding_)) :
+ convolution_forward::desc(prop_kind::forward,
+ convolution_direct, fwd_input_md,
+ fwd_filter_md, fwd_out_md,
+ strides, padding_l, padding_r,
+ TFPaddingToMklDnnPadding(padding_));
auto fwd_pd = convolution_forward::primitive_desc(fwd_desc, cpu_engine);
// Create memory for user data. Describe how the inputs and outputs of
@@ -485,8 +534,9 @@
// Operator-specific call to create and execute primitive.
CreatePrimitive(context, cpu_engine, fwd_pd, &input, &filter,
- &outbackprop, &output, &output_tensor, strides, padding_l,
- padding_r, TFPaddingToMklDnnPadding(padding_),
+ &outbackprop, &output, &output_tensor,
+ strides, dilations, padding_l, padding_r,
+ TFPaddingToMklDnnPadding(padding_),
bwd_output_dims, bwd_output_format);
} catch (mkldnn::error& e) {
string error_msg = "Status: " + std::to_string(e.status) +
@@ -535,20 +585,21 @@
virtual memory::format GetOutputFormat(const memory::format data_format) = 0;
/// Create and execute the primitive storing output in the output_tensor.
- virtual void CreatePrimitive(
- OpKernelContext* context, const engine& cpu_engine,
- const convolution_forward::primitive_desc& conv_fwd_pd,
- MklDnnData<T>* input, MklDnnData<T>* filter, MklDnnData<T>* outbackprop,
- MklDnnData<T>* output, Tensor** output_tensor,
- const memory::dims& strides, const memory::dims& padding_l,
- const memory::dims& padding_r, padding_kind padding,
- const memory::dims& bwd_output_dims,
- memory::format bwd_output_format) = 0;
+ virtual void CreatePrimitive(OpKernelContext* context,
+ const engine& cpu_engine,
+ const convolution_forward::primitive_desc& conv_fwd_pd,
+ MklDnnData<T>* input, MklDnnData<T>* filter, MklDnnData<T>* outbackprop,
+ MklDnnData<T>* output, Tensor** output_tensor, const memory::dims& strides,
+ const memory::dims& dilations, const memory::dims& padding_l,
+ const memory::dims& padding_r, padding_kind padding,
+ const memory::dims& bwd_output_dims,
+ memory::format bwd_output_format) = 0;
// Get the data_format {NCHW, NHWC}
TensorFormat GetTFDataFormat() { return data_format_; }
private:
+ std::vector<int32> dilations_;
std::vector<int32> strides_;
Padding padding_;
TensorFormat data_format_;
diff --git a/tensorflow/core/kernels/mkl_input_conversion_op.cc b/tensorflow/core/kernels/mkl_input_conversion_op.cc
index e9a2376..d91f710 100644
--- a/tensorflow/core/kernels/mkl_input_conversion_op.cc
+++ b/tensorflow/core/kernels/mkl_input_conversion_op.cc
@@ -442,12 +442,11 @@
auto input_tf_md = mkl_output_mkl_shape.GetTfLayout();
tf_input.SetUsrMem(input_tf_md, tf_tensor);
- // Create reorder between tensorflow layout and Mkl layout.
+ // Create reorder between tensorflow layout and Mkl layout if necessary
std::vector<primitive> net;
- CHECK_EQ(tf_input.CheckReorderToOpMem(
+ tf_input.CheckReorderToOpMem(
memory::primitive_desc(output_mkl_md, cpu_engine),
- tensor_out, &net),
- true);
+ tensor_out, &net);
stream(stream::kind::eager).submit(net).wait();
// -- The tensor in MKL format passes through --
diff --git a/tensorflow/core/kernels/mkl_relu_op.cc b/tensorflow/core/kernels/mkl_relu_op.cc
index 267f4f8..0a0f695 100644
--- a/tensorflow/core/kernels/mkl_relu_op.cc
+++ b/tensorflow/core/kernels/mkl_relu_op.cc
@@ -392,7 +392,7 @@
Tensor* dst_tensor = nullptr;
if (src_tensor.dims() == 0) {
- Compute_Scalar(context);
+ Compute_Scalar(context); // scalar case doesn't use in-place operation
return;
}
@@ -437,11 +437,15 @@
dnn_shape_dst.SetMklTensor(false);
tf_shape_dst = src_tensor.shape();
}
- AllocateOutputSetMklShape(context, dst_index, &dst_tensor, tf_shape_dst,
- dnn_shape_dst);
+
+ // Allocate output and MklDnnShape tensors separately for possible
+ // in-place operation
+ OP_REQUIRES_OK(context, context->forward_input_or_allocate_output(
+ {src_index}, dst_index, tf_shape_dst, &dst_tensor));
+ AllocateOutputSetMklShape(context, dst_index, dnn_shape_dst);
// Destination memory descriptor is same as source memory descriptor.
- auto dst_md = src_md;
+ auto &dst_md = src_md;
dst.SetUsrMem(dst_md, dst_tensor);
// execute net
@@ -492,7 +496,7 @@
int src_dims_size = src_tensor.dims();
if (src_dims_size == 0) {
- Compute_Scalar(context);
+ Compute_Scalar(context); // scalar case doesn't use in-place operation
return;
}
@@ -603,8 +607,13 @@
// so it is ok to get TensorFlow shape.
tf_shape_diff_src = src_tensor.shape();
}
- AllocateOutputSetMklShape(context, diff_src_index, &diff_src_tensor,
- tf_shape_diff_src, dnn_shape_diff_src);
+
+ // Allocate diff_src and MklDnnShape tensors separately for possible
+ // in-place operation
+ OP_REQUIRES_OK(context, context->forward_input_or_allocate_output(
+ {diff_dst_index}, diff_src_index, tf_shape_diff_src,
+ &diff_src_tensor));
+ AllocateOutputSetMklShape(context, diff_src_index, dnn_shape_diff_src);
// diff_src memory descriptor is same as memory descriptor for both
// inputs.
diff --git a/tensorflow/core/kernels/segment_reduction_ops.h b/tensorflow/core/kernels/segment_reduction_ops.h
index 4abfbfb..d0703d7 100644
--- a/tensorflow/core/kernels/segment_reduction_ops.h
+++ b/tensorflow/core/kernels/segment_reduction_ops.h
@@ -13,8 +13,16 @@
limitations under the License.
==============================================================================*/
-#ifndef TENSORFLOW_CORE_KERNELS_SEGMENT_REDUCTION_OPS_H_
-#define TENSORFLOW_CORE_KERNELS_SEGMENT_REDUCTION_OPS_H_
+#ifndef THIRD_PARTY_TENSORFLOW_CORE_KERNELS_SEGMENT_REDUCTION_OPS_H_
+#define THIRD_PARTY_TENSORFLOW_CORE_KERNELS_SEGMENT_REDUCTION_OPS_H_
+
+
+// This file requires the following include because it uses CudaAtomicMax:
+// #include "tensorflow/core/util/cuda_kernel_helper.h"
+
+// Unfortunately we can't add the #include, since it breaks compilation for
+// non-GPU targets. This only breaks in clang, because it's more strict for
+// template code and CudaAtomicMax is used in template context.
// This file requires the following include because it uses CudaAtomicMax:
// #include "tensorflow/core/util/cuda_kernel_helper.h"
@@ -130,4 +138,4 @@
} // namespace functor
} // namespace tensorflow
-#endif // TENSORFLOW_CORE_KERNELS_SEGMENT_REDUCTION_OPS_H_
+#endif // THIRD_PARTY_TENSORFLOW_CORE_KERNELS_SEGMENT_REDUCTION_OPS_H_
diff --git a/tensorflow/core/kernels/spacetodepth_op.cc b/tensorflow/core/kernels/spacetodepth_op.cc
index 23df1c3..e59adfc 100644
--- a/tensorflow/core/kernels/spacetodepth_op.cc
+++ b/tensorflow/core/kernels/spacetodepth_op.cc
@@ -188,6 +188,9 @@
Name("SpaceToDepth").Device(DEVICE_GPU).TypeConstraint<float>("T"),
SpaceToDepthOp<GPUDevice, float>);
REGISTER_KERNEL_BUILDER(
+ Name("SpaceToDepth").Device(DEVICE_GPU).TypeConstraint<Eigen::half>("T"),
+ SpaceToDepthOp<GPUDevice, Eigen::half>);
+REGISTER_KERNEL_BUILDER(
Name("SpaceToDepth").Device(DEVICE_GPU).TypeConstraint<qint8>("T"),
SpaceToDepthOp<GPUDevice, qint8>);
#endif // GOOGLE_CUDA
diff --git a/tensorflow/core/kernels/spacetodepth_op_gpu.cu.cc b/tensorflow/core/kernels/spacetodepth_op_gpu.cu.cc
index db05ca1..f384597 100644
--- a/tensorflow/core/kernels/spacetodepth_op_gpu.cu.cc
+++ b/tensorflow/core/kernels/spacetodepth_op_gpu.cu.cc
@@ -234,6 +234,12 @@
template struct functor::SpaceToDepthOpFunctor<GPUDevice, float, FORMAT_NCHW>;
template struct functor::SpaceToDepthOpFunctor<GPUDevice, float, FORMAT_NHWC>;
+// Instantiate the GPU implementations for Eigen::half.
+template struct functor::SpaceToDepthOpFunctor<GPUDevice, Eigen::half,
+ FORMAT_NCHW>;
+template struct functor::SpaceToDepthOpFunctor<GPUDevice, Eigen::half,
+ FORMAT_NHWC>;
+
// NCHW_VECT_C with 4 x qint8 can be treated as NCHW int32.
template struct functor::SpaceToDepthOpFunctor<GPUDevice, int32, FORMAT_NCHW>;
diff --git a/tensorflow/core/lib/core/stringpiece.cc b/tensorflow/core/lib/core/stringpiece.cc
index 5bd7977..0b006fa 100644
--- a/tensorflow/core/lib/core/stringpiece.cc
+++ b/tensorflow/core/lib/core/stringpiece.cc
@@ -55,6 +55,4 @@
return StringPiece(data_ + pos, n);
}
-const StringPiece::size_type StringPiece::npos = size_type(-1);
-
} // namespace tensorflow
diff --git a/tensorflow/core/lib/core/stringpiece.h b/tensorflow/core/lib/core/stringpiece.h
index 910e4d9..2d00f71 100644
--- a/tensorflow/core/lib/core/stringpiece.h
+++ b/tensorflow/core/lib/core/stringpiece.h
@@ -65,7 +65,7 @@
iterator begin() const { return data_; }
iterator end() const { return data_ + size_; }
- static const size_t npos;
+ static const size_t npos = size_type(-1);
// Return the ith byte in the referenced data.
// REQUIRES: n < size()
diff --git a/tensorflow/core/lib/io/record_reader.cc b/tensorflow/core/lib/io/record_reader.cc
index 254fdf1..6de850b 100644
--- a/tensorflow/core/lib/io/record_reader.cc
+++ b/tensorflow/core/lib/io/record_reader.cc
@@ -205,7 +205,9 @@
if (options_.buffer_size > 0) {
TF_RETURN_IF_ERROR(input_stream_->SkipNBytes(offset));
}
+#if !defined(IS_SLIM_BUILD)
}
+#endif
return Status::OK();
} // namespace io
diff --git a/tensorflow/core/lib/io/record_reader.h b/tensorflow/core/lib/io/record_reader.h
index 62dd2ef..26278e0 100644
--- a/tensorflow/core/lib/io/record_reader.h
+++ b/tensorflow/core/lib/io/record_reader.h
@@ -16,10 +16,10 @@
#ifndef TENSORFLOW_LIB_IO_RECORD_READER_H_
#define TENSORFLOW_LIB_IO_RECORD_READER_H_
-#include "tensorflow/core/lib/core/status.h"
+#include "tensorflow/core/lib/core/errors.h"
#include "tensorflow/core/lib/core/stringpiece.h"
-#if !defined(IS_SLIM_BUILD)
#include "tensorflow/core/lib/io/inputstream_interface.h"
+#if !defined(IS_SLIM_BUILD)
#include "tensorflow/core/lib/io/zlib_compression_options.h"
#include "tensorflow/core/lib/io/zlib_inputstream.h"
#endif // IS_SLIM_BUILD
diff --git a/tensorflow/core/ops/dataset_ops.cc b/tensorflow/core/ops/dataset_ops.cc
index bdbbf6d..9a4b616 100644
--- a/tensorflow/core/ops/dataset_ops.cc
+++ b/tensorflow/core/ops/dataset_ops.cc
@@ -1,4 +1,4 @@
-/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
@@ -265,6 +265,16 @@
.Attr("output_shapes: list(shape) >= 1")
.SetShapeFn(shape_inference::ScalarShape);
+// TODO(mrry): move SlideDataset to contrib in the future.
+REGISTER_OP("SlideDataset")
+ .Input("input_dataset: variant")
+ .Input("window_size: int64")
+ .Input("stride: int64")
+ .Output("handle: variant")
+ .Attr("output_types: list(type) >= 1")
+ .Attr("output_shapes: list(shape) >= 1")
+ .SetShapeFn(shape_inference::ScalarShape);
+
REGISTER_OP("PaddedBatchDataset")
.Input("input_dataset: variant")
.Input("batch_size: int64")
diff --git a/tensorflow/core/ops/nn_ops.cc b/tensorflow/core/ops/nn_ops.cc
index 910fbac..d6a0f38 100644
--- a/tensorflow/core/ops/nn_ops.cc
+++ b/tensorflow/core/ops/nn_ops.cc
@@ -1498,6 +1498,7 @@
.Attr("use_cudnn_on_gpu: bool = true")
.Attr(GetPaddingAttrString())
.Attr(GetConvnetDataFormatAttrString())
+ .Attr("dilations: list(int) = [1, 1, 1, 1]")
.SetShapeFn(shape_inference::Conv2DShape)
.Doc(R"doc(
MKL version of Conv2D operator. Uses MKL DNN APIs to perform 2D convolution.
@@ -1516,6 +1517,7 @@
.Attr("use_cudnn_on_gpu: bool = true")
.Attr(GetPaddingAttrString())
.Attr(GetConvnetDataFormatAttrString())
+ .Attr("dilations: list(int) = [1, 1, 1, 1]")
.Doc(R"doc(
Dummy node that enables fusing Conv2D and BiasAdd operator for MKL. This node
does not perform anything. It is just created as an intermediate output of
@@ -1541,6 +1543,7 @@
.Attr("use_cudnn_on_gpu: bool = true")
.Attr(GetPaddingAttrString())
.Attr(GetConvnetDataFormatAttrString())
+ .Attr("dilations: list(int) = [1, 1, 1, 1]")
.Doc(R"doc(
MKL version of Conv2D and BiasAdd operator. Uses MKL DNN APIs to perform
2D convolution and add Bias to the output of convolution.
@@ -1563,6 +1566,7 @@
.Attr("use_cudnn_on_gpu: bool = true")
.Attr(GetPaddingAttrString())
.Attr(GetConvnetDataFormatAttrString())
+ .Attr("dilations: list(int) = [1, 1, 1, 1]")
.SetShapeFn([](InferenceContext* c) {
ShapeHandle s;
TF_RETURN_IF_ERROR(c->MakeShapeFromShapeTensor(1, &s));
@@ -1589,6 +1593,7 @@
.Attr("use_cudnn_on_gpu: bool = true")
.Attr(GetPaddingAttrString())
.Attr(GetConvnetDataFormatAttrString())
+ .Attr("dilations: list(int) = [1, 1, 1, 1]")
.SetShapeFn([](InferenceContext* c) {
ShapeHandle input_shape;
// Fetch the data_format attribute, which may not exist.
@@ -1633,6 +1638,7 @@
.Attr("use_cudnn_on_gpu: bool = true")
.Attr(GetPaddingAttrString())
.Attr(GetConvnetDataFormatAttrString())
+ .Attr("dilations: list(int) = [1, 1, 1, 1]")
.SetShapeFn([](InferenceContext* c) {
ShapeHandle input_shape;
// Fetch the data_format attribute, which may not exist.
@@ -1668,6 +1674,7 @@
.Attr("T: {half, float, double}")
.Attr("strides: list(int)")
.Attr(GetConvnetDataFormatAttrString())
+ .Attr("dilations: list(int) = [1, 1, 1, 1]")
.Doc(R"doc(
MKL version of Conv2DBackpropBias. Uses MKL DNN APIs to compute the
gradients of convolution with respect to the bias.
@@ -1690,6 +1697,7 @@
.Attr("use_cudnn_on_gpu: bool = true")
.Attr(GetPaddingAttrString())
.Attr(GetConvnetDataFormatAttrString())
+ .Attr("dilations: list(int) = [1, 1, 1, 1]")
.SetShapeFn([](InferenceContext* c) {
ShapeHandle s;
TF_RETURN_IF_ERROR(c->MakeShapeFromShapeTensor(0, &s));
diff --git a/tensorflow/core/platform/tracing.h b/tensorflow/core/platform/tracing.h
index 8f7bff1..eebbeae 100644
--- a/tensorflow/core/platform/tracing.h
+++ b/tensorflow/core/platform/tracing.h
@@ -103,7 +103,7 @@
friend class ScopedAnnotation;
friend class TraceMe;
- static std::atomic<Tracing::Engine*> tracing_engine_;
+ TF_EXPORT static std::atomic<Tracing::Engine*> tracing_engine_;
static Tracing::Engine* engine() {
return tracing_engine_.load(std::memory_order_acquire);
}
diff --git a/tensorflow/core/platform/windows/windows_file_system.cc b/tensorflow/core/platform/windows/windows_file_system.cc
index b6b3722..682e46e 100644
--- a/tensorflow/core/platform/windows/windows_file_system.cc
+++ b/tensorflow/core/platform/windows/windows_file_system.cc
@@ -382,7 +382,8 @@
Status WindowsFileSystem::FileExists(const string& fname) {
constexpr int kOk = 0;
- if (_access(TranslateName(fname).c_str(), kOk) == 0) {
+ std::wstring ws_translated_fname = Utf8ToWideChar(TranslateName(fname));
+ if (_waccess(ws_translated_fname.c_str(), kOk) == 0) {
return Status::OK();
}
return errors::NotFound(fname, " not found");
diff --git a/tensorflow/core/public/version.h b/tensorflow/core/public/version.h
index 7405e01..22f2c02 100644
--- a/tensorflow/core/public/version.h
+++ b/tensorflow/core/public/version.h
@@ -24,7 +24,7 @@
// TF_VERSION_SUFFIX is non-empty for pre-releases (e.g. "-alpha", "-alpha.1",
// "-beta", "-rc", "-rc.1")
-#define TF_VERSION_SUFFIX "-rc1"
+#define TF_VERSION_SUFFIX ""
#define TF_STR_HELPER(x) #x
#define TF_STR(x) TF_STR_HELPER(x)
diff --git a/tensorflow/docs_src/community/welcome.md b/tensorflow/docs_src/community/welcome.md
index 9f6fe91..6d0458e 100644
--- a/tensorflow/docs_src/community/welcome.md
+++ b/tensorflow/docs_src/community/welcome.md
@@ -51,6 +51,8 @@
TensorFlow provides multiple communication paths. To pick the right path,
please read the following list carefully:
+ * For new release announcements and security updates, subscribe to
+ [announce@tensorflow.org](https://groups.google.com/a/tensorflow.org/forum/#!forum/announce).
* To ask or answer technical questions about TensorFlow, use
[Stack Overflow](https://stackoverflow.com/questions/tagged/tensorflow).
For example, ask or search Stack Overflow about a particular error message
@@ -65,5 +67,5 @@
on GitHub. For example, use the issue tracker to request a
new operation in TensorFlow.
* To report vulnerabilities, please follow our
- [vulnerability disclosure guidelines](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/SECURITY.md).
+ [vulnerability disclosure guidelines](https://github.com/tensorflow/tensorflow/blob/master/SECURITY.md).
diff --git a/tensorflow/docs_src/install/install_c.md b/tensorflow/docs_src/install/install_c.md
index 8187985..0481c97 100644
--- a/tensorflow/docs_src/install/install_c.md
+++ b/tensorflow/docs_src/install/install_c.md
@@ -38,7 +38,7 @@
OS="linux" # Change to "darwin" for macOS
TARGET_DIRECTORY="/usr/local"
curl -L \
- "https://storage.googleapis.com/tensorflow/libtensorflow/libtensorflow-${TF_TYPE}-${OS}-x86_64-1.6.0-rc1.tar.gz" |
+ "https://storage.googleapis.com/tensorflow/libtensorflow/libtensorflow-${TF_TYPE}-${OS}-x86_64-1.6.0.tar.gz" |
sudo tar -C $TARGET_DIRECTORY -xz
The `tar` command extracts the TensorFlow C library into the `lib`
diff --git a/tensorflow/docs_src/install/install_go.md b/tensorflow/docs_src/install/install_go.md
index 4c6dfa8..8f89898 100644
--- a/tensorflow/docs_src/install/install_go.md
+++ b/tensorflow/docs_src/install/install_go.md
@@ -38,7 +38,7 @@
TF_TYPE="cpu" # Change to "gpu" for GPU support
TARGET_DIRECTORY='/usr/local'
curl -L \
- "https://storage.googleapis.com/tensorflow/libtensorflow/libtensorflow-${TF_TYPE}-$(go env GOOS)-x86_64-1.6.0-rc1.tar.gz" |
+ "https://storage.googleapis.com/tensorflow/libtensorflow/libtensorflow-${TF_TYPE}-$(go env GOOS)-x86_64-1.6.0.tar.gz" |
sudo tar -C $TARGET_DIRECTORY -xz
The `tar` command extracts the TensorFlow C library into the `lib`
diff --git a/tensorflow/docs_src/install/install_java.md b/tensorflow/docs_src/install/install_java.md
index 5278848..0ee9c84 100644
--- a/tensorflow/docs_src/install/install_java.md
+++ b/tensorflow/docs_src/install/install_java.md
@@ -36,7 +36,7 @@
<dependency>
<groupId>org.tensorflow</groupId>
<artifactId>tensorflow</artifactId>
- <version>1.6.0-rc1</version>
+ <version>1.6.0</version>
</dependency>
```
@@ -65,7 +65,7 @@
<dependency>
<groupId>org.tensorflow</groupId>
<artifactId>tensorflow</artifactId>
- <version>1.6.0-rc1</version>
+ <version>1.6.0</version>
</dependency>
</dependencies>
</project>
@@ -123,12 +123,12 @@
<dependency>
<groupId>org.tensorflow</groupId>
<artifactId>libtensorflow</artifactId>
- <version>1.6.0-rc1</version>
+ <version>1.6.0</version>
</dependency>
<dependency>
<groupId>org.tensorflow</groupId>
<artifactId>libtensorflow_jni_gpu</artifactId>
- <version>1.6.0-rc1</version>
+ <version>1.6.0</version>
</dependency>
```
@@ -147,7 +147,7 @@
Take the following steps to install TensorFlow for Java on Linux or macOS:
1. Download
- [libtensorflow.jar](https://storage.googleapis.com/tensorflow/libtensorflow/libtensorflow-1.6.0-rc1.jar),
+ [libtensorflow.jar](https://storage.googleapis.com/tensorflow/libtensorflow/libtensorflow-1.6.0.jar),
which is the TensorFlow Java Archive (JAR).
2. Decide whether you will run TensorFlow for Java on CPU(s) only or with
@@ -166,7 +166,7 @@
OS=$(uname -s | tr '[:upper:]' '[:lower:]')
mkdir -p ./jni
curl -L \
- "https://storage.googleapis.com/tensorflow/libtensorflow/libtensorflow_jni-${TF_TYPE}-${OS}-x86_64-1.6.0-rc1.tar.gz" |
+ "https://storage.googleapis.com/tensorflow/libtensorflow/libtensorflow_jni-${TF_TYPE}-${OS}-x86_64-1.6.0.tar.gz" |
tar -xz -C ./jni
### Install on Windows
@@ -174,10 +174,10 @@
Take the following steps to install TensorFlow for Java on Windows:
1. Download
- [libtensorflow.jar](https://storage.googleapis.com/tensorflow/libtensorflow/libtensorflow-1.6.0-rc1.jar),
+ [libtensorflow.jar](https://storage.googleapis.com/tensorflow/libtensorflow/libtensorflow-1.6.0.jar),
which is the TensorFlow Java Archive (JAR).
2. Download the following Java Native Interface (JNI) file appropriate for
- [TensorFlow for Java on Windows](https://storage.googleapis.com/tensorflow/libtensorflow/libtensorflow_jni-cpu-windows-x86_64-1.6.0-rc1.zip).
+ [TensorFlow for Java on Windows](https://storage.googleapis.com/tensorflow/libtensorflow/libtensorflow_jni-cpu-windows-x86_64-1.6.0.zip).
3. Extract this .zip file.
@@ -225,7 +225,7 @@
downloaded `.jar` in your `classpath` by using the `-cp` compilation flag
as follows:
-<pre><b>javac -cp libtensorflow-1.6.0-rc1.jar HelloTF.java</b></pre>
+<pre><b>javac -cp libtensorflow-1.6.0.jar HelloTF.java</b></pre>
### Running
@@ -239,11 +239,11 @@
For example, the following command line executes the `HelloTF` program on Linux
and macOS X:
-<pre><b>java -cp libtensorflow-1.6.0-rc1.jar:. -Djava.library.path=./jni HelloTF</b></pre>
+<pre><b>java -cp libtensorflow-1.6.0.jar:. -Djava.library.path=./jni HelloTF</b></pre>
And the following command line executes the `HelloTF` program on Windows:
-<pre><b>java -cp libtensorflow-1.6.0-rc1.jar;. -Djava.library.path=jni HelloTF</b></pre>d
+<pre><b>java -cp libtensorflow-1.6.0.jar;. -Djava.library.path=jni HelloTF</b></pre>
If the program prints <tt>Hello from <i>version</i></tt>, you've successfully
installed TensorFlow for Java and are ready to use the API. If the program
diff --git a/tensorflow/docs_src/install/install_linux.md b/tensorflow/docs_src/install/install_linux.md
index e3e115d..3e8744b 100644
--- a/tensorflow/docs_src/install/install_linux.md
+++ b/tensorflow/docs_src/install/install_linux.md
@@ -41,7 +41,8 @@
[NVIDIA's documentation](https://developer.nvidia.com/cudnn).
Ensure that you create the `CUDA_HOME` environment variable as
described in the NVIDIA documentation.
- * GPU card with CUDA Compute Capability 3.0 or higher. See
+ * GPU card with CUDA Compute Capability 3.0 or higher for building
+ from source and 3.5 or higher for our binaries. See
[NVIDIA documentation](https://developer.nvidia.com/cuda-gpus) for
a list of supported GPU cards.
* The libcupti-dev library, which is the NVIDIA CUDA Profile Tools Interface.
@@ -188,7 +189,7 @@
Virtualenv environment:
<pre>(tensorflow)$ <b>pip3 install --upgrade \
- https://storage.googleapis.com/tensorflow/linux/cpu/tensorflow-1.6.0rc1-cp34-cp34m-linux_x86_64.whl</b></pre>
+ https://storage.googleapis.com/tensorflow/linux/cpu/tensorflow-1.6.0-cp34-cp34m-linux_x86_64.whl</b></pre>
If you encounter installation problems, see
[Common Installation Problems](#common_installation_problems).
@@ -293,7 +294,7 @@
<pre>
$ <b>sudo pip3 install --upgrade \
- https://storage.googleapis.com/tensorflow/linux/cpu/tensorflow-1.6.0rc1-cp34-cp34m-linux_x86_64.whl</b>
+ https://storage.googleapis.com/tensorflow/linux/cpu/tensorflow-1.6.0-cp34-cp34m-linux_x86_64.whl</b>
</pre>
If this step fails, see
@@ -356,24 +357,23 @@
to 6006.
* <tt><i>TensorFlowCPUImage</i></tt> is required. It identifies the Docker
container. Specify one of the following values:
- * <tt>gcr.io/tensorflow/tensorflow</tt>, which is the TensorFlow CPU binary image.
- * <tt>gcr.io/tensorflow/tensorflow:latest-devel</tt>, which is the latest
+ * <tt>tensorflow/tensorflow</tt>, which is the TensorFlow CPU binary image.
+ * <tt>tensorflow/tensorflow:latest-devel</tt>, which is the latest
TensorFlow CPU Binary image plus source code.
- * <tt>gcr.io/tensorflow/tensorflow:<i>version</i></tt>, which is the
+ * <tt>tensorflow/tensorflow:<i>version</i></tt>, which is the
specified version (for example, 1.1.0rc1) of TensorFlow CPU binary image.
- * <tt>gcr.io/tensorflow/tensorflow:<i>version</i>-devel</tt>, which is
+ * <tt>tensorflow/tensorflow:<i>version</i>-devel</tt>, which is
the specified version (for example, 1.1.0rc1) of the TensorFlow GPU
binary image plus source code.
- <tt>gcr.io</tt> is the Google Container Registry. Note that some
- TensorFlow images are also available at
+ TensorFlow images are available at
[dockerhub](https://hub.docker.com/r/tensorflow/tensorflow/).
For example, the following command launches the latest TensorFlow CPU binary image
in a Docker container from which you can run TensorFlow programs in a shell:
<pre>
-$ <b>docker run -it gcr.io/tensorflow/tensorflow bash</b>
+$ <b>docker run -it tensorflow/tensorflow bash</b>
</pre>
The following command also launches the latest TensorFlow CPU binary image in a
@@ -381,7 +381,7 @@
programs in a Jupyter notebook:
<pre>
-$ <b>docker run -it -p 8888:8888 gcr.io/tensorflow/tensorflow</b>
+$ <b>docker run -it -p 8888:8888 tensorflow/tensorflow</b>
</pre>
Docker will download the TensorFlow binary image the first time you launch it.
@@ -405,14 +405,14 @@
<tt><i>hostPort</i></tt> and <code><em>containerPort</em></code> to `8888`.
* <i>TensorFlowGPUImage</i> specifies the Docker container. You must
specify one of the following values:
- * <tt>gcr.io/tensorflow/tensorflow:latest-gpu</tt>, which is the latest
+ * <tt>tensorflow/tensorflow:latest-gpu</tt>, which is the latest
TensorFlow GPU binary image.
- * <tt>gcr.io/tensorflow/tensorflow:latest-devel-gpu</tt>, which is
+ * <tt>tensorflow/tensorflow:latest-devel-gpu</tt>, which is
the latest TensorFlow GPU Binary image plus source code.
- * <tt>gcr.io/tensorflow/tensorflow:<i>version</i>-gpu</tt>, which is the
+ * <tt>tensorflow/tensorflow:<i>version</i>-gpu</tt>, which is the
specified version (for example, 0.12.1) of the TensorFlow GPU
binary image.
- * <tt>gcr.io/tensorflow/tensorflow:<i>version</i>-devel-gpu</tt>, which is
+ * <tt>tensorflow/tensorflow:<i>version</i>-devel-gpu</tt>, which is
the specified version (for example, 0.12.1) of the TensorFlow GPU
binary image plus source code.
@@ -421,7 +421,7 @@
Docker container from which you can run TensorFlow programs in a shell:
<pre>
-$ <b>nvidia-docker run -it gcr.io/tensorflow/tensorflow:latest-gpu bash</b>
+$ <b>nvidia-docker run -it tensorflow/tensorflow:latest-gpu bash</b>
</pre>
The following command also launches the latest TensorFlow GPU binary image
@@ -429,13 +429,13 @@
programs in a Jupyter notebook:
<pre>
-$ <b>nvidia-docker run -it -p 8888:8888 gcr.io/tensorflow/tensorflow:latest-gpu</b>
+$ <b>nvidia-docker run -it -p 8888:8888 tensorflow/tensorflow:latest-gpu</b>
</pre>
The following command installs an older TensorFlow version (0.12.1):
<pre>
-$ <b>nvidia-docker run -it -p 8888:8888 gcr.io/tensorflow/tensorflow:0.12.1-gpu</b>
+$ <b>nvidia-docker run -it -p 8888:8888 tensorflow/tensorflow:0.12.1-gpu</b>
</pre>
Docker will download the TensorFlow binary image the first time you launch it.
@@ -480,7 +480,7 @@
<pre>
(tensorflow)$ <b>pip install --ignore-installed --upgrade \
- https://storage.googleapis.com/tensorflow/linux/cpu/tensorflow-1.6.0rc1-cp34-cp34m-linux_x86_64.whl</b></pre>
+ https://storage.googleapis.com/tensorflow/linux/cpu/tensorflow-1.6.0-cp34-cp34m-linux_x86_64.whl</b></pre>
<a name="ValidateYourInstallation"></a>
## Validate your installation
@@ -505,7 +505,7 @@
from which you can run bash. For example:
<pre>
-$ <b>docker run -it gcr.io/tensorflow/tensorflow bash</b>
+$ <b>docker run -it tensorflow/tensorflow bash</b>
</pre>
@@ -647,14 +647,14 @@
CPU only:
<pre>
-https://storage.googleapis.com/tensorflow/linux/cpu/tensorflow-1.6.0rc1-cp27-none-linux_x86_64.whl
+https://storage.googleapis.com/tensorflow/linux/cpu/tensorflow-1.6.0-cp27-none-linux_x86_64.whl
</pre>
GPU support:
<pre>
-https://storage.googleapis.com/tensorflow/linux/gpu/tensorflow_gpu-1.6.0rc1-cp27-none-linux_x86_64.whl
+https://storage.googleapis.com/tensorflow/linux/gpu/tensorflow_gpu-1.6.0-cp27-none-linux_x86_64.whl
</pre>
Note that GPU support requires the NVIDIA hardware and software described in
@@ -666,14 +666,14 @@
CPU only:
<pre>
-https://storage.googleapis.com/tensorflow/linux/cpu/tensorflow-1.6.0rc1-cp34-cp34m-linux_x86_64.whl
+https://storage.googleapis.com/tensorflow/linux/cpu/tensorflow-1.6.0-cp34-cp34m-linux_x86_64.whl
</pre>
GPU support:
<pre>
-https://storage.googleapis.com/tensorflow/linux/gpu/tensorflow_gpu-1.6.0rc1-cp34-cp34m-linux_x86_64.whl
+https://storage.googleapis.com/tensorflow/linux/gpu/tensorflow_gpu-1.6.0-cp34-cp34m-linux_x86_64.whl
</pre>
Note that GPU support requires the NVIDIA hardware and software described in
@@ -685,14 +685,14 @@
CPU only:
<pre>
-https://storage.googleapis.com/tensorflow/linux/cpu/tensorflow-1.6.0rc1-cp35-cp35m-linux_x86_64.whl
+https://storage.googleapis.com/tensorflow/linux/cpu/tensorflow-1.6.0-cp35-cp35m-linux_x86_64.whl
</pre>
GPU support:
<pre>
-https://storage.googleapis.com/tensorflow/linux/gpu/tensorflow_gpu-1.6.0rc1-cp35-cp35m-linux_x86_64.whl
+https://storage.googleapis.com/tensorflow/linux/gpu/tensorflow_gpu-1.6.0-cp35-cp35m-linux_x86_64.whl
</pre>
@@ -704,14 +704,14 @@
CPU only:
<pre>
-https://storage.googleapis.com/tensorflow/linux/cpu/tensorflow-1.6.0rc1-cp36-cp36m-linux_x86_64.whl
+https://storage.googleapis.com/tensorflow/linux/cpu/tensorflow-1.6.0-cp36-cp36m-linux_x86_64.whl
</pre>
GPU support:
<pre>
-https://storage.googleapis.com/tensorflow/linux/gpu/tensorflow_gpu-1.6.0rc1-cp36-cp36m-linux_x86_64.whl
+https://storage.googleapis.com/tensorflow/linux/gpu/tensorflow_gpu-1.6.0-cp36-cp36m-linux_x86_64.whl
</pre>
diff --git a/tensorflow/docs_src/install/install_mac.md b/tensorflow/docs_src/install/install_mac.md
index 623ca6b..205db8e 100644
--- a/tensorflow/docs_src/install/install_mac.md
+++ b/tensorflow/docs_src/install/install_mac.md
@@ -119,7 +119,7 @@
TensorFlow in the active Virtualenv is as follows:
<pre> $ <b>pip install --upgrade \
- https://storage.googleapis.com/tensorflow/mac/cpu/tensorflow-1.6.0rc1-py2-none-any.whl</b></pre>
+ https://storage.googleapis.com/tensorflow/mac/cpu/tensorflow-1.6.0-py2-none-any.whl</b></pre>
If you encounter installation problems, see
[Common Installation Problems](#common-installation-problems).
@@ -242,7 +242,7 @@
issue the following command:
<pre> $ <b>sudo pip install --upgrade \
- https://storage.googleapis.com/tensorflow/mac/cpu/tensorflow-1.6.0rc1-py2-none-any.whl</b> </pre>
+ https://storage.googleapis.com/tensorflow/mac/cpu/tensorflow-1.6.0-py2-none-any.whl</b> </pre>
If the preceding command fails, see
[installation problems](#common-installation-problems).
@@ -292,24 +292,23 @@
to 6006.
* <i>TensorFlowImage</i> is required. It identifies the Docker container.
You must specify one of the following values:
- * <code>gcr.io/tensorflow/tensorflow</code>: TensorFlow binary image.
- * <code>gcr.io/tensorflow/tensorflow:latest-devel</code>: TensorFlow
+ * <code>tensorflow/tensorflow</code>: TensorFlow binary image.
+ * <code>tensorflow/tensorflow:latest-devel</code>: TensorFlow
Binary image plus source code.
-<code>gcr.io</code> is the Google Container Registry. Note that some
-TensorFlow images are also available at
+The TensorFlow images are available at
[dockerhub](https://hub.docker.com/r/tensorflow/tensorflow/).
For example, the following command launches a TensorFlow CPU binary image
in a Docker container from which you can run TensorFlow programs in a shell:
-<pre>$ <b>docker run -it gcr.io/tensorflow/tensorflow bash</b></pre>
+<pre>$ <b>docker run -it tensorflow/tensorflow bash</b></pre>
The following command also launches a TensorFlow CPU binary image in a
Docker container. However, in this Docker container, you can run
TensorFlow programs in a Jupyter notebook:
-<pre>$ <b>docker run -it -p 8888:8888 gcr.io/tensorflow/tensorflow</b></pre>
+<pre>$ <b>docker run -it -p 8888:8888 tensorflow/tensorflow</b></pre>
Docker will download the TensorFlow binary image the first time you launch it.
@@ -351,7 +350,7 @@
TensorFlow for Python 2.7:
<pre> (<i>targetDirectory</i>)$ <b>pip install --ignore-installed --upgrade \
- https://storage.googleapis.com/tensorflow/mac/cpu/tensorflow-1.6.0rc1-py2-none-any.whl</b></pre>
+ https://storage.googleapis.com/tensorflow/mac/cpu/tensorflow-1.6.0-py2-none-any.whl</b></pre>
<a name="ValidateYourInstallation"></a>
@@ -376,7 +375,7 @@
If you installed through Docker, start a Docker container that runs bash.
For example:
-<pre>$ <b>docker run -it gcr.io/tensorflow/tensorflow bash</b></pre>
+<pre>$ <b>docker run -it tensorflow/tensorflow bash</b></pre>
@@ -524,7 +523,7 @@
<pre>
-https://storage.googleapis.com/tensorflow/mac/cpu/tensorflow-1.6.0rc1-py2-none-any.whl
+https://storage.googleapis.com/tensorflow/mac/cpu/tensorflow-1.6.0-py2-none-any.whl
</pre>
@@ -532,5 +531,5 @@
<pre>
-https://storage.googleapis.com/tensorflow/mac/cpu/tensorflow-1.6.0rc1-py3-none-any.whl
+https://storage.googleapis.com/tensorflow/mac/cpu/tensorflow-1.6.0-py3-none-any.whl
</pre>
diff --git a/tensorflow/docs_src/install/install_sources.md b/tensorflow/docs_src/install/install_sources.md
index acf0af0..c09c9c2 100644
--- a/tensorflow/docs_src/install/install_sources.md
+++ b/tensorflow/docs_src/install/install_sources.md
@@ -359,10 +359,10 @@
The filename of the `.whl` file depends on your platform.
For example, the following command will install the pip package
-for TensorFlow 1.6.0rc1 on Linux:
+for TensorFlow 1.6.0 on Linux:
<pre>
-$ <b>sudo pip install /tmp/tensorflow_pkg/tensorflow-1.6.0rc1-py2-none-any.whl</b>
+$ <b>sudo pip install /tmp/tensorflow_pkg/tensorflow-1.6.0-py2-none-any.whl</b>
</pre>
## Validate your installation
diff --git a/tensorflow/docs_src/install/install_windows.md b/tensorflow/docs_src/install/install_windows.md
index f0a30ee..2413bc9 100644
--- a/tensorflow/docs_src/install/install_windows.md
+++ b/tensorflow/docs_src/install/install_windows.md
@@ -17,7 +17,7 @@
NVIDIA® GPU, you must install this version. Note that this version of
TensorFlow is typically much easier to install (typically,
in 5 or 10 minutes), so even if you have an NVIDIA GPU, we recommend
- installing this version first.
+ installing this version first. Prebuilt binaries will use AVX instructions.
* **TensorFlow with GPU support**. TensorFlow programs typically run
significantly faster on a GPU than on a CPU. Therefore, if your
system has a NVIDIA® GPU meeting the prerequisites shown below
@@ -41,7 +41,8 @@
Note that cuDNN is typically installed in a different location from the
other CUDA DLLs. Ensure that you add the directory where you installed
the cuDNN DLL to your `%PATH%` environment variable.
- * GPU card with CUDA Compute Capability 3.0 or higher. See
+ * GPU card with CUDA Compute Capability 3.0 or higher for building
+ from source and 3.5 or higher for our binaries. See
[NVIDIA documentation](https://developer.nvidia.com/cuda-gpus) for a
list of supported GPU cards.
diff --git a/tensorflow/docs_src/performance/xla/jit.md b/tensorflow/docs_src/performance/xla/jit.md
index d4dc3e5..d9a979c 100644
--- a/tensorflow/docs_src/performance/xla/jit.md
+++ b/tensorflow/docs_src/performance/xla/jit.md
@@ -157,7 +157,7 @@
in succession.
To Render the .dot file into a png, install
-[GraphViz](http://www.graphviz.org/Download..php) and run:
+[GraphViz](https://www.graphviz.org/download/) and run:
```shell
dot -Tpng hlo_graph_80.dot -o hlo_graph_80.png
diff --git a/tensorflow/docs_src/programmers_guide/debugger.md b/tensorflow/docs_src/programmers_guide/debugger.md
index 5fb1c2d..d139981 100644
--- a/tensorflow/docs_src/programmers_guide/debugger.md
+++ b/tensorflow/docs_src/programmers_guide/debugger.md
@@ -459,7 +459,7 @@
[debug_tflearn_iris.py](https://www.tensorflow.org/code/tensorflow/python/debug/examples/debug_tflearn_iris.py),
-based on {$tflearn$tf-learn's iris tutorial}, contains a full example of how to
+based on [tf-learn's iris tutorial](https://www.tensorflow.org/versions/r1.2/get_started/tflearn), contains a full example of how to
use the tfdbg with `Estimator`s. To run this example, do:
```none
@@ -753,6 +753,7 @@
# For LocalCLIDebugHook
hooks = [tf_debug.LocalCLIDebugHook(dump_root="/with/lots/of/space")]
```
+
Make sure that the directory pointed to by dump_root is empty or nonexistent.
tfdbg cleans up the dump directories before exiting.
* Reduce the batch size used during the runs.
diff --git a/tensorflow/docs_src/programmers_guide/faq.md b/tensorflow/docs_src/programmers_guide/faq.md
index 1548d43..392ac6f 100644
--- a/tensorflow/docs_src/programmers_guide/faq.md
+++ b/tensorflow/docs_src/programmers_guide/faq.md
@@ -159,7 +159,7 @@
@{$reading_data$input pipelines}, at the cost of making the
TensorFlow computation somewhat more complicated. See the how-to documentation
for
-@{$reading_data#creating-threads-to-prefetch-using-queuerunner-objects$using `QueueRunner` objects to drive queues and readers}
+@{$reading_data#creating_threads_to_prefetch_using_queuerunner_objects$using `QueueRunner` objects to drive queues and readers}
for more information on how to use them.
## Variables
@@ -272,7 +272,7 @@
If your data is not easily parsable with the built-in TensorFlow operations,
consider converting it, offline, to a format that is easily parsable, such
-as ${tf.python_io.TFRecordWriter$`TFRecord`} format.
+as @{tf.python_io.TFRecordWriter$`TFRecord`} format.
The more efficient method to customize the parsing behavior is to
@{$adding_an_op$add a new op written in C++} that parses your
diff --git a/tensorflow/docs_src/programmers_guide/summaries_and_tensorboard.md b/tensorflow/docs_src/programmers_guide/summaries_and_tensorboard.md
index 79280d2..fadfa03 100644
--- a/tensorflow/docs_src/programmers_guide/summaries_and_tensorboard.md
+++ b/tensorflow/docs_src/programmers_guide/summaries_and_tensorboard.md
@@ -83,7 +83,7 @@
every `n` steps.
The code example below is a modification of the
-@{$layers$simple MNIST tutorial},
+[simple MNIST tutorial](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/examples/tutorials/mnist/mnist.py),
in which we have added some summary ops, and run them every ten steps. If you
run this and then launch `tensorboard --logdir=/tmp/tensorflow/mnist`, you'll be able
to visualize statistics, such as how the weights or accuracy varied during
diff --git a/tensorflow/docs_src/programmers_guide/using_tpu.md b/tensorflow/docs_src/programmers_guide/using_tpu.md
index d74d7f3..a9c2cb3 100644
--- a/tensorflow/docs_src/programmers_guide/using_tpu.md
+++ b/tensorflow/docs_src/programmers_guide/using_tpu.md
@@ -129,10 +129,9 @@
Typically the `FLAGS` would be set by command line arguments. To switch from
training locally to training on a cloud TPU you would need to:
- 1) Set `FLAGS.use_tpu` to `True`
- 1) Set `FLAGS.tpu_name` so the
- `tf.contrib.cluster_resolver.TPUClusterResolver` can find it
- 1) Set `FLAGS.model_dir` to a Google Cloud Storage bucket url (`gs://`).
+* Set `FLAGS.use_tpu` to `True`
+* Set `FLAGS.tpu_name` so the `tf.contrib.cluster_resolver.TPUClusterResolver` can find it
+* Set `FLAGS.model_dir` to a Google Cloud Storage bucket url (`gs://`).
## Optimizer
diff --git a/tensorflow/docs_src/tutorials/layers.md b/tensorflow/docs_src/tutorials/layers.md
index ee03f44..9b17d0d 100644
--- a/tensorflow/docs_src/tutorials/layers.md
+++ b/tensorflow/docs_src/tutorials/layers.md
@@ -193,7 +193,7 @@
you're already experienced with CNNs and @{$get_started/custom_estimators$TensorFlow `Estimator`s},
and find the above code intuitive, you may want to skim these sections or just
skip ahead to ["Training and Evaluating the CNN MNIST
-Classifier"](#training-and-evaluating-the-cnn-mnist-classifier).
+Classifier"](#training_and_evaluating_the_cnn_mnist_classifier).
### Input Layer
@@ -446,7 +446,7 @@
> Note: We use the `name` argument to explicitly name this operation
> `softmax_tensor`, so we can reference it later. (We'll set up logging for the
-> softmax values in ["Set Up a Logging Hook"](#set-up-a-logging-hook).
+> softmax values in ["Set Up a Logging Hook"](#set-up-a-logging-hook)).
We compile our predictions in a dict, and return an `EstimatorSpec` object:
@@ -534,9 +534,8 @@
```
> Note: For a more in-depth look at configuring training ops for Estimator model
-> functions, see @{$get_started/custom_estimators#defining-the-training-op-for-the-model$"Defining
-> the training op for the model"} in the @{$get_started/custom_estimators$"Creating Estimations in
-> tf.estimator"} tutorial.
+> functions, see @{$get_started/custom_estimators#defining_the_training_op_for_the_model$"Defining the training op for the model"}
+> in the @{$get_started/custom_estimators$"Creating Estimators in tf.estimator."} tutorial.
### Add evaluation metrics
@@ -625,7 +624,8 @@
> Note: If you don't explicitly assign a name to an operation via the `name`
> argument, TensorFlow will assign a default name. A couple easy ways to
> discover the names applied to operations are to visualize your graph on
-> @{$graph_viz$TensorBoard}) or to enable the @{$debugger$TensorFlow Debugger (tfdbg)}.
+> @{$graph_viz$TensorBoard}) or to enable the
+> @{$programmers_guide/debugger$TensorFlow Debugger (tfdbg)}.
Next, we create the `LoggingTensorHook`, passing `tensors_to_log` to the
`tensors` argument. We set `every_n_iter=50`, which specifies that probabilities
diff --git a/tensorflow/docs_src/tutorials/recurrent_quickdraw.md b/tensorflow/docs_src/tutorials/recurrent_quickdraw.md
index e22536a..7584a76 100644
--- a/tensorflow/docs_src/tutorials/recurrent_quickdraw.md
+++ b/tensorflow/docs_src/tutorials/recurrent_quickdraw.md
@@ -109,7 +109,8 @@
To convert the `ndjson` files to
@{$python/python_io#tfrecords_format_details$TFRecord} files containing
-${tf.train.Example} protos run the following command.
+[`tf.train.Example`](https://www.tensorflow.org/code/tensorflow/core/example/example.proto)
+protos run the following command.
```shell
python create_dataset.py --ndjson_path rnn_tutorial_data \
diff --git a/tensorflow/docs_src/tutorials/wide.md b/tensorflow/docs_src/tutorials/wide.md
index 005dc02..27ce75a 100644
--- a/tensorflow/docs_src/tutorials/wide.md
+++ b/tensorflow/docs_src/tutorials/wide.md
@@ -74,8 +74,8 @@
| relationship | Categorical | Wife, Own-child, Husband, |
: : : Not-in-family, Other-relative, :
: : : Unmarried. :
-| race | Categorical | White, Asian-Pac-Islander, |
-: : : Amer-Indian-Eskimo, Other, Black. :
+| race | Categorical | Amer-Indian-Eskimo, Asian-Pac- |
+: : : Islander, Black, White, Other. :
| gender | Categorical | Female, Male. |
| capital_gain | Continuous | Capital gains recorded. |
| capital_loss | Continuous | Capital Losses recorded. |
@@ -247,7 +247,7 @@
### Making Continuous Features Categorical through Bucketization
Sometimes the relationship between a continuous feature and the label is not
-linear. As an hypothetical example, a person's income may grow with age in the
+linear. As a hypothetical example, a person's income may grow with age in the
early stage of one's career, then the growth may slow at some point, and finally
the income decreases after retirement. In this scenario, using the raw `age` as
a real-valued feature column might not be a good choice because the model can
@@ -361,6 +361,16 @@
`accuracy: 0.83557522`, which means the accuracy is 83.6%. Feel free to try more
features and transformations and see if you can do even better!
+After the model is evaluated, we can use the model to predict whether an individual has an annual income of over
+50,000 dollars given an individual's information input.
+```python
+ pred_iter = model.predict(input_fn=lambda: input_fn(FLAGS.test_data, 1, False, 1))
+ for pred in pred_iter:
+ print(pred['classes'])
+```
+
+The model prediction output would be like `[b'1']` or `[b'0']` which means whether corresponding individual has an annual income of over 50,000 dollars or not.
+
If you'd like to see a working end-to-end example, you can download our
[example code](https://github.com/tensorflow/models/tree/master/official/wide_deep/wide_deep.py)
and set the `model_type` flag to `wide`.
diff --git a/tensorflow/examples/android/AndroidManifest.xml b/tensorflow/examples/android/AndroidManifest.xml
index bb75431..5c47ce6 100644
--- a/tensorflow/examples/android/AndroidManifest.xml
+++ b/tensorflow/examples/android/AndroidManifest.xml
@@ -40,6 +40,7 @@
<intent-filter>
<action android:name="android.intent.action.MAIN" />
<category android:name="android.intent.category.LAUNCHER" />
+ <category android:name="android.intent.category.LEANBACK_LAUNCHER" />
</intent-filter>
</activity>
@@ -49,6 +50,7 @@
<intent-filter>
<action android:name="android.intent.action.MAIN" />
<category android:name="android.intent.category.LAUNCHER" />
+ <category android:name="android.intent.category.LEANBACK_LAUNCHER" />
</intent-filter>
</activity>
@@ -58,6 +60,7 @@
<intent-filter>
<action android:name="android.intent.action.MAIN" />
<category android:name="android.intent.category.LAUNCHER" />
+ <category android:name="android.intent.category.LEANBACK_LAUNCHER" />
</intent-filter>
</activity>
@@ -67,6 +70,7 @@
<intent-filter>
<action android:name="android.intent.action.MAIN" />
<category android:name="android.intent.category.LAUNCHER" />
+ <category android:name="android.intent.category.LEANBACK_LAUNCHER" />
</intent-filter>
</activity>
</application>
diff --git a/tensorflow/examples/android/src/org/tensorflow/demo/CameraActivity.java b/tensorflow/examples/android/src/org/tensorflow/demo/CameraActivity.java
index 8bd4abb..429138a 100644
--- a/tensorflow/examples/android/src/org/tensorflow/demo/CameraActivity.java
+++ b/tensorflow/examples/android/src/org/tensorflow/demo/CameraActivity.java
@@ -351,6 +351,10 @@
protected void setFragment() {
String cameraId = chooseCamera();
+ if (cameraId == null) {
+ Toast.makeText(this, "No Camera Detected", Toast.LENGTH_SHORT).show();
+ finish();
+ }
Fragment fragment;
if (useCamera2API) {
@@ -416,7 +420,8 @@
@Override
public boolean onKeyDown(final int keyCode, final KeyEvent event) {
- if (keyCode == KeyEvent.KEYCODE_VOLUME_DOWN || keyCode == KeyEvent.KEYCODE_VOLUME_UP) {
+ if (keyCode == KeyEvent.KEYCODE_VOLUME_DOWN || keyCode == KeyEvent.KEYCODE_VOLUME_UP
+ || keyCode == KeyEvent.KEYCODE_BUTTON_L1 || keyCode == KeyEvent.KEYCODE_DPAD_CENTER) {
debug = !debug;
requestRender();
onSetDebug(debug);
diff --git a/tensorflow/examples/android/src/org/tensorflow/demo/StylizeActivity.java b/tensorflow/examples/android/src/org/tensorflow/demo/StylizeActivity.java
index 6a66ec3..33ec65e 100644
--- a/tensorflow/examples/android/src/org/tensorflow/demo/StylizeActivity.java
+++ b/tensorflow/examples/android/src/org/tensorflow/demo/StylizeActivity.java
@@ -16,8 +16,10 @@
package org.tensorflow.demo;
+import android.app.UiModeManager;
import android.content.Context;
import android.content.res.AssetManager;
+import android.content.res.Configuration;
import android.graphics.Bitmap;
import android.graphics.Bitmap.Config;
import android.graphics.BitmapFactory;
@@ -31,9 +33,11 @@
import android.media.ImageReader.OnImageAvailableListener;
import android.os.Bundle;
import android.os.SystemClock;
+import android.util.DisplayMetrics;
import android.util.Size;
import android.util.TypedValue;
import android.view.Display;
+import android.view.KeyEvent;
import android.view.MotionEvent;
import android.view.View;
import android.view.View.OnClickListener;
@@ -43,6 +47,7 @@
import android.widget.Button;
import android.widget.GridView;
import android.widget.ImageView;
+import android.widget.RelativeLayout;
import android.widget.Toast;
import java.io.IOException;
import java.io.InputStream;
@@ -381,6 +386,27 @@
grid = (GridView) findViewById(R.id.grid_layout);
grid.setAdapter(adapter);
grid.setOnTouchListener(gridTouchAdapter);
+
+ // Change UI on Android TV
+ UiModeManager uiModeManager = (UiModeManager) getSystemService(UI_MODE_SERVICE);
+ if (uiModeManager.getCurrentModeType() == Configuration.UI_MODE_TYPE_TELEVISION) {
+ DisplayMetrics displayMetrics = new DisplayMetrics();
+ getWindowManager().getDefaultDisplay().getMetrics(displayMetrics);
+ int styleSelectorHeight = displayMetrics.heightPixels;
+ int styleSelectorWidth = displayMetrics.widthPixels - styleSelectorHeight;
+ RelativeLayout.LayoutParams layoutParams = new RelativeLayout.LayoutParams(styleSelectorWidth, ViewGroup.LayoutParams.MATCH_PARENT);
+
+ // Calculate number of style in a row, so all the style can show up without scrolling
+ int numOfStylePerRow = 3;
+ while (styleSelectorWidth / numOfStylePerRow * Math.ceil((float) (adapter.getCount() - 2) / numOfStylePerRow) > styleSelectorHeight) {
+ numOfStylePerRow++;
+ }
+ grid.setNumColumns(numOfStylePerRow);
+ layoutParams.addRule(RelativeLayout.ALIGN_PARENT_RIGHT);
+ grid.setLayoutParams(layoutParams);
+ adapter.buttons.clear();
+ }
+
setStyle(adapter.items[0], 1.0f);
}
@@ -602,4 +628,38 @@
borderedText.drawLines(canvas, 10, canvas.getHeight() - 10, lines);
}
+
+ @Override
+ public boolean onKeyDown(int keyCode, KeyEvent event) {
+ int moveOffset = 0;
+ switch (keyCode) {
+ case KeyEvent.KEYCODE_DPAD_LEFT:
+ moveOffset = -1;
+ break;
+ case KeyEvent.KEYCODE_DPAD_RIGHT:
+ moveOffset = 1;
+ break;
+ case KeyEvent.KEYCODE_DPAD_UP:
+ moveOffset = -1 * grid.getNumColumns();
+ break;
+ case KeyEvent.KEYCODE_DPAD_DOWN:
+ moveOffset = grid.getNumColumns();
+ break;
+ default:
+ return super.onKeyDown(keyCode, event);
+ }
+
+ // get the highest selected style
+ int currentSelect = 0;
+ float highestValue = 0;
+ for (int i = 0; i < adapter.getCount(); i++) {
+ if (adapter.items[i].value > highestValue) {
+ currentSelect = i;
+ highestValue = adapter.items[i].value;
+ }
+ }
+ setStyle(adapter.items[(currentSelect + moveOffset + adapter.getCount()) % adapter.getCount()], 1);
+
+ return true;
+ }
}
diff --git a/tensorflow/python/BUILD b/tensorflow/python/BUILD
index a206685..3cbeb34 100644
--- a/tensorflow/python/BUILD
+++ b/tensorflow/python/BUILD
@@ -28,6 +28,7 @@
load("//tensorflow:tensorflow.bzl", "tf_py_build_info_genrule")
load("//tensorflow:tensorflow.bzl", "tf_py_wrap_cc")
load("//tensorflow:tensorflow.bzl", "tf_cc_shared_object")
+load("//tensorflow:tensorflow.bzl", "tf_custom_op_library_additional_deps_impl")
load("//tensorflow:tensorflow.bzl", "cuda_py_test")
load("//tensorflow:tensorflow.bzl", "cuda_py_tests")
load("//tensorflow/core:platform/default/build_config.bzl", "pyx_library")
@@ -106,20 +107,19 @@
":training",
":util",
":weights_broadcast_ops",
- "//third_party/py/numpy",
+ "//tensorflow/contrib:contrib_py",
"//tensorflow/core:protos_all_py",
"//tensorflow/python/data",
"//tensorflow/python/estimator:estimator_py",
"//tensorflow/python/feature_column:feature_column_py",
"//tensorflow/python/keras",
- "//tensorflow/python/ops/losses",
"//tensorflow/python/ops/distributions",
"//tensorflow/python/ops/linalg",
+ "//tensorflow/python/ops/losses",
"//tensorflow/python/profiler",
"//tensorflow/python/saved_model",
- ] + if_not_windows([
- "//tensorflow/contrib:contrib_py",
- ]),
+ "//third_party/py/numpy",
+ ],
)
tf_py_build_info_genrule()
@@ -947,7 +947,6 @@
srcs = ["framework/contrib_test.py"],
main = "framework/contrib_test.py",
srcs_version = "PY2AND3",
- tags = ["no_windows"],
deps = [
"//tensorflow:tensorflow_py",
"//tensorflow/python:client_testlib",
@@ -1312,7 +1311,6 @@
srcs = ["framework/dtypes_test.py"],
main = "framework/dtypes_test.py",
srcs_version = "PY2AND3",
- tags = ["no_windows"],
deps = [
":framework_for_generated_wrappers",
":framework_test_lib",
@@ -1654,7 +1652,6 @@
size = "small",
srcs = ["ops/clip_ops_test.py"],
srcs_version = "PY2AND3",
- tags = ["no_windows"],
deps = [
":client_testlib",
":clip_ops",
@@ -2721,7 +2718,6 @@
],
data = ["//tensorflow/core:image_testdata"],
shard_count = 5,
- tags = ["no_windows"],
)
cuda_py_test(
@@ -3305,6 +3301,65 @@
tf_additional_gdr_deps()),
)
+# ** Targets for Windows build (start) **
+# We need the following targets to expose symbols from _pywrap_tensorflow.dll
+
+# Build a cc_binary from tf_custom_op_library_additional_deps_impl,
+# it contains all object code from its dependencies.
+cc_binary(
+ name = "tf_custom_op_library_additional_deps.so",
+ linkshared = 1,
+ linkstatic = 1,
+ deps = tf_custom_op_library_additional_deps_impl(),
+)
+
+# Get a DEF file generated by parsing all object files
+# of tf_custom_op_library_additional_deps.so
+filegroup(
+ name = "pywrap_tensorflow_def_file",
+ srcs = [":tf_custom_op_library_additional_deps.so"],
+ output_group = "def_file",
+)
+
+# Filter the DEF file to reduce the number of symbols to 64K or less.
+# Note that we also write the name of the pyd file into DEF file so that
+# the dynamic libraries of custom ops can find it at runtime.
+genrule(
+ name = "pywrap_tensorflow_filtered_def_file",
+ srcs = [":pywrap_tensorflow_def_file"],
+ outs = ["pywrap_tensorflow_filtered_def_file.def"],
+ cmd = select({
+ "//tensorflow:windows": """
+ $(location @local_config_def_file_filter//:def_file_filter) \\
+ --input $(location :pywrap_tensorflow_def_file) \\
+ --output $@ \\
+ --target _pywrap_tensorflow_internal.pyd
+ """,
+ "//conditions:default": "touch $@", # Just a placeholder for Unix platforms
+ }),
+ tools = ["@local_config_def_file_filter//:def_file_filter"],
+)
+
+# Get the import library of _pywrap_tensorflow_internal.dll
+filegroup(
+ name = "pywrap_tensorflow_import_lib_file",
+ srcs = [":_pywrap_tensorflow_internal.so"],
+ output_group = "interface_library",
+)
+
+# Create a cc_import rule for the import library of _pywrap_tensorflow_internal.dll
+# so that custom ops' dynamic libraries can link against it.
+cc_import(
+ name = "pywrap_tensorflow_import_lib",
+ interface_library = select({
+ "//tensorflow:windows": ":pywrap_tensorflow_import_lib_file",
+ "//conditions:default": "not_exsiting_on_unix.lib", # Just a placeholder for Unix platforms
+ }),
+ system_provided = 1,
+)
+
+# ** Targets for Windows build (end) **
+
py_library(
name = "lib",
srcs = [
@@ -3677,7 +3732,6 @@
size = "small",
srcs = ["lib/core/bfloat16_test.py"],
srcs_version = "PY2AND3",
- tags = ["no_windows"],
deps = [
":client_testlib",
":lib",
@@ -3958,7 +4012,11 @@
srcs = ["training/checkpoint_utils_test.py"],
srcs_version = "PY2AND3",
tags = [
+ "manual",
+ "no_cuda_on_cpu_tap",
+ "no_oss",
"no_windows",
+ "notap",
],
deps = [
":client",
@@ -3981,7 +4039,6 @@
size = "small",
srcs = ["training/checkpoint_ops_test.py"],
srcs_version = "PY2AND3",
- tags = ["no_windows"],
deps = [
":checkpoint_ops_gen",
":client",
@@ -4022,10 +4079,7 @@
size = "medium",
srcs = ["training/monitored_session_test.py"],
srcs_version = "PY2AND3",
- tags = [
- "no_windows",
- "notsan", # b/67945581
- ],
+ tags = ["notsan"], # b/67945581
deps = [
":array_ops",
":client_testlib",
diff --git a/tensorflow/python/debug/BUILD b/tensorflow/python/debug/BUILD
index 512d292..c60f692 100644
--- a/tensorflow/python/debug/BUILD
+++ b/tensorflow/python/debug/BUILD
@@ -913,6 +913,7 @@
"//tensorflow/python:util",
"//tensorflow/python:variables",
],
+ tags = ["no_windows"], # TODO: needs investigation on Windows
)
py_test(
diff --git a/tensorflow/python/estimator/estimator.py b/tensorflow/python/estimator/estimator.py
index 9fcbd4f..5245a05 100644
--- a/tensorflow/python/estimator/estimator.py
+++ b/tensorflow/python/estimator/estimator.py
@@ -139,8 +139,8 @@
to configure Estimators from hyper parameter tuning.
* `config`: Optional configuration object. Will receive what is passed
to Estimator in `config` parameter, or the default `config`.
- Allows updating things in your model_fn based on configuration
- such as `num_ps_replicas`, or `model_dir`.
+ Allows updating things in your `model_fn` based on
+ configuration such as `num_ps_replicas`, or `model_dir`.
* Returns:
`EstimatorSpec`
@@ -301,11 +301,11 @@
* A 'tf.data.Dataset' object: Outputs of `Dataset` object must be a
tuple (features, labels) with same constraints as below.
- * A tuple (features, labels): Where features is a `Tensor` or a
- dictionary of string feature name to `Tensor` and labels is a
+ * A tuple (features, labels): Where `features` is a `Tensor` or a
+ dictionary of string feature name to `Tensor` and `labels` is a
`Tensor` or a dictionary of string label name to `Tensor`. Both
- features and labels are consumed by `model_fn`. They should satisfy
- the expectation of `model_fn` from inputs.
+ `features` and `labels` are consumed by `model_fn`. They should
+ satisfy the expectation of `model_fn` from inputs.
hooks: List of `SessionRunHook` subclass instances. Used for callbacks
inside the training loop.
@@ -381,11 +381,11 @@
* A 'tf.data.Dataset' object: Outputs of `Dataset` object must be a
tuple (features, labels) with same constraints as below.
- * A tuple (features, labels): Where features is a `Tensor` or a
- dictionary of string feature name to `Tensor` and labels is a
+ * A tuple (features, labels): Where `features` is a `Tensor` or a
+ dictionary of string feature name to `Tensor` and `labels` is a
`Tensor` or a dictionary of string label name to `Tensor`. Both
- features and labels are consumed by `model_fn`. They should satisfy
- the expectation of `model_fn` from inputs.
+ `features` and `labels` are consumed by `model_fn`. They should
+ satisfy the expectation of `model_fn` from inputs.
steps: Number of steps for which to evaluate model. If `None`, evaluates
until `input_fn` raises an end-of-input exception.
@@ -457,17 +457,17 @@
checkpoint_path: Path of a specific checkpoint to predict. If `None`, the
latest checkpoint in `model_dir` is used.
yield_single_examples: If False, yield the whole batch as returned by the
- model_fn instead of decomposing the batch into individual elements. This
- is useful if model_fn return some tensor with first dimension not
- equal to the batch size
+ `model_fn` instead of decomposing the batch into individual elements.
+ This is useful if `model_fn` returns some tensors whose first dimension
+ is not equal to the batch size.
Yields:
Evaluated values of `predictions` tensors.
Raises:
- ValueError: Could not find a trained model in model_dir.
- ValueError: if batch length of predictions are not same and
- yield_single_examples is True.
+ ValueError: Could not find a trained model in `model_dir`.
+ ValueError: If batch length of predictions is not the same and
+ `yield_single_examples` is True.
ValueError: If there is a conflict between `predict_keys` and
`predictions`. For example if `predict_keys` is not `None` but
`EstimatorSpec.predictions` is not a `dict`.
diff --git a/tensorflow/python/estimator/training.py b/tensorflow/python/estimator/training.py
index 2cc3331..e38b765 100644
--- a/tensorflow/python/estimator/training.py
+++ b/tensorflow/python/estimator/training.py
@@ -128,9 +128,16 @@
"""Creates a validated `TrainSpec` instance.
Args:
- input_fn: Training input function returning a tuple of:
- features - `Tensor` or dictionary of string feature name to `Tensor`.
- labels - `Tensor` or dictionary of `Tensor` with labels.
+ input_fn: A function that provides input data for training as minibatches.
+ See @{$get_started/premade_estimators#create_input_functions} for more
+ information. The function should construct and return one of
+ the following:
+ * A 'tf.data.Dataset' object: Outputs of `Dataset` object must be a
+ tuple (features, labels) with same constraints as below.
+ * A tuple (features, labels): Where features is a `Tensor` or a
+ dictionary of string feature name to `Tensor` and labels is a
+ `Tensor` or a dictionary of string label name to `Tensor`.
+
max_steps: Int. Positive number of total steps for which to train model.
If `None`, train forever. The training `input_fn` is not expected to
generate `OutOfRangeError` or `StopIteration` exceptions. See the
@@ -185,9 +192,16 @@
"""Creates a validated `EvalSpec` instance.
Args:
- input_fn: Evaluation input function returning a tuple of:
- features - `Tensor` or dictionary of string feature name to `Tensor`.
- labels - `Tensor` or dictionary of `Tensor` with labels.
+ input_fn: A function that constructs the input data for evaluation.
+ See @{$get_started/premade_estimators#create_input_functions} for more
+ information. The function should construct and return one of
+ the following:
+ * A 'tf.data.Dataset' object: Outputs of `Dataset` object must be a
+ tuple (features, labels) with same constraints as below.
+ * A tuple (features, labels): Where features is a `Tensor` or a
+ dictionary of string feature name to `Tensor` and labels is a
+ `Tensor` or a dictionary of string label name to `Tensor`.
+
steps: Int. Positive number of steps for which to evaluate model. If
`None`, evaluates until `input_fn` raises an end-of-input exception.
See `Estimator.evaluate` for details.
diff --git a/tensorflow/python/keras/BUILD b/tensorflow/python/keras/BUILD
index eef91e9..f74881f 100755
--- a/tensorflow/python/keras/BUILD
+++ b/tensorflow/python/keras/BUILD
@@ -636,7 +636,10 @@
size = "small",
srcs = ["_impl/keras/utils/io_utils_test.py"],
srcs_version = "PY2AND3",
- tags = ["notsan"],
+ tags = [
+ "no_windows", # TODO: needs investigation on Windows
+ "notsan",
+ ],
deps = [
":keras",
"//tensorflow/python:client_testlib",
diff --git a/tensorflow/python/kernel_tests/BUILD b/tensorflow/python/kernel_tests/BUILD
index 5b0c38f..f27ca5c 100644
--- a/tensorflow/python/kernel_tests/BUILD
+++ b/tensorflow/python/kernel_tests/BUILD
@@ -295,7 +295,6 @@
"//tensorflow/python:nn_grad",
],
data = ["//tensorflow/core:image_testdata"],
- tags = ["no_windows"],
)
tf_py_test(
@@ -1139,7 +1138,6 @@
"//tensorflow/python:variables",
],
data = ["//tensorflow/core:lmdb_testdata"],
- tags = ["no_windows"],
)
cuda_py_test(
@@ -2329,7 +2327,6 @@
"//tensorflow/python:variables",
],
shard_count = 4,
- tags = ["no_windows"],
)
cuda_py_test(
@@ -2460,7 +2457,6 @@
"//tensorflow/python/eager:context",
],
shard_count = 10,
- tags = ["no_windows"],
)
cuda_py_test(
diff --git a/tensorflow/python/kernel_tests/conv_ops_test.py b/tensorflow/python/kernel_tests/conv_ops_test.py
index f4fe01f..25525cc 100644
--- a/tensorflow/python/kernel_tests/conv_ops_test.py
+++ b/tensorflow/python/kernel_tests/conv_ops_test.py
@@ -970,7 +970,7 @@
self.assertArrayNear(value_2.flatten(), value.flatten(), err)
def testConv2D2x2Depth3ValidBackpropFilterStride1x1Dilation2x1(self):
- if test.is_gpu_available(cuda_only=True):
+ if test.is_gpu_available(cuda_only=True) or test_util.IsMklEnabled():
for (data_format, use_gpu) in GetTestConfigs():
self._RunAndVerifyBackpropFilterDilation(
input_sizes=[1, 3, 6, 1],
@@ -984,7 +984,7 @@
err=1e-5)
def testConv2D2x2Depth1ValidBackpropFilterDilation1x2(self):
- if test.is_gpu_available(cuda_only=True):
+ if test.is_gpu_available(cuda_only=True) or test_util.IsMklEnabled():
for (data_format, use_gpu) in GetTestConfigs():
self._RunAndVerifyBackpropFilterDilation(
input_sizes=[1, 2, 3, 1],
@@ -998,7 +998,7 @@
err=1e-5)
def testConv2DEmptyBackpropFilterDilation1x2(self):
- if test.is_gpu_available(cuda_only=True):
+ if test.is_gpu_available(cuda_only=True) or test_util.IsMklEnabled():
for (data_format, use_gpu) in GetTestConfigs():
self._RunAndVerifyBackpropFilterDilation(
input_sizes=[1, 2, 3, 1],
@@ -1012,7 +1012,7 @@
err=1e-5)
def testConv2D2x2Depth3ValidBackpropFilterDilation2x2(self):
- if test.is_gpu_available(cuda_only=True):
+ if test.is_gpu_available(cuda_only=True) or test_util.IsMklEnabled():
for (data_format, use_gpu) in GetTestConfigs():
self._RunAndVerifyBackpropFilterDilation(
input_sizes=[1, 3, 4, 3],
@@ -1026,7 +1026,7 @@
err=1e-5)
def testConv2DKernelSizeMatchesInputSizeBackpropFilterDilation2x2(self):
- if test.is_gpu_available(cuda_only=True):
+ if test.is_gpu_available(cuda_only=True) or test_util.IsMklEnabled():
for (data_format, use_gpu) in GetTestConfigs():
self._RunAndVerifyBackpropFilterDilation(
input_sizes=[1, 3, 3, 1],
@@ -1040,7 +1040,7 @@
err=1e-5)
def testConv2D2x2Depth3ValidBackpropInputStride1x1Dilation2x1(self):
- if test.is_gpu_available(cuda_only=True):
+ if test.is_gpu_available(cuda_only=True) or test_util.IsMklEnabled():
for (data_format, use_gpu) in GetTestConfigs():
self._RunAndVerifyBackpropInputDilation(
input_sizes=[1, 3, 6, 1],
@@ -1054,7 +1054,7 @@
err=1e-5)
def testConv2D2x2Depth1ValidBackpropInputDilation1x2(self):
- if test.is_gpu_available(cuda_only=True):
+ if test.is_gpu_available(cuda_only=True) or test_util.IsMklEnabled():
for (data_format, use_gpu) in GetTestConfigs():
self._RunAndVerifyBackpropInputDilation(
input_sizes=[1, 2, 3, 1],
@@ -1068,7 +1068,7 @@
err=1e-5)
def testConv2DEmptyBackpropInputDilation1x2(self):
- if test.is_gpu_available(cuda_only=True):
+ if test.is_gpu_available(cuda_only=True) or test_util.IsMklEnabled():
for (data_format, use_gpu) in GetTestConfigs():
self._RunAndVerifyBackpropInputDilation(
input_sizes=[0, 2, 3, 1],
@@ -1082,7 +1082,7 @@
err=1e-5)
def testConv2D2x2Depth3ValidBackpropInputDilation2x1(self):
- if test.is_gpu_available(cuda_only=True):
+ if test.is_gpu_available(cuda_only=True) or test_util.IsMklEnabled():
for (data_format, use_gpu) in GetTestConfigs():
# The GPU version of this test is not very stable. So adjusting the
# error threshold to 1e-4.
@@ -1098,7 +1098,7 @@
err=1e-4)
def testConv2DKernelSizeMatchesInputSizeBackpropInputDilation2x2(self):
- if test.is_gpu_available(cuda_only=True):
+ if test.is_gpu_available(cuda_only=True) or test_util.IsMklEnabled():
for (data_format, use_gpu) in GetTestConfigs():
self._RunAndVerifyBackpropInputDilation(
input_sizes=[1, 3, 3, 1],
diff --git a/tensorflow/python/kernel_tests/depthtospace_op_test.py b/tensorflow/python/kernel_tests/depthtospace_op_test.py
index 96c9718..f0beabb 100644
--- a/tensorflow/python/kernel_tests/depthtospace_op_test.py
+++ b/tensorflow/python/kernel_tests/depthtospace_op_test.py
@@ -35,8 +35,8 @@
class DepthToSpaceTest(test.TestCase):
- def _testOne(self, inputs, block_size, outputs):
- input_nhwc = math_ops.to_float(inputs)
+ def _testOne(self, inputs, block_size, outputs, dtype=dtypes.float32):
+ input_nhwc = math_ops.cast(inputs, dtype)
with self.test_session(use_gpu=False):
# test NHWC (default) on CPU
x_tf = array_ops.depth_to_space(input_nhwc, block_size)
@@ -59,6 +59,12 @@
x_out = [[[[1], [2]], [[3], [4]]]]
self._testOne(x_np, block_size, x_out)
+ def testBasicFloat16(self):
+ x_np = [[[[1, 2, 3, 4]]]]
+ block_size = 2
+ x_out = [[[[1], [2]], [[3], [4]]]]
+ self._testOne(x_np, block_size, x_out, dtype=dtypes.float16)
+
# Tests for larger input dimensions. To make sure elements are
# correctly ordered spatially.
def testBlockSize2(self):
diff --git a/tensorflow/python/kernel_tests/spacetodepth_op_test.py b/tensorflow/python/kernel_tests/spacetodepth_op_test.py
index b761357..cd90d16 100644
--- a/tensorflow/python/kernel_tests/spacetodepth_op_test.py
+++ b/tensorflow/python/kernel_tests/spacetodepth_op_test.py
@@ -34,8 +34,8 @@
class SpaceToDepthTest(test.TestCase):
- def _testOne(self, inputs, block_size, outputs):
- input_nhwc = math_ops.to_float(inputs)
+ def _testOne(self, inputs, block_size, outputs, dtype=dtypes.float32):
+ input_nhwc = math_ops.cast(inputs, dtype)
with self.test_session(use_gpu=False):
# test NHWC (default) on CPU
x_tf = array_ops.space_to_depth(input_nhwc, block_size)
@@ -58,6 +58,12 @@
x_out = [[[[1, 2, 3, 4]]]]
self._testOne(x_np, block_size, x_out)
+ def testBasicFloat16(self):
+ x_np = [[[[1], [2]], [[3], [4]]]]
+ block_size = 2
+ x_out = [[[[1, 2, 3, 4]]]]
+ self._testOne(x_np, block_size, x_out, dtype=dtypes.float16)
+
# Tests for larger input dimensions. To make sure elements are
# correctly ordered spatially.
def testLargerInput2x2(self):
diff --git a/tensorflow/python/lib/io/file_io_test.py b/tensorflow/python/lib/io/file_io_test.py
index a751607..223858e 100644
--- a/tensorflow/python/lib/io/file_io_test.py
+++ b/tensorflow/python/lib/io/file_io_test.py
@@ -485,6 +485,11 @@
f.flush()
self.assertEqual(content, f.read(len(content) + 1))
+ def testUTF8StringPathExists(self):
+ file_path = os.path.join(self._base_dir, "UTF8测试_file_exist")
+ file_io.write_string_to_file(file_path, "testing")
+ v = file_io.file_exists(file_path)
+ self.assertEqual(v, True)
if __name__ == "__main__":
test.main()
diff --git a/tensorflow/python/ops/nn_ops.py b/tensorflow/python/ops/nn_ops.py
index fb3fe77..87fe253 100644
--- a/tensorflow/python/ops/nn_ops.py
+++ b/tensorflow/python/ops/nn_ops.py
@@ -699,7 +699,7 @@
`padded_input` is obtained by zero padding the input using an effective
spatial filter shape of `(spatial_filter_shape-1) * dilation_rate + 1` and
output striding `strides` as described in the
- @{tf.nn.convolution$comment here}.
+ @{$python/nn#Convolution$comment here}.
In the case that `data_format` does start with `"NC"`, the `input` and output
(but not the `filter`) are simply transposed as follows:
diff --git a/tensorflow/python/ops/rnn.py b/tensorflow/python/ops/rnn.py
index c59eccc..42af7f8 100644
--- a/tensorflow/python/ops/rnn.py
+++ b/tensorflow/python/ops/rnn.py
@@ -867,7 +867,7 @@
```python
time = tf.constant(0, dtype=tf.int32)
- (finished, next_input, initial_state, _, loop_state) = loop_fn(
+ (finished, next_input, initial_state, emit_structure, loop_state) = loop_fn(
time=time, cell_output=None, cell_state=None, loop_state=None)
emit_ta = TensorArray(dynamic_size=True, dtype=initial_state.dtype)
state = initial_state
@@ -878,7 +878,7 @@
loop_state=loop_state)
# Emit zeros and copy forward state for minibatch entries that are finished.
state = tf.where(finished, state, next_state)
- emit = tf.where(finished, tf.zeros_like(emit), emit)
+ emit = tf.where(finished, tf.zeros_like(emit_structure), emit)
emit_ta = emit_ta.write(time, emit)
# If any new minibatch entries are marked as finished, mark these.
finished = tf.logical_or(finished, next_finished)
@@ -938,10 +938,15 @@
and `emit_output`: the output to store for this iteration.
Note that `emit_output` should be a `Tensor` or (possibly nested)
- tuple of tensors with shapes and structure matching `cell.output_size`
- and `cell_output` above. The parameter `cell_state` and output
- `next_cell_state` may be either a single or (possibly nested) tuple
- of tensors. The parameter `loop_state` and
+ tuple of tensors which is aggregated in the `emit_ta` inside the
+ `while_loop`. For the first call to `loop_fn`, the `emit_output`
+ corresponds to the `emit_structure` which is then used to determine the
+ size of the `zero_tensor` for the `emit_ta` (defaults to
+ `cell.output_size`). For the subsequent calls to the `loop_fn`, the
+ `emit_output` corresponds to the actual output tensor
+ that is to be aggregated in the `emit_ta`. The parameter `cell_state`
+ and output `next_cell_state` may be either a single or (possibly nested)
+ tuple of tensors. The parameter `loop_state` and
output `next_loop_state` may be either a single or (possibly nested) tuple
of `Tensor` and `TensorArray` objects. This last parameter
may be ignored by `loop_fn` and the return value may be `None`. If it
diff --git a/tensorflow/python/ops/special_math_ops.py b/tensorflow/python/ops/special_math_ops.py
index 6d7eaab..5e2146b 100644
--- a/tensorflow/python/ops/special_math_ops.py
+++ b/tensorflow/python/ops/special_math_ops.py
@@ -163,7 +163,7 @@
if '...' in equation:
raise ValueError('Subscripts with ellipses are not yet supported.')
- match = re.match('([a-z,]+)(->[a-z]*)?', equation)
+ match = re.match('^([a-zA-Z,]+)(->[a-zA-Z]*)?$', equation)
if not match:
raise ValueError('Indices have incorrect format: %s' % equation)
@@ -402,7 +402,7 @@
if '...' in equation:
raise ValueError('Subscripts with ellipses are not yet supported.')
- match = re.match('([a-z,]+)(->[a-z]*)?', equation)
+ match = re.match('^([a-zA-Z,]+)(->[a-zA-Z]*)?$', equation)
if not match:
raise ValueError('Indices have incorrect format: %s' % equation)
diff --git a/tensorflow/python/ops/special_math_ops_test.py b/tensorflow/python/ops/special_math_ops_test.py
index 2c212f4..d7c3a7e8 100644
--- a/tensorflow/python/ops/special_math_ops_test.py
+++ b/tensorflow/python/ops/special_math_ops_test.py
@@ -192,6 +192,9 @@
'abc,cba',
'dba,ead,cad->bce',
'aef,fbc,dca->bde',
+ 'iJ,Jk->ik',
+ 'iJ,Ki->JK',
+ 'iJk,Jklm->Jk'
]
long_cases = [
@@ -208,6 +211,8 @@
'ijk ijk',
'ij.jk->ik',
'ij...,jk...->ik...',
+ 'ij,k ->kji',
+ 'ij,k-> kji',
# axis in output that does not exist
'ij,jk->im',
diff --git a/tensorflow/python/tools/freeze_graph.py b/tensorflow/python/tools/freeze_graph.py
index a52f325..e9f1def 100644
--- a/tensorflow/python/tools/freeze_graph.py
+++ b/tensorflow/python/tools/freeze_graph.py
@@ -56,8 +56,6 @@
from tensorflow.python.tools import saved_model_utils
from tensorflow.python.training import saver as saver_lib
-FLAGS = None
-
def freeze_graph_with_def_protos(input_graph_def,
input_saver_def,
@@ -256,25 +254,24 @@
checkpoint_version=checkpoint_version)
-def main(unused_args):
- if FLAGS.checkpoint_version == 1:
+def main(unused_args, flags):
+ if flags.checkpoint_version == 1:
checkpoint_version = saver_pb2.SaverDef.V1
- elif FLAGS.checkpoint_version == 2:
+ elif flags.checkpoint_version == 2:
checkpoint_version = saver_pb2.SaverDef.V2
else:
print("Invalid checkpoint version (must be '1' or '2'): %d" %
- FLAGS.checkpoint_version)
+ flags.checkpoint_version)
return -1
- freeze_graph(FLAGS.input_graph, FLAGS.input_saver, FLAGS.input_binary,
- FLAGS.input_checkpoint, FLAGS.output_node_names,
- FLAGS.restore_op_name, FLAGS.filename_tensor_name,
- FLAGS.output_graph, FLAGS.clear_devices, FLAGS.initializer_nodes,
- FLAGS.variable_names_whitelist, FLAGS.variable_names_blacklist,
- FLAGS.input_meta_graph, FLAGS.input_saved_model_dir,
- FLAGS.saved_model_tags, checkpoint_version)
+ freeze_graph(flags.input_graph, flags.input_saver, flags.input_binary,
+ flags.input_checkpoint, flags.output_node_names,
+ flags.restore_op_name, flags.filename_tensor_name,
+ flags.output_graph, flags.clear_devices, flags.initializer_nodes,
+ flags.variable_names_whitelist, flags.variable_names_blacklist,
+ flags.input_meta_graph, flags.input_saved_model_dir,
+ flags.saved_model_tags, checkpoint_version)
-
-if __name__ == "__main__":
+def run_main():
parser = argparse.ArgumentParser()
parser.register("type", "bool", lambda v: v.lower() == "true")
parser.add_argument(
@@ -376,5 +373,10 @@
separated by \',\'. For tag-set contains multiple tags, all tags \
must be passed in.\
""")
- FLAGS, unparsed = parser.parse_known_args()
- app.run(main=main, argv=[sys.argv[0]] + unparsed)
+ flags, unparsed = parser.parse_known_args()
+
+ my_main = lambda unused_args: main(unused_args, flags)
+ app.run(main=my_main, argv=[sys.argv[0]] + unparsed)
+
+if __name__ == '__main__':
+ run_main()
diff --git a/tensorflow/python/tools/saved_model_cli.py b/tensorflow/python/tools/saved_model_cli.py
index b0e9e3e..b88be4a 100644
--- a/tensorflow/python/tools/saved_model_cli.py
+++ b/tensorflow/python/tools/saved_model_cli.py
@@ -38,11 +38,15 @@
from tensorflow.core.framework import types_pb2
from tensorflow.python.client import session
from tensorflow.python.debug.wrappers import local_cli_wrapper
+from tensorflow.python.framework import meta_graph as meta_graph_lib
from tensorflow.python.framework import ops as ops_lib
from tensorflow.python.platform import app # pylint: disable=unused-import
from tensorflow.python.saved_model import loader
from tensorflow.python.tools import saved_model_utils
+# Set of ops to blacklist.
+_OP_BLACKLIST = set(['WriteFile', 'ReadFile'])
+
def _show_tag_sets(saved_model_dir):
"""Prints the tag-sets stored in SavedModel directory.
@@ -242,6 +246,27 @@
return meta_graph.signature_def
+def scan_meta_graph_def(meta_graph_def):
+ """Scans meta_graph_def and reports if there are ops on blacklist.
+
+ Print ops if they are on black list, or print success if no blacklisted ops
+ found.
+
+ Args:
+ meta_graph_def: MetaGraphDef protocol buffer.
+ """
+ all_ops_set = set(
+ meta_graph_lib.ops_used_by_graph_def(meta_graph_def.graph_def))
+ blacklisted_ops = _OP_BLACKLIST & all_ops_set
+ if blacklisted_ops:
+ # TODO(yifeif): print more warnings
+ print('MetaGraph with tag set %s contains the following blacklisted ops:' %
+ meta_graph_def.meta_info_def.tags, blacklisted_ops)
+ else:
+ print('MetaGraph with tag set %s does not contain blacklisted ops.' %
+ meta_graph_def.meta_info_def.tags)
+
+
def run_saved_model_with_feed_dict(saved_model_dir, tag_set, signature_def_key,
input_tensor_key_feed_dict, outdir,
overwrite_flag, tf_debug=False):
@@ -609,6 +634,21 @@
args.overwrite, tf_debug=args.tf_debug)
+def scan(args):
+ """Function triggered by scan command.
+
+ Args:
+ args: A namespace parsed from command line.
+ """
+ if args.tag_set:
+ scan_meta_graph_def(
+ saved_model_utils.get_meta_graph_def(args.dir, args.tag_set))
+ else:
+ saved_model = reader.read_saved_model(args.dir)
+ for meta_graph_def in saved_model.meta_graphs:
+ scan_meta_graph_def(meta_graph_def)
+
+
def create_parser():
"""Creates a parser that parse the command line arguments.
@@ -730,6 +770,26 @@
'SavedModel.')
parser_run.set_defaults(func=run)
+ # scan command
+ scan_msg = ('Usage example:\n'
+ 'To scan for blacklisted ops in SavedModel:\n'
+ '$saved_model_cli scan --dir /tmp/saved_model\n'
+ 'To scan a specific MetaGraph, pass in --tag_set\n')
+ parser_scan = subparsers.add_parser(
+ 'scan',
+ description=scan_msg,
+ formatter_class=argparse.RawTextHelpFormatter)
+ parser_scan.add_argument(
+ '--dir',
+ type=str,
+ required=True,
+ help='directory containing the SavedModel to execute')
+ parser_scan.add_argument(
+ '--tag_set',
+ type=str,
+ help='tag-set of graph in SavedModel to scan, separated by \',\'')
+ parser_scan.set_defaults(func=scan)
+
return parser
diff --git a/tensorflow/python/tools/saved_model_cli_test.py b/tensorflow/python/tools/saved_model_cli_test.py
index f99c844..eedc893 100644
--- a/tensorflow/python/tools/saved_model_cli_test.py
+++ b/tensorflow/python/tools/saved_model_cli_test.py
@@ -525,6 +525,28 @@
y_expected = np.array([[2.5], [3.0]])
self.assertAllClose(y_expected, y_actual)
+ def testScanCommand(self):
+ self.parser = saved_model_cli.create_parser()
+ base_path = test.test_src_dir_path(SAVED_MODEL_PATH)
+ args = self.parser.parse_args(['scan', '--dir', base_path])
+ with captured_output() as (out, _):
+ saved_model_cli.scan(args)
+ output = out.getvalue().strip()
+ self.assertTrue('does not contain blacklisted ops' in output)
+
+ def testScanCommandFoundBlacklistedOp(self):
+ self.parser = saved_model_cli.create_parser()
+ base_path = test.test_src_dir_path(SAVED_MODEL_PATH)
+ args = self.parser.parse_args(
+ ['scan', '--dir', base_path, '--tag_set', 'serve'])
+ op_blacklist = saved_model_cli._OP_BLACKLIST
+ saved_model_cli._OP_BLACKLIST = set(['VariableV2'])
+ with captured_output() as (out, _):
+ saved_model_cli.scan(args)
+ saved_model_cli._OP_BLACKLIST = op_blacklist
+ output = out.getvalue().strip()
+ self.assertTrue('\'VariableV2\'' in output)
+
if __name__ == '__main__':
test.main()
diff --git a/tensorflow/stream_executor/cuda/cuda_dnn.cc b/tensorflow/stream_executor/cuda/cuda_dnn.cc
index 0b3b060..03e3e08 100644
--- a/tensorflow/stream_executor/cuda/cuda_dnn.cc
+++ b/tensorflow/stream_executor/cuda/cuda_dnn.cc
@@ -274,7 +274,8 @@
// clang-format off
#if CUDNN_VERSION >= 7000
#define CUDNN_DNN_ROUTINE_EACH_R7(__macro) \
- __macro(cudnnSetConvolutionMathType)
+ __macro(cudnnSetConvolutionMathType) \
+ __macro(cudnnSetRNNMatrixMathType)
// clang-format on
CUDNN_DNN_ROUTINE_EACH_R7(PERFTOOLS_GPUTOOLS_CUDNN_WRAP)
@@ -586,6 +587,19 @@
return is_enabled;
}
+// A helper function to decide whether to enable the TENSOR_OP_MATH math type
+// for RNNs.
+static bool RnnTensorOpMathEnabled() {
+ static bool is_enabled = [] {
+ bool is_disabled = false;
+ TF_CHECK_OK(
+ tensorflow::ReadBoolFromEnvVar("TF_DISABLE_CUDNN_RNN_TENSOR_OP_MATH",
+ /*default_val=*/false, &is_disabled));
+ return !is_disabled;
+ }();
+ return is_enabled;
+}
+
// A helper function to decide whether to use CUDNN_BATCHNORM_SPATIAL_PERSISTENT
// in batchnorm. This mode can be faster in some tasks because an optimized path
// may be selected for CUDNN_DATA_FLOAT and CUDNN_DATA_HALF data types, compute
@@ -1124,6 +1138,9 @@
SetFailure(cudnn_params_desc_->Status());
return;
}
+ if (data_type == CUDNN_DATA_HALF) {
+ set_use_tensor_op_math(true);
+ }
}
~CudnnRnnDescriptor() override {
if (rnn_desc_) {
@@ -1132,6 +1149,20 @@
CUDNN_RETURN_IF_FAIL(status, "Unable to destroy RNN descriptor");
}
}
+ void set_use_tensor_op_math(bool use_tensor_op_math) {
+#if CUDNN_VERSION >= 7000
+ cudnnMathType_t math_type =
+ (use_tensor_op_math ? CUDNN_TENSOR_OP_MATH : CUDNN_DEFAULT_MATH);
+ if (RnnTensorOpMathEnabled()) {
+ cudnnStatus_t status =
+ wrap::cudnnSetRNNMatrixMathType(parent_, rnn_desc_, math_type);
+ if (status != CUDNN_STATUS_SUCCESS) {
+ LOG(FATAL) << "could not set cudnn RNN math type: "
+ << ToString(status);
+ }
+ }
+#endif
+ }
cudnnRNNDescriptor_t handle() const {
if (!ok()) return nullptr;
return rnn_desc_;
diff --git a/tensorflow/tensorflow.bzl b/tensorflow/tensorflow.bzl
index ae305a2..a0ec170 100644
--- a/tensorflow/tensorflow.bzl
+++ b/tensorflow/tensorflow.bzl
@@ -1168,6 +1168,20 @@
"@protobuf_archive//:protobuf_headers",
clean_dep("//third_party/eigen3"),
clean_dep("//tensorflow/core:framework_headers_lib"),
+ ] + if_windows(["//tensorflow/python:pywrap_tensorflow_import_lib"])
+
+# A list of targets that contains the implemenation of
+# tf_custom_op_library_additional_deps. It's used to generate a DEF file for
+# exporting symbols from _pywrap_tensorflow.dll on Windows.
+def tf_custom_op_library_additional_deps_impl():
+ return [
+ # for @nsync//:nsync_headers
+ "//third_party/nsync:nsync_cpp",
+ # for //third_party/eigen3
+ clean_dep("//third_party/eigen3"),
+ # for //tensorflow/core:framework_headers_lib
+ clean_dep("//tensorflow/core:framework"),
+ clean_dep("//tensorflow/core:reader_base"),
]
# Traverse the dependency graph along the "deps" attribute of the
@@ -1254,6 +1268,7 @@
deps=deps + if_cuda(cuda_deps),
data=[name + "_check_deps"],
copts=tf_copts(is_external=True),
+ features = ["windows_export_all_symbols"],
linkopts=linkopts + select({
"//conditions:default": [
"-lm",
@@ -1400,7 +1415,8 @@
]) + tf_extension_copts()),
linkopts=tf_extension_linkopts() + extra_linkopts,
linkstatic=1,
- deps=deps + extra_deps)
+ deps=deps + extra_deps,
+ **kwargs)
native.genrule(
name="gen_" + cc_library_pyd_name,
srcs=[":" + cc_library_name],
diff --git a/tensorflow/tools/api/tests/api_compatibility_test.py b/tensorflow/tools/api/tests/api_compatibility_test.py
index 5268bba..baa7a08 100644
--- a/tensorflow/tools/api/tests/api_compatibility_test.py
+++ b/tensorflow/tools/api/tests/api_compatibility_test.py
@@ -247,6 +247,8 @@
public_api_visitor = public_api.PublicAPIVisitor(visitor)
public_api_visitor.do_not_descend_map['tf'].append('contrib')
public_api_visitor.do_not_descend_map['tf.GPUOptions'] = ['Experimental']
+ # TODO(annarev): Make slide_dataset available in API.
+ public_api_visitor.private_map['tf'] = ['slide_dataset']
traverse.traverse(api, public_api_visitor)
proto_dict = visitor.GetProtos()
diff --git a/tensorflow/tools/ci_build/Dockerfile.cmake b/tensorflow/tools/ci_build/Dockerfile.cmake
index ec90c83..d5dea4f 100644
--- a/tensorflow/tools/ci_build/Dockerfile.cmake
+++ b/tensorflow/tools/ci_build/Dockerfile.cmake
@@ -23,11 +23,12 @@
RUN apt-get update
RUN apt-get install -y --no-install-recommends python-pip
+RUN pip install --upgrade wheel
RUN pip install --upgrade astor
RUN pip install --upgrade gast
RUN pip install --upgrade numpy
RUN pip install --upgrade termcolor
# Install golang
-RUN add-apt-repository -y ppa:ubuntu-lxc/lxd-stable
-RUN apt-get install -y golang
+RUN apt-get install -t xenial-backports -y golang-1.9
+ENV PATH=${PATH}:/usr/lib/go-1.9/bin
diff --git a/tensorflow/tools/ci_build/windows/cpu/pip/build_tf_windows.sh b/tensorflow/tools/ci_build/windows/cpu/pip/build_tf_windows.sh
index 8b8ba31..40189a6 100644
--- a/tensorflow/tools/ci_build/windows/cpu/pip/build_tf_windows.sh
+++ b/tensorflow/tools/ci_build/windows/cpu/pip/build_tf_windows.sh
@@ -65,4 +65,5 @@
--define=no_tensorflow_py_deps=true --test_lang_filters=py \
--test_tag_filters=-no_pip,-no_windows,-no_oss \
--build_tag_filters=-no_pip,-no_windows,-no_oss --build_tests_only \
- //${PY_TEST_DIR}/tensorflow/python/...
+ //${PY_TEST_DIR}/tensorflow/python/... \
+ //${PY_TEST_DIR}/tensorflow/contrib/...
diff --git a/tensorflow/tools/def_file_filter/BUILD b/tensorflow/tools/def_file_filter/BUILD
new file mode 100644
index 0000000..e69de29
--- /dev/null
+++ b/tensorflow/tools/def_file_filter/BUILD
diff --git a/tensorflow/tools/def_file_filter/BUILD.tpl b/tensorflow/tools/def_file_filter/BUILD.tpl
new file mode 100644
index 0000000..3cb72f4
--- /dev/null
+++ b/tensorflow/tools/def_file_filter/BUILD.tpl
@@ -0,0 +1,15 @@
+# Description:
+# Tools for filtering DEF file for TensorFlow on Windows
+#
+# On Windows, we use a DEF file generated by Bazel to export
+# symbols from the tensorflow dynamic library(_pywrap_tensorflow.dll).
+# The maximum number of symbols that can be exported per DLL is 64K,
+# so we have to filter some useless symbols through this python script.
+
+package(default_visibility = ["//visibility:public"])
+
+py_binary(
+ name = "def_file_filter",
+ srcs = ["def_file_filter.py"],
+ srcs_version = "PY2AND3",
+)
diff --git a/tensorflow/tools/def_file_filter/def_file_filter.py.tpl b/tensorflow/tools/def_file_filter/def_file_filter.py.tpl
new file mode 100644
index 0000000..8bdc03e
--- /dev/null
+++ b/tensorflow/tools/def_file_filter/def_file_filter.py.tpl
@@ -0,0 +1,168 @@
+# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
+"""def_file_filter.py - tool to filter a windows def file.
+
+The def file can be used to export symbols from the tensorflow dll to enable
+tf.load_library().
+
+Because the linker allows only 64K symbols to be exported per dll
+we filter the symbols down to the essentials. The regular expressions
+we use for this are specific to tensorflow.
+
+TODO: this works fine but there is an issue with exporting
+'const char * const' and importing it from a user_ops. The problem is
+on the importing end and using __declspec(dllimport) works around it.
+"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import argparse
+import io
+import os
+import re
+import subprocess
+import sys
+import tempfile
+
+# External tools we use that come with visual studio sdk
+UNDNAME = "%{undname_bin_path}"
+
+# Exclude if matched
+EXCLUDE_RE = re.compile(r"RTTI|deleting destructor|::internal::")
+
+# Include if matched before exclude
+INCLUDEPRE_RE = re.compile(r"google::protobuf::internal::ExplicitlyConstructed|"
+ r"google::protobuf::internal::ArenaImpl::AllocateAligned|" # for contrib/data/_prefetching_ops
+ r"google::protobuf::internal::ArenaImpl::AddCleanup|" # for contrib/data/_prefetching_ops
+ r"google::protobuf::Arena::OnArenaAllocation|" # for contrib/data/_prefetching_ops
+ r"tensorflow::internal::LogMessage|"
+ r"tensorflow::internal::LogString|"
+ r"tensorflow::internal::CheckOpMessageBuilder|"
+ r"tensorflow::internal::MakeCheckOpValueString|"
+ r"tensorflow::internal::PickUnusedPortOrDie|"
+ r"tensorflow::internal::ValidateDevice|"
+ r"tensorflow::ops::internal::Enter|"
+ r"tensorflow::strings::internal::AppendPieces|"
+ r"tensorflow::strings::internal::CatPieces|"
+ r"tensorflow::io::internal::JoinPathImpl")
+
+# Include if matched after exclude
+INCLUDE_RE = re.compile(r"^(TF_\w*)$|"
+ r"^(TFE_\w*)$|"
+ r"nsync::|"
+ r"tensorflow::|"
+ r"functor::|"
+ r"perftools::gputools")
+
+# We want to identify data members explicitly in the DEF file, so that no one
+# can implicitly link against the DLL if they use one of the variables exported
+# from the DLL and the header they use does not decorate the symbol with
+# __declspec(dllimport). It is easier to detect what a data symbol does
+# NOT look like, so doing it with the below regex.
+DATA_EXCLUDE_RE = re.compile(r"[)(]|"
+ r"vftable|"
+ r"vbtable|"
+ r"vcall|"
+ r"RTTI|"
+ r"protobuf::internal::ExplicitlyConstructed")
+
+def get_args():
+ """Parse command line."""
+ filename_list = lambda x: x.split(";")
+ parser = argparse.ArgumentParser()
+ parser.add_argument("--input", type=filename_list,
+ help="paths to input def file",
+ required=True)
+ parser.add_argument("--output", help="output deffile", required=True)
+ parser.add_argument("--target", help="name of the target", required=True)
+ args = parser.parse_args()
+ return args
+
+
+def main():
+ """main."""
+ args = get_args()
+
+ # Pipe dumpbin to extract all linkable symbols from libs.
+ # Good symbols are collected in candidates and also written to
+ # a temp file.
+ candidates = []
+ tmpfile = tempfile.NamedTemporaryFile(mode="w", delete=False)
+ for def_file_path in args.input:
+ def_file = open(def_file_path, 'r')
+ for line in def_file:
+ cols = line.split()
+ sym = cols[0]
+ tmpfile.file.write(sym + "\n")
+ candidates.append(sym)
+ tmpfile.file.close()
+
+ # Run the symbols through undname to get their undecorated name
+ # so we can filter on something readable.
+ with open(args.output, "w") as def_fp:
+ # track dupes
+ taken = set()
+
+ # Header for the def file.
+ def_fp.write("LIBRARY " + args.target + "\n")
+ def_fp.write("EXPORTS\n")
+ def_fp.write("\t ??1OpDef@tensorflow@@UEAA@XZ\n")
+
+ # Each symbols returned by undname matches the same position in candidates.
+ # We compare on undname but use the decorated name from candidates.
+ dupes = 0
+ proc = subprocess.Popen([UNDNAME, tmpfile.name], stdout=subprocess.PIPE)
+ for idx, line in enumerate(io.TextIOWrapper(proc.stdout, encoding="utf-8")):
+ decorated = candidates[idx]
+ if decorated in taken:
+ # Symbol is already in output, done.
+ dupes += 1
+ continue
+
+ if not INCLUDEPRE_RE.search(line):
+ if EXCLUDE_RE.search(line):
+ continue
+ if not INCLUDE_RE.search(line):
+ continue
+
+ if "deleting destructor" in line:
+ # Some of the symbols convered by INCLUDEPRE_RE export deleting
+ # destructor symbols, which is a bad idea.
+ # So we filter out such symbols here.
+ continue
+
+ if DATA_EXCLUDE_RE.search(line):
+ def_fp.write("\t" + decorated + "\n")
+ else:
+ def_fp.write("\t" + decorated + " DATA\n")
+ taken.add(decorated)
+ def_fp.close()
+
+ exit_code = proc.wait()
+ if exit_code != 0:
+ print("{} failed, exit={}".format(UNDNAME, exit_code))
+ return exit_code
+
+ os.unlink(tmpfile.name)
+
+ print("symbols={}, taken={}, dupes={}"
+ .format(len(candidates), len(taken), dupes))
+ return 0
+
+
+if __name__ == "__main__":
+ sys.exit(main())
diff --git a/tensorflow/tools/def_file_filter/def_file_filter_configure.bzl b/tensorflow/tools/def_file_filter/def_file_filter_configure.bzl
new file mode 100644
index 0000000..47539b2
--- /dev/null
+++ b/tensorflow/tools/def_file_filter/def_file_filter_configure.bzl
@@ -0,0 +1,56 @@
+"""Repository rule for def file filter autoconfiguration.
+
+This repository reuses Bazel's VC detect mechanism to find undname.exe,
+which is a tool used in def_file_filter.py.
+
+def_file_filter.py is for filtering the DEF file for TensorFlow on Windows.
+On Windows, we use a DEF file generated by Bazel to export symbols from the
+tensorflow dynamic library(_pywrap_tensorflow.dll). The maximum number of
+symbols that can be exported per DLL is 64K, so we have to filter some useless
+symbols through this python script.
+
+`def_file_filter_config` depends on the following environment variables:
+ * `BAZEL_VC`
+ * `BAZEL_VS`
+ * `VS90COMNTOOLS`
+ * `VS100COMNTOOLS`
+ * `VS110COMNTOOLS`
+ * `VS120COMNTOOLS`
+ * `VS140COMNTOOLS`
+"""
+
+load("@bazel_tools//tools/cpp:windows_cc_configure.bzl", "find_vc_path")
+load("@bazel_tools//tools/cpp:windows_cc_configure.bzl", "find_msvc_tool")
+load("@bazel_tools//tools/cpp:lib_cc_configure.bzl", "auto_configure_fail")
+
+def _def_file_filter_configure_impl(repository_ctx):
+ if repository_ctx.os.name.lower().find("windows") == -1:
+ repository_ctx.symlink(Label("//tensorflow/tools/def_file_filter:BUILD.tpl"), "BUILD")
+ repository_ctx.file("def_file_filter.py", "")
+ return
+ vc_path = find_vc_path(repository_ctx)
+ if vc_path == "visual-studio-not-found":
+ auto_configure_fail("Visual C++ build tools not found on your machine")
+ undname_bin_path = find_msvc_tool(repository_ctx, vc_path, "undname.exe").replace("\\", "\\\\")
+
+ repository_ctx.template(
+ "def_file_filter.py",
+ Label("//tensorflow/tools/def_file_filter:def_file_filter.py.tpl"),
+ {
+ "%{undname_bin_path}": undname_bin_path,
+ })
+ repository_ctx.symlink(Label("//tensorflow/tools/def_file_filter:BUILD.tpl"), "BUILD")
+
+
+def_file_filter_configure = repository_rule(
+ implementation = _def_file_filter_configure_impl,
+ environ = [
+ "BAZEL_VC",
+ "BAZEL_VS",
+ "VS90COMNTOOLS",
+ "VS100COMNTOOLS",
+ "VS110COMNTOOLS",
+ "VS120COMNTOOLS",
+ "VS140COMNTOOLS"
+ ],
+)
diff --git a/tensorflow/tools/dist_test/README.md b/tensorflow/tools/dist_test/README.md
index c1b1f79..228d5ee 100644
--- a/tensorflow/tools/dist_test/README.md
+++ b/tensorflow/tools/dist_test/README.md
@@ -17,6 +17,14 @@
./local_test.sh --model_name CENSUS_WIDENDEEP
+You can test specify version of TensorFlow:
+
+```shell
+./local_test.sh ${whl_file_url}
+```
+
+For example, you can find these TensorFlow python package URLs from [here](https://www.tensorflow.org/install/install_linux#the_url_of_the_tensorflow_python_package) for Ubuntu.
+
**2) Launch a remote k8s cluster on Google Kubernetes Engine (GKE) and run the
test suite on it**
diff --git a/tensorflow/tools/dist_test/local_test.sh b/tensorflow/tools/dist_test/local_test.sh
index 435f9d0..caae7fd 100755
--- a/tensorflow/tools/dist_test/local_test.sh
+++ b/tensorflow/tools/dist_test/local_test.sh
@@ -16,12 +16,11 @@
#
# Tests distributed TensorFlow on a locally running TF GRPC cluster.
#
-# This script peforms the following steps:
-# 1) Build the docker-in-docker (dind) image capable of running docker and
-# Kubernetes (k8s) cluster inside.
+# This script performs the following steps:
+# 1) Build the docker image capable of running distributed TensorFlow in docker.
# 2) Run a container from the aforementioned image and start docker service
# in it
-# 3) Call a script to launch a k8s TensorFlow GRPC cluster inside the container
+# 3) Call a script to launch a distributed TensorFlow GRPC cluster inside the container
# and run the distributed test suite.
#
# Usage: local_test.sh <whl_file_location>
@@ -64,15 +63,9 @@
# Configurations
DOCKER_IMG_NAME="tensorflow/tf-dist-test-local-cluster"
-LOCAL_K8S_CACHE=${HOME}/kubernetes
-# Helper function
-get_container_id_by_image_name() {
- # Get the id of a container by image name
- # Usage: get_docker_container_id_by_image_name <img_name>
-
- docker ps | grep $1 | awk '{print $1}'
-}
+# Use TensorFlow v1.5.0 for Python 2.7 and CPU only as we set num_gpus to 0 in the below
+DEFAULT_WHL_FILE_LOCATION="https://storage.googleapis.com/tensorflow/linux/cpu/tensorflow-1.5.0-cp27-none-linux_x86_64.whl"
# Parse input arguments
LEAVE_CONTAINER_RUNNING=0
@@ -84,7 +77,8 @@
WHL_FILE_LOCATION=${1}
if [[ -z "${WHL_FILE_LOCATION}" ]]; then
- die "whl file location is not specified"
+ WHL_FILE_LOCATION=${DEFAULT_WHL_FILE_LOCATION}
+ echo "use default whl file location"
fi
while true; do
@@ -121,7 +115,7 @@
# Get utility functions
source ${DIR}/scripts/utils.sh
-# Build docker-in-docker image for local k8s cluster.
+# Build docker image for local distributed TensorFlow cluster.
NO_CACHE_FLAG=""
if [[ ! -z "${TF_DIST_DOCKER_NO_CACHE}" ]] &&
[[ "${TF_DIST_DOCKER_NO_CACHE}" != "0" ]]; then
diff --git a/tensorflow/tools/git/gen_git_source.py b/tensorflow/tools/git/gen_git_source.py
index 3630dbd..cbcdbf5 100755
--- a/tensorflow/tools/git/gen_git_source.py
+++ b/tensorflow/tools/git/gen_git_source.py
@@ -114,6 +114,13 @@
for target, src in link_map.items():
if src is None:
open(os.path.join(gen_path, target), "w").write("")
+ elif not os.path.exists(src):
+ # Git repo is configured in a way we don't support such as having
+ # packed refs. Even though in a git repo, tf.__git_version__ will not
+ # be accurate.
+ # TODO(mikecase): Support grabbing git info when using packed refs.
+ open(os.path.join(gen_path, target), "w").write("")
+ spec["git"] = False
else:
try:
# In python 3.5, symlink function exists even on Windows. But requires
diff --git a/tensorflow/tools/graph_transforms/BUILD b/tensorflow/tools/graph_transforms/BUILD
index b7d7fac..6e21aa2 100644
--- a/tensorflow/tools/graph_transforms/BUILD
+++ b/tensorflow/tools/graph_transforms/BUILD
@@ -178,6 +178,7 @@
"//tensorflow/core:test",
"//tensorflow/core:test_main",
"//tensorflow/core:testlib",
+ "//tensorflow/core/kernels:quantization_utils",
"//tensorflow/core/kernels:quantized_ops",
"//tensorflow/core/util/tensor_bundle",
],
diff --git a/tensorflow/tools/graph_transforms/fold_old_batch_norms.cc b/tensorflow/tools/graph_transforms/fold_old_batch_norms.cc
index d89afe8..d86f653 100644
--- a/tensorflow/tools/graph_transforms/fold_old_batch_norms.cc
+++ b/tensorflow/tools/graph_transforms/fold_old_batch_norms.cc
@@ -182,6 +182,36 @@
return Status::OK();
}
+Status FuseBatchNormWithBatchToSpace(const NodeMatch& match,
+ std::vector<NodeDef>* new_nodes) {
+ // Calculate the scale and offset values to apply.
+ std::vector<float> scale_values;
+ std::vector<float> offset_values;
+ TF_RETURN_IF_ERROR(
+ GetScaleAndOffsetValues(match, &scale_values, &offset_values));
+
+ // Fuse conv weights, and set the final output node name as batch_norm_node.
+ const NodeDef& batch_norm_node = match.node;
+ const NodeMatch& batch_to_space_node_match = match.inputs[0];
+ const NodeMatch& conv_node_match = batch_to_space_node_match.inputs[0];
+ const NodeDef& batch_to_space_node = batch_to_space_node_match.node;
+ const NodeDef& conv_node = conv_node_match.node;
+
+ string biasadd_name = conv_node.name() + "/biasadd";
+ TF_RETURN_IF_ERROR(
+ FuseScaleOffsetToConvWeights(scale_values, offset_values, conv_node_match,
+ biasadd_name , new_nodes));
+
+ NodeDef new_batch_to_space_node = batch_to_space_node;
+ // reuse batch_norm node name
+ new_batch_to_space_node.set_name(batch_norm_node.name());
+ new_batch_to_space_node.set_input(0, biasadd_name);
+ new_nodes->push_back(batch_to_space_node_match.inputs[1].node);
+ new_nodes->push_back(batch_to_space_node_match.inputs[2].node);
+ new_nodes->push_back(new_batch_to_space_node);
+ return Status::OK();
+}
+
Status FuseBatchNormWithConvConcat(const NodeMatch& match,
std::vector<NodeDef>* new_nodes) {
// Calculate the scale and offset values to apply.
@@ -287,6 +317,43 @@
do {
did_graph_change = false;
GraphDef replaced_graph_def;
+ TF_RETURN_IF_ERROR(ReplaceMatchingOpTypes(
+ current_graph_def, // clang-format off
+ {"BatchNormWithGlobalNormalization|FusedBatchNorm", // batch_norm_node
+ {
+ {"BatchToSpaceND", // batch_to_space_node
+ {
+ {"Conv2D", // conv_node
+ {
+ {"*"}, // input_node
+ {"Const"}, // weights_node
+ }
+ },
+ {"Const"}, // block_shape
+ {"Const"}, // crops
+ }
+ },
+ {"Const"}, // mean_node
+ {"Const"}, // variance_node
+ {"Const"}, // beta_node
+ {"Const"}, // gamma_node
+ }
+ }, // clang-format on
+ [&did_graph_change](const NodeMatch& match,
+ const std::set<string>& input_nodes,
+ const std::set<string>& output_nodes,
+ std::vector<NodeDef>* new_nodes) {
+ TF_RETURN_IF_ERROR(FuseBatchNormWithBatchToSpace(match, new_nodes));
+ did_graph_change = true;
+ return Status::OK();
+ },
+ {}, &replaced_graph_def));
+ current_graph_def = replaced_graph_def;
+ } while (did_graph_change);
+
+ do {
+ did_graph_change = false;
+ GraphDef replaced_graph_def;
// Replace BatchNorm with concat as input.
TF_RETURN_IF_ERROR(ReplaceMatchingOpTypes(
current_graph_def, // clang-format off
diff --git a/tensorflow/tools/graph_transforms/fold_old_batch_norms_test.cc b/tensorflow/tools/graph_transforms/fold_old_batch_norms_test.cc
index b30ba9a..272410c 100644
--- a/tensorflow/tools/graph_transforms/fold_old_batch_norms_test.cc
+++ b/tensorflow/tools/graph_transforms/fold_old_batch_norms_test.cc
@@ -16,6 +16,7 @@
#include "tensorflow/cc/ops/const_op.h"
#include "tensorflow/cc/ops/image_ops.h"
#include "tensorflow/cc/ops/nn_ops.h"
+#include "tensorflow/cc/ops/array_ops.h"
#include "tensorflow/cc/ops/sendrecv_ops.h"
#include "tensorflow/cc/ops/standard_ops.h"
#include "tensorflow/core/framework/tensor_testutil.h"
@@ -298,6 +299,96 @@
}
};
+void TestFoldFusedBatchNormsWithBatchToSpace() {
+ auto root = tensorflow::Scope::NewRootScope();
+ using namespace ::tensorflow::ops; // NOLINT(build/namespaces)
+
+ Tensor input_data(DT_FLOAT, TensorShape({2, 1, 3, 2}));
+ test::FillValues<float>(
+ &input_data, {1.0f, 4.0f, 2.0f, 5.0f, 3.0f, 6.0f, -1.0f, -4.0f, -2.0f,
+ -5.0f, -3.0f, -6.0f});
+ Output input_op =
+ Const(root.WithOpName("input_op"), Input::Initializer(input_data));
+
+ Tensor weights_data(DT_FLOAT, TensorShape({1, 2, 2, 2}));
+ test::FillValues<float>(&weights_data,
+ {1.0f, 2.0f, 3.0f, 4.0f, 0.1f, 0.2f, 0.3f, 0.4f});
+ Output weights_op =
+ Const(root.WithOpName("weights_op"), Input::Initializer(weights_data));
+
+ Output conv_op = Conv2D(root.WithOpName("conv_op"), input_op, weights_op,
+ {1, 1, 1, 1}, "VALID");
+
+ Tensor block_shape_data(DT_INT32, TensorShape({2}));
+ test::FillValues<int32>(&block_shape_data, {1, 2});
+ Output block_shape_op =
+ Const(root.WithOpName("block_shape_op"), Input::Initializer(block_shape_data));
+
+ Tensor crops_data(DT_INT32, TensorShape({2, 2}));
+ test::FillValues<int32>(&crops_data, {0, 0, 0, 1});
+ Output crops_op =
+ Const(root.WithOpName("crops_op"), Input::Initializer(crops_data));
+
+ Output batch_to_space_op = BatchToSpaceND(root.WithOpName("batch_to_space_op"),
+ conv_op, block_shape_op, crops_data);
+
+ Tensor mean_data(DT_FLOAT, TensorShape({2}));
+ test::FillValues<float>(&mean_data, {10.0f, 20.0f});
+ Output mean_op =
+ Const(root.WithOpName("mean_op"), Input::Initializer(mean_data));
+
+ Tensor variance_data(DT_FLOAT, TensorShape({2}));
+ test::FillValues<float>(&variance_data, {0.25f, 0.5f});
+ Output variance_op = Const(root.WithOpName("variance_op"),
+ Input::Initializer(variance_data));
+
+ Tensor beta_data(DT_FLOAT, TensorShape({2}));
+ test::FillValues<float>(&beta_data, {0.1f, 0.6f});
+ Output beta_op =
+ Const(root.WithOpName("beta_op"), Input::Initializer(beta_data));
+
+ Tensor gamma_data(DT_FLOAT, TensorShape({2}));
+ test::FillValues<float>(&gamma_data, {1.0f, 2.0f});
+ Output gamma_op =
+ Const(root.WithOpName("gamma_op"), Input::Initializer(gamma_data));
+
+ GraphDef original_graph_def;
+ TF_ASSERT_OK(root.ToGraphDef(&original_graph_def));
+
+ NodeDef batch_norm_node;
+ batch_norm_node.set_op("FusedBatchNorm");
+ batch_norm_node.set_name("output");
+ AddNodeInput("batch_to_space_op", &batch_norm_node);
+ AddNodeInput("gamma_op", &batch_norm_node);
+ AddNodeInput("beta_op", &batch_norm_node);
+ AddNodeInput("mean_op", &batch_norm_node);
+ AddNodeInput("variance_op", &batch_norm_node);
+ SetNodeAttr("T", DT_FLOAT, &batch_norm_node);
+ SetNodeAttr("epsilon", 0.00001f, &batch_norm_node);
+ SetNodeAttr("is_training", false, &batch_norm_node);
+ *(original_graph_def.mutable_node()->Add()) = batch_norm_node;
+
+ std::unique_ptr<Session> original_session(NewSession(SessionOptions()));
+ TF_ASSERT_OK(original_session->Create(original_graph_def));
+ std::vector<Tensor> original_outputs;
+ TF_ASSERT_OK(original_session->Run({}, {"output"}, {}, &original_outputs));
+
+ GraphDef fused_graph_def;
+ TF_ASSERT_OK(FoldOldBatchNorms(original_graph_def, {{}, {"output"}},
+ &fused_graph_def));
+
+ std::unique_ptr<Session> fused_session(NewSession(SessionOptions()));
+ TF_ASSERT_OK(fused_session->Create(fused_graph_def));
+ std::vector<Tensor> fused_outputs;
+ TF_ASSERT_OK(fused_session->Run({}, {"output"}, {}, &fused_outputs));
+
+ test::ExpectTensorNear<float>(original_outputs[0], fused_outputs[0], 1e-5);
+
+ for (const NodeDef& node : fused_graph_def.node()) {
+ EXPECT_NE("FusedBatchNormWithBatchToSpace", node.op());
+ }
+}
+
TEST_F(FoldOldBatchNormsTest, TestFoldOldBatchNorms) {
TestFoldOldBatchNorms();
}
@@ -315,5 +406,9 @@
TestFoldFusedBatchNormsWithConcat(/*split=*/false);
}
+TEST_F(FoldOldBatchNormsTest, TestFoldFusedBatchNormsWithBatchToSpace) {
+ TestFoldFusedBatchNormsWithBatchToSpace();
+}
+
} // namespace graph_transforms
} // namespace tensorflow
diff --git a/tensorflow/tools/pip_package/BUILD b/tensorflow/tools/pip_package/BUILD
index 1833d67..2607b9d 100644
--- a/tensorflow/tools/pip_package/BUILD
+++ b/tensorflow/tools/pip_package/BUILD
@@ -48,36 +48,65 @@
deps = ["//tensorflow:tensorflow_py"],
)
+COMMON_PIP_DEPS = [
+ ":licenses",
+ "MANIFEST.in",
+ "README",
+ "setup.py",
+ ":included_headers",
+ "//tensorflow:tensorflow_py",
+ "//tensorflow/contrib/boosted_trees:boosted_trees_pip",
+ "//tensorflow/contrib/cluster_resolver:cluster_resolver_pip",
+ "//tensorflow/contrib/data/python/kernel_tests:dataset_serialization_test",
+ "//tensorflow/contrib/data/python/ops:contrib_op_loader",
+ "//tensorflow/contrib/eager/python/examples:examples_pip",
+ "//tensorflow/contrib/eager/python:checkpointable_utils",
+ "//tensorflow/contrib/eager/python:evaluator",
+ "//tensorflow/contrib/gan:gan",
+ "//tensorflow/contrib/graph_editor:graph_editor_pip",
+ "//tensorflow/contrib/keras:keras",
+ "//tensorflow/contrib/labeled_tensor:labeled_tensor_pip",
+ "//tensorflow/contrib/nn:nn_py",
+ "//tensorflow/contrib/predictor:predictor_pip",
+ "//tensorflow/contrib/py2tf:py2tf",
+ "//tensorflow/contrib/py2tf/converters:converters",
+ "//tensorflow/contrib/py2tf/converters:test_lib",
+ "//tensorflow/contrib/py2tf/impl:impl",
+ "//tensorflow/contrib/py2tf/pyct:pyct",
+ "//tensorflow/contrib/py2tf/pyct/static_analysis:static_analysis",
+ "//tensorflow/contrib/receptive_field:receptive_field_pip",
+ "//tensorflow/contrib/session_bundle:session_bundle_pip",
+ "//tensorflow/contrib/signal:signal_py",
+ "//tensorflow/contrib/signal:test_util",
+ "//tensorflow/contrib/slim:slim",
+ "//tensorflow/contrib/slim/python/slim/data:data_pip",
+ "//tensorflow/contrib/slim/python/slim/nets:nets_pip",
+ "//tensorflow/contrib/specs:specs",
+ "//tensorflow/contrib/summary:summary_test_util",
+ "//tensorflow/contrib/tensor_forest:init_py",
+ "//tensorflow/contrib/tensor_forest/hybrid:hybrid_pip",
+ "//tensorflow/contrib/timeseries:timeseries_pip",
+ "//tensorflow/contrib/tpu",
+ "//tensorflow/examples/tutorials/mnist:package",
+ "//tensorflow/python:distributed_framework_test_lib",
+ "//tensorflow/python:meta_graph_testdata",
+ "//tensorflow/python:spectral_ops_test_util",
+ "//tensorflow/python:util_example_parser_configuration",
+ "//tensorflow/python/debug:debug_pip",
+ "//tensorflow/python/eager:eager_pip",
+ "//tensorflow/python/saved_model:saved_model",
+ "//tensorflow/python/tools:tools_pip",
+ "//tensorflow/python:test_ops",
+ "//tensorflow/tools/dist_test/server:grpc_tensorflow_server",
+]
+
# On Windows, python binary is a zip file of runfiles tree.
# Add everything to its data dependency for generating a runfiles tree
# for building the pip package on Windows.
py_binary(
name = "simple_console_for_windows",
srcs = ["simple_console_for_windows.py"],
- data = [
- "MANIFEST.in",
- "README",
- "setup.py",
- ":included_headers",
- "//tensorflow/contrib/nn:nn_py",
- "//tensorflow/contrib/session_bundle:session_bundle_pip",
- "//tensorflow/contrib/signal:signal_py",
- "//tensorflow/contrib/slim/python/slim/data:data_pip",
- "//tensorflow/python:util_example_parser_configuration",
- "//tensorflow/python/debug:debug_pip",
- "//tensorflow/python/saved_model",
- "//tensorflow/python:spectral_ops_test_util",
- "//tensorflow/python/tools:tools_pip",
- "//tensorflow/python/eager:eager_pip",
- "//tensorflow/contrib/summary:summary_test_util",
- # These targets don't build on Windows yet. Exclude them for now.
- # "//tensorflow/contrib/slim",
- # "//tensorflow/contrib/slim/python/slim/nets:nets_pip",
- # "//tensorflow/contrib/specs",
- # "//tensorflow/contrib/tensor_forest:init_py",
- # "//tensorflow/contrib/tensor_forest/hybrid:hybrid_pip",
- # "//tensorflow/examples/tutorials/mnist:package",
- ],
+ data = COMMON_PIP_DEPS,
srcs_version = "PY2AND3",
deps = ["//tensorflow:tensorflow_py"],
)
@@ -108,6 +137,7 @@
"@highwayhash//:LICENSE",
"@jemalloc//:COPYING",
"@jpeg//:LICENSE.md",
+ "@kafka//:LICENSE",
"@libxsmm_archive//:LICENSE",
"@lmdb//:LICENSE",
"@local_config_sycl//sycl:LICENSE.text",
@@ -137,61 +167,12 @@
data = select({
"//tensorflow:windows": [":simple_console_for_windows"],
"//tensorflow:windows_msvc": [":simple_console_for_windows"],
- "//conditions:default": [
- ":licenses",
- "MANIFEST.in",
- "README",
- "setup.py",
- ":included_headers",
+ "//conditions:default": COMMON_PIP_DEPS + [
":simple_console",
- "//tensorflow:tensorflow_py",
- "//tensorflow/contrib/boosted_trees:boosted_trees_pip",
- "//tensorflow/contrib/cluster_resolver:cluster_resolver_pip",
- "//tensorflow/contrib/data/python/kernel_tests:dataset_serialization_test",
- "//tensorflow/contrib/data/python/ops:contrib_op_loader",
- "//tensorflow/contrib/eager/python/examples:examples_pip",
- "//tensorflow/contrib/eager/python:checkpointable_utils",
- "//tensorflow/contrib/eager/python:evaluator",
- "//tensorflow/contrib/gan:gan",
- "//tensorflow/contrib/graph_editor:graph_editor_pip",
- "//tensorflow/contrib/keras:keras",
- "//tensorflow/contrib/labeled_tensor:labeled_tensor_pip",
"//tensorflow/contrib/lite/python:interpreter_test_data",
"//tensorflow/contrib/lite/toco:toco",
"//tensorflow/contrib/lite/toco/python:toco_wrapper",
"//tensorflow/contrib/lite/toco/python:toco_from_protos",
- "//tensorflow/contrib/nn:nn_py",
- "//tensorflow/contrib/predictor:predictor_pip",
- "//tensorflow/contrib/py2tf:py2tf",
- "//tensorflow/contrib/py2tf/converters:converters",
- "//tensorflow/contrib/py2tf/converters:test_lib",
- "//tensorflow/contrib/py2tf/impl:impl",
- "//tensorflow/contrib/py2tf/pyct:pyct",
- "//tensorflow/contrib/py2tf/pyct/static_analysis:static_analysis",
- "//tensorflow/contrib/receptive_field:receptive_field_pip",
- "//tensorflow/contrib/session_bundle:session_bundle_pip",
- "//tensorflow/contrib/signal:signal_py",
- "//tensorflow/contrib/signal:test_util",
- "//tensorflow/contrib/slim:slim",
- "//tensorflow/contrib/slim/python/slim/data:data_pip",
- "//tensorflow/contrib/slim/python/slim/nets:nets_pip",
- "//tensorflow/contrib/specs:specs",
- "//tensorflow/contrib/summary:summary_test_util",
- "//tensorflow/contrib/tensor_forest:init_py",
- "//tensorflow/contrib/tensor_forest/hybrid:hybrid_pip",
- "//tensorflow/contrib/timeseries:timeseries_pip",
- "//tensorflow/contrib/tpu",
- "//tensorflow/examples/tutorials/mnist:package",
- "//tensorflow/python:distributed_framework_test_lib",
- "//tensorflow/python:meta_graph_testdata",
- "//tensorflow/python:spectral_ops_test_util",
- "//tensorflow/python:util_example_parser_configuration",
- "//tensorflow/python/debug:debug_pip",
- "//tensorflow/python/eager:eager_pip",
- "//tensorflow/python/saved_model:saved_model",
- "//tensorflow/python/tools:tools_pip",
- "//tensorflow/python:test_ops",
- "//tensorflow/tools/dist_test/server:grpc_tensorflow_server",
],
}) + if_mkl(["//third_party/mkl:intel_binary_blob"]) + if_tensorrt([
"//tensorflow/contrib/tensorrt:init_py",
diff --git a/tensorflow/tools/pip_package/setup.py b/tensorflow/tools/pip_package/setup.py
index e1a5f09..e0152da 100644
--- a/tensorflow/tools/pip_package/setup.py
+++ b/tensorflow/tools/pip_package/setup.py
@@ -29,7 +29,7 @@
# This version string is semver compatible, but incompatible with pip.
# For pip, we will remove all '-' characters from this string, and use the
# result for pip.
-_VERSION = '1.6.0-rc1'
+_VERSION = '1.6.0'
REQUIRED_PACKAGES = [
'absl-py >= 0.1.6',
@@ -72,7 +72,7 @@
# pylint: disable=line-too-long
CONSOLE_SCRIPTS = [
- 'freeze_graph = tensorflow.python.tools.freeze_graph:main',
+ 'freeze_graph = tensorflow.python.tools.freeze_graph:run_main',
'toco_from_protos = tensorflow.contrib.lite.toco.python.toco_from_protos:main',
'toco = tensorflow.contrib.lite.toco.python.toco_wrapper:main',
'saved_model_cli = tensorflow.python.tools.saved_model_cli:main',
diff --git a/tensorflow/tools/test/upload_test_benchmarks.py b/tensorflow/tools/test/upload_test_benchmarks.py
index 77cc9f7..edd0935 100644
--- a/tensorflow/tools/test/upload_test_benchmarks.py
+++ b/tensorflow/tools/test/upload_test_benchmarks.py
@@ -88,6 +88,7 @@
import shutil
from google.cloud import datastore
+from six import text_type
def is_real_file(dirpath, fname):
@@ -150,7 +151,7 @@
"""
test_result = json.loads(data)
- test_name = unicode(test_result["name"])
+ test_name = text_type(test_result["name"])
start_time = datetime.datetime.utcfromtimestamp(
float(test_result["startTime"]))
batch = []
@@ -162,7 +163,7 @@
t_val.update({
"test": test_name,
"start": start_time,
- "info": unicode(data)
+ "info": text_type(data)
})
batch.append(t_val)
@@ -170,7 +171,7 @@
# the attribute to be fetched and displayed. The full entry information is
# also stored as a non-indexed JSON blob.
for ent in test_result["entries"].get("entry", []):
- ent_name = unicode(ent["name"])
+ ent_name = text_type(ent["name"])
e_key = client.key("Entry")
e_val = datastore.Entity(e_key, exclude_from_indexes=["info"])
e_val.update({
@@ -178,7 +179,7 @@
"start": start_time,
"entry": ent_name,
"timing": ent["wallTime"],
- "info": unicode(json.dumps(ent))
+ "info": text_type(json.dumps(ent))
})
batch.append(e_val)
diff --git a/tensorflow/workspace.bzl b/tensorflow/workspace.bzl
index abc9eb9..a922808 100644
--- a/tensorflow/workspace.bzl
+++ b/tensorflow/workspace.bzl
@@ -12,6 +12,8 @@
load("//third_party:repo.bzl", "tf_http_archive")
load("@io_bazel_rules_closure//closure/private:java_import_external.bzl", "java_import_external")
load("@io_bazel_rules_closure//closure:defs.bzl", "filegroup_external")
+load("//tensorflow/tools/def_file_filter:def_file_filter_configure.bzl",
+ "def_file_filter_configure")
def _extract_version_number(bazel_version):
"""Extracts the semantic version number from a version string
@@ -67,7 +69,7 @@
# We must check the bazel version before trying to parse any other BUILD
# files, in case the parsing of those build files depends on the bazel
# version we require here.
- check_bazel_version_at_least("0.5.4")
+ check_bazel_version_at_least("0.10.0")
clang6_configure(name="local_config_clang6")
cuda_configure(name="local_config_cuda")
tensorrt_configure(name="local_config_tensorrt")
@@ -75,6 +77,10 @@
sycl_configure(name="local_config_sycl")
python_configure(name="local_config_python")
+ # For windows bazel build
+ # TODO: Remove def file filter when TensorFlow can export symbols properly on Windows.
+ def_file_filter_configure(name = "local_config_def_file_filter")
+
# Point //external/local_config_arm_compiler to //external/arm_compiler
arm_compiler_configure(
name="local_config_arm_compiler",
diff --git a/third_party/jpeg/jpeg.BUILD b/third_party/jpeg/jpeg.BUILD
index 87a2392..4418ac3 100644
--- a/third_party/jpeg/jpeg.BUILD
+++ b/third_party/jpeg/jpeg.BUILD
@@ -526,12 +526,12 @@
config_setting(
name = "armeabi-v7a",
- values = {"android_cpu": "armeabi-v7a"},
+ values = {"cpu": "armeabi-v7a"},
)
config_setting(
name = "arm64-v8a",
- values = {"android_cpu": "arm64-v8a"},
+ values = {"cpu": "arm64-v8a"},
)
config_setting(
diff --git a/third_party/kafka/BUILD b/third_party/kafka/BUILD
index a61a9e1..a839ca7 100644
--- a/third_party/kafka/BUILD
+++ b/third_party/kafka/BUILD
@@ -130,18 +130,23 @@
],
hdrs = [
"config.h",
+ "src-cpp/rdkafkacpp.h",
+ "src-cpp/rdkafkacpp_int.h",
+ "src/lz4.c",
+ "src/snappy_compat.h",
+ ],
+ copts = [
+ "-Iexternal/kafka/src",
+ "-Iexternal/kafka/src-cpp",
],
defines = [
],
- includes = [
- "src",
- "src-cpp",
- ],
linkopts = [
"-lpthread",
],
visibility = ["//visibility:public"],
deps = [
"@boringssl//:ssl",
+ "@zlib_archive//:zlib",
],
)
diff --git a/third_party/py/BUILD.tpl b/third_party/py/BUILD.tpl
index de06ad5..1dd8ab4 100644
--- a/third_party/py/BUILD.tpl
+++ b/third_party/py/BUILD.tpl
@@ -2,20 +2,26 @@
package(default_visibility = ["//visibility:public"])
+# To build Python C/C++ extension on Windows, we need to link to python import library pythonXY.lib
+# See https://docs.python.org/3/extending/windows.html
+cc_import(
+ name = "python_lib",
+ interface_library = select({
+ ":windows": ":python_import_lib",
+ # A placeholder for Unix platforms which makes --no_build happy.
+ "//conditions:default": "not-existing.lib",
+ }),
+ system_provided = 1,
+)
+
cc_library(
name = "python_headers",
hdrs = [":python_include"],
- data = select({
- ":windows": [":python_import_lib"],
+ deps = select({
+ ":windows": [":python_lib"],
"//conditions:default": [],
}),
includes = ["python_include"],
- linkopts = select({
- # TODO(pcloudy): Ideally, this should just go into deps after resolving
- # https://github.com/bazelbuild/bazel/issues/3237,
- ":windows": ["$(locations :python_import_lib)"],
- "//conditions:default": [],
- }),
)
cc_library(