arm_compute v18.05
diff --git a/examples/SConscript b/examples/SConscript
index 9be9fa9..c3576fb 100644
--- a/examples/SConscript
+++ b/examples/SConscript
@@ -24,75 +24,65 @@
Import('env')
-if env['opencl']:
- Import('opencl')
-
-if env['gles_compute'] and env['os'] != 'android':
- Import('egl')
- Import('glesv2')
-
examples_env = env.Clone()
examples_env.Append(CPPPATH = ["#"])
-examples_env.Append(LIBPATH = ["#build/%s/opencl-1.2-stubs" % env['build_dir']])
# Build examples
utils = examples_env.Object("../utils/Utils.cpp")
if env['os'] in ['android', 'bare_metal'] or env['standalone']:
+ Import('arm_compute_graph_a')
Import('arm_compute_a')
Import('arm_compute_core_a')
arm_compute_libs = [ arm_compute_a, arm_compute_core_a ]
arm_compute_dependency = arm_compute_a
+ graph_dependency = [arm_compute_graph_a]
else:
+ Import('arm_compute_graph_so')
Import('arm_compute_so')
- arm_compute_libs = ["arm_compute", "arm_compute_core"]
+ arm_compute_libs = ["arm_compute_graph", "arm_compute", "arm_compute_core"]
arm_compute_dependency = arm_compute_so
+ graph_dependency = [arm_compute_graph_so]
-if env['opencl'] and env['neon']:
+# Build graph examples
+graph_utils = examples_env.Object("../utils/GraphUtils.cpp")
+examples_libs = examples_env.get("LIBS",[])
+for file in Glob("./graph_*.cpp"):
+ example = os.path.basename(os.path.splitext(str(file))[0])
+ prog = None
+ arm_compute_graph_libs = arm_compute_libs
+
+ if env['os'] in ['android', 'bare_metal'] or env['standalone']:
+ prog = examples_env.Program(example, ["{}.cpp".format(example), utils, graph_utils], LIBS = examples_libs + arm_compute_graph_libs, LINKFLAGS=examples_env["LINKFLAGS"]+['-Wl,--whole-archive',graph_dependency,'-Wl,--no-whole-archive'])
+ Depends(prog, graph_dependency)
+ else:
+ #-Wl,--allow-shlib-undefined: Ignore dependencies of dependencies
+ prog = examples_env.Program(example, ["{}.cpp".format(example), utils, graph_utils], LIBS = examples_libs + arm_compute_graph_libs, LINKFLAGS=examples_env["LINKFLAGS"]+['-Wl,--allow-shlib-undefined'] )
+ Depends(prog, graph_dependency)
+ alias = examples_env.Alias(example, prog)
+ Default(alias)
+
+if env['opencl'] and env['neon']:
for file in Glob("./neoncl_*.cpp"):
example = os.path.basename(os.path.splitext(str(file))[0])
- prog = examples_env.Program(example, ["{}.cpp".format(example), utils], CPPDEFINES=['ARM_COMPUTE_CL'], LIBS = arm_compute_libs +["OpenCL"])
- Depends(prog, [arm_compute_dependency, opencl])
- alias = examples_env.Alias(example, prog)
- Default(alias)
- if env['os'] == 'android':
- Import('arm_compute_graph_a')
- Import('arm_compute_core_a')
- Import('arm_compute_a')
- arm_compute_graph_libs = [ arm_compute_a, arm_compute_core_a, "OpenCL"]
- graph_dependency = arm_compute_graph_a
- else:
- Import('arm_compute_graph_so')
- arm_compute_graph_libs = ["arm_compute_graph", "arm_compute", "arm_compute_core"]
- graph_dependency = arm_compute_graph_so
-
- graph_utils = examples_env.Object("../utils/GraphUtils.cpp")
- for file in Glob("./graph_*.cpp"):
- example = os.path.basename(os.path.splitext(str(file))[0])
- prog = None
- if env['os'] == 'android':
- prog = examples_env.Program(example, ["{}.cpp".format(example), utils, graph_utils], LIBS = arm_compute_graph_libs + ["OpenCL"], LINKFLAGS=examples_env["LINKFLAGS"]+['-Wl,--whole-archive',graph_dependency,'-Wl,--no-whole-archive'])
- Depends(prog, [graph_dependency, opencl])
- else:
- #-Wl,--allow-shlib-undefined: Ignore dependencies of dependencies
- prog = examples_env.Program(example, ["{}.cpp".format(example), utils, graph_utils], LIBS = arm_compute_graph_libs, LINKFLAGS=examples_env["LINKFLAGS"]+['-Wl,--allow-shlib-undefined'] )
- Depends(prog, graph_dependency)
+ prog = examples_env.Program(example, ["{}.cpp".format(example), utils], CPPDEFINES=['ARM_COMPUTE_CL'], LIBS = examples_libs + arm_compute_libs)
+ Depends(prog, arm_compute_dependency)
alias = examples_env.Alias(example, prog)
Default(alias)
if env['opencl']:
for file in Glob("./cl_*.cpp"):
example = os.path.basename(os.path.splitext(str(file))[0])
- prog = examples_env.Program(example, ["{}.cpp".format(example), utils], CPPDEFINES=['ARM_COMPUTE_CL'], LIBS = arm_compute_libs +["OpenCL"])
- Depends(prog, [arm_compute_dependency, opencl])
+ prog = examples_env.Program(example, ["{}.cpp".format(example), utils], CPPDEFINES=['ARM_COMPUTE_CL'], LIBS = examples_libs + arm_compute_libs)
+ Depends(prog, arm_compute_dependency)
alias = examples_env.Alias(example, prog)
Default(alias)
if env['neon']:
for file in Glob("./neon_*.cpp"):
example = os.path.basename(os.path.splitext(str(file))[0])
- prog = examples_env.Program(example, ["{}.cpp".format(example), utils], LIBS = arm_compute_libs)
+ prog = examples_env.Program(example, ["{}.cpp".format(example), utils], LIBS = examples_libs + arm_compute_libs)
Depends(prog, arm_compute_dependency)
alias = examples_env.Alias(example, prog)
Default(alias)
@@ -100,14 +90,7 @@
if env['gles_compute']:
for file in Glob("./gc_*.cpp"):
example = os.path.basename(os.path.splitext(str(file))[0])
- if env['os'] != 'android':
- prog = examples_env.Program(example, ["{}.cpp".format(example), utils], CPPDEFINES=['ARM_COMPUTE_GC'], LIBS = [arm_compute_libs, "EGL", "GLESv2"])
- Depends(prog, [arm_compute_dependency, egl, glesv2])
- else:
- if env['arch'] != 'armv7a':
- prog = examples_env.Program(example, ["{}.cpp".format(example), utils], CPPDEFINES=['ARM_COMPUTE_GC'], LIBS = [arm_compute_libs, "EGL", "GLESv3"])
- else:
- prog = examples_env.Program(example, ["{}.cpp".format(example), utils], CPPDEFINES=['ARM_COMPUTE_GC'], LIBS = [arm_compute_libs, "EGL", "GLESv2"])
- Depends(prog, [arm_compute_dependency])
+ prog = examples_env.Program(example, ["{}.cpp".format(example), utils], CPPDEFINES=['ARM_COMPUTE_GC'], LIBS = examples_libs + arm_compute_libs)
+ Depends(prog, arm_compute_dependency)
alias = examples_env.Alias(example, prog)
Default(alias)
diff --git a/examples/cl_sgemm.cpp b/examples/cl_sgemm.cpp
index 966661b..fa57885 100644
--- a/examples/cl_sgemm.cpp
+++ b/examples/cl_sgemm.cpp
@@ -198,4 +198,4 @@
int main(int argc, char **argv)
{
return utils::run_example<CLSGEMMExample>(argc, argv);
-}
\ No newline at end of file
+}
diff --git a/examples/graph_alexnet.cpp b/examples/graph_alexnet.cpp
index a396c76..9e6d919 100644
--- a/examples/graph_alexnet.cpp
+++ b/examples/graph_alexnet.cpp
@@ -21,8 +21,7 @@
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
-#include "arm_compute/graph/Graph.h"
-#include "arm_compute/graph/Nodes.h"
+#include "arm_compute/graph.h"
#include "support/ToolchainSupport.h"
#include "utils/GraphUtils.h"
#include "utils/Utils.h"
@@ -32,13 +31,13 @@
#include <memory>
using namespace arm_compute::utils;
-using namespace arm_compute::graph;
+using namespace arm_compute::graph::frontend;
using namespace arm_compute::graph_utils;
/** Example demonstrating how to implement AlexNet's network using the Compute Library's graph API
*
* @param[in] argc Number of arguments
- * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels )
+ * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) )
*/
class GraphAlexnetExample : public Example
{
@@ -54,56 +53,69 @@
std::unique_ptr<IPreprocessor> preprocessor = arm_compute::support::cpp14::make_unique<CaffePreproccessor>(mean_rgb);
// Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON
- const int int_target_hint = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
- TargetHint target_hint = set_target_hint(int_target_hint);
+ const int target = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
+ Target target_hint = set_target_hint(target);
- const bool is_gemm_convolution5x5 = Graph::gpu_target() == arm_compute::GPUTarget::MIDGARD || target_hint == TargetHint::NEON;
- ConvolutionMethodHint convolution_5x5_hint = is_gemm_convolution5x5 ? ConvolutionMethodHint::GEMM : ConvolutionMethodHint::DIRECT;
+ const bool is_neon = (target_hint == Target::NEON);
+ ConvolutionMethod convolution_5x5_hint = is_neon ? ConvolutionMethod::GEMM : ConvolutionMethod::DIRECT;
+ ConvolutionMethod convolution_3x3_hint = ConvolutionMethod::DEFAULT;
+ FastMathHint fast_math_hint = FastMathHint::DISABLED;
// Parse arguments
if(argc < 2)
{
// Print help
- std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels] [fast_math_hint]\n\n";
std::cout << "No data folder provided: using random values\n\n";
}
else if(argc == 2)
{
- std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels] [fast_math_hint]\n\n";
std::cout << "No data folder provided: using random values\n\n";
}
else if(argc == 3)
{
data_path = argv[2];
- std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels] [fast_math_hint]\n\n";
std::cout << "No image provided: using random values\n\n";
}
else if(argc == 4)
{
data_path = argv[2];
image = argv[3];
- std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels] [fast_math_hint]\n\n";
std::cout << "No text file with labels provided: skipping output accessor\n\n";
}
- else
+ else if(argc == 5)
{
data_path = argv[2];
image = argv[3];
label = argv[4];
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " " << argv[4] << " [fast_math_hint]\n\n";
+ std::cout << "No fast math info provided: disabling fast math\n\n";
+ }
+ else
+ {
+ data_path = argv[2];
+ image = argv[3];
+ label = argv[4];
+ fast_math_hint = (std::strtol(argv[5], nullptr, 1) == 0) ? FastMathHint::DISABLED : FastMathHint::ENABLED;
}
graph << target_hint
- << Tensor(TensorInfo(TensorShape(227U, 227U, 3U, 1U), 1, DataType::F32),
- get_input_accessor(image, std::move(preprocessor)))
+ << fast_math_hint
+ << InputLayer(TensorDescriptor(TensorShape(227U, 227U, 3U, 1U), DataType::F32),
+ get_input_accessor(image, std::move(preprocessor)))
// Layer 1
<< ConvolutionLayer(
11U, 11U, 96U,
get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv1_w.npy"),
get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv1_b.npy"),
PadStrideInfo(4, 4, 0, 0))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
- << NormalizationLayer(NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f))
- << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0)))
+ .set_name("conv1")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu1")
+ << NormalizationLayer(NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f)).set_name("norm1")
+ << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0))).set_name("pool1")
// Layer 2
<< convolution_5x5_hint
<< ConvolutionLayer(
@@ -111,55 +123,64 @@
get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv2_w.npy"),
get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv2_b.npy"),
PadStrideInfo(1, 1, 2, 2), 2)
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
- << NormalizationLayer(NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f))
- << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0)))
- << ConvolutionMethodHint::GEMM
+ .set_name("conv2")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu2")
+ << NormalizationLayer(NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f)).set_name("norm2")
+ << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0))).set_name("pool2")
+ << convolution_3x3_hint
// Layer 3
<< ConvolutionLayer(
3U, 3U, 384U,
get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv3_w.npy"),
get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv3_b.npy"),
PadStrideInfo(1, 1, 1, 1))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ .set_name("conv3")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu3")
// Layer 4
<< ConvolutionLayer(
3U, 3U, 384U,
get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv4_w.npy"),
get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv4_b.npy"),
PadStrideInfo(1, 1, 1, 1), 2)
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ .set_name("conv4")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu4")
// Layer 5
<< ConvolutionLayer(
3U, 3U, 256U,
get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv5_w.npy"),
get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv5_b.npy"),
PadStrideInfo(1, 1, 1, 1), 2)
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
- << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0)))
+ .set_name("conv5")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu5")
+ << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0))).set_name("pool5")
// Layer 6
<< FullyConnectedLayer(
4096U,
get_weights_accessor(data_path, "/cnn_data/alexnet_model/fc6_w.npy"),
get_weights_accessor(data_path, "/cnn_data/alexnet_model/fc6_b.npy"))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ .set_name("fc6")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu6")
// Layer 7
<< FullyConnectedLayer(
4096U,
get_weights_accessor(data_path, "/cnn_data/alexnet_model/fc7_w.npy"),
get_weights_accessor(data_path, "/cnn_data/alexnet_model/fc7_b.npy"))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ .set_name("fc7")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu7")
// Layer 8
<< FullyConnectedLayer(
1000U,
get_weights_accessor(data_path, "/cnn_data/alexnet_model/fc8_w.npy"),
get_weights_accessor(data_path, "/cnn_data/alexnet_model/fc8_b.npy"))
+ .set_name("fc8")
// Softmax
- << SoftmaxLayer()
- << Tensor(get_output_accessor(label, 5));
+ << SoftmaxLayer().set_name("prob")
+ << OutputLayer(get_output_accessor(label, 5));
- // In order to enable the OpenCL tuner, graph_init() has to be called only when all nodes have been instantiated
- graph.graph_init(int_target_hint == 2);
+ // Finalize graph
+ GraphConfig config;
+ config.use_tuner = (target == 2);
+ graph.finalize(target_hint, config);
}
void do_run() override
{
@@ -168,13 +189,13 @@
}
private:
- Graph graph{};
+ Stream graph{ 0, "AlexNet" };
};
/** Main program for AlexNet
*
* @param[in] argc Number of arguments
- * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL), [optional] Path to the weights folder, [optional] image, [optional] labels )
+ * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) )
*/
int main(int argc, char **argv)
{
diff --git a/examples/graph_googlenet.cpp b/examples/graph_googlenet.cpp
index de4afa2..2dba67f 100644
--- a/examples/graph_googlenet.cpp
+++ b/examples/graph_googlenet.cpp
@@ -21,9 +21,7 @@
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
-#include "arm_compute/graph/Graph.h"
-#include "arm_compute/graph/Nodes.h"
-#include "arm_compute/graph/SubGraph.h"
+#include "arm_compute/graph.h"
#include "support/ToolchainSupport.h"
#include "utils/GraphUtils.h"
#include "utils/Utils.h"
@@ -32,13 +30,13 @@
#include <tuple>
using namespace arm_compute::utils;
-using namespace arm_compute::graph;
+using namespace arm_compute::graph::frontend;
using namespace arm_compute::graph_utils;
/** Example demonstrating how to implement Googlenet's network using the Compute Library's graph API
*
* @param[in] argc Number of arguments
- * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL), [optional] Path to the weights folder, [optional] image, [optional] labels )
+ * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) )
*/
class GraphGooglenetExample : public Example
{
@@ -54,45 +52,55 @@
std::unique_ptr<IPreprocessor> preprocessor = arm_compute::support::cpp14::make_unique<CaffePreproccessor>(mean_rgb);
// Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON
- const int int_target_hint = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
- TargetHint target_hint = set_target_hint(int_target_hint);
- ConvolutionMethodHint convolution_hint = ConvolutionMethodHint::GEMM;
+ const int target = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
+ Target target_hint = set_target_hint(target);
+ FastMathHint fast_math_hint = FastMathHint::DISABLED;
// Parse arguments
if(argc < 2)
{
// Print help
- std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels] [fast_math_hint]\n\n";
std::cout << "No data folder provided: using random values\n\n";
}
else if(argc == 2)
{
- std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels] [fast_math_hint]\n\n";
std::cout << "No data folder provided: using random values\n\n";
}
else if(argc == 3)
{
data_path = argv[2];
- std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels] [fast_math_hint]\n\n";
std::cout << "No image provided: using random values\n\n";
}
else if(argc == 4)
{
data_path = argv[2];
image = argv[3];
- std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels] [fast_math_hint]\n\n";
std::cout << "No text file with labels provided: skipping output accessor\n\n";
}
- else
+ else if(argc == 5)
{
data_path = argv[2];
image = argv[3];
label = argv[4];
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " " << argv[4] << " [fast_math_hint]\n\n";
+ std::cout << "No fast math info provided: disabling fast math\n\n";
+ }
+ else
+ {
+ data_path = argv[2];
+ image = argv[3];
+ label = argv[4];
+ fast_math_hint = (std::strtol(argv[5], nullptr, 1) == 0) ? FastMathHint::DISABLED : FastMathHint::ENABLED;
}
graph << target_hint
- << Tensor(TensorInfo(TensorShape(224U, 224U, 3U, 1U), 1, DataType::F32),
- get_input_accessor(image, std::move(preprocessor)))
+ << fast_math_hint
+ << InputLayer(TensorDescriptor(TensorShape(224U, 224U, 3U, 1U), DataType::F32),
+ get_input_accessor(image, std::move(preprocessor)))
<< ConvolutionLayer(
7U, 7U, 64U,
get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv1/conv1_7x7_s2_w.npy"),
@@ -101,7 +109,6 @@
<< ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
<< PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
<< NormalizationLayer(NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f))
- << convolution_hint
<< ConvolutionLayer(
1U, 1U, 64U,
get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv2/conv2_3x3_reduce_w.npy"),
@@ -115,28 +122,30 @@
PadStrideInfo(1, 1, 1, 1))
<< ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
<< NormalizationLayer(NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f))
- << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
- << get_inception_node(data_path, "inception_3a", 64, std::make_tuple(96U, 128U), std::make_tuple(16U, 32U), 32U)
- << get_inception_node(data_path, "inception_3b", 128, std::make_tuple(128U, 192U), std::make_tuple(32U, 96U), 64U)
- << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
- << get_inception_node(data_path, "inception_4a", 192, std::make_tuple(96U, 208U), std::make_tuple(16U, 48U), 64U)
- << get_inception_node(data_path, "inception_4b", 160, std::make_tuple(112U, 224U), std::make_tuple(24U, 64U), 64U)
- << get_inception_node(data_path, "inception_4c", 128, std::make_tuple(128U, 256U), std::make_tuple(24U, 64U), 64U)
- << get_inception_node(data_path, "inception_4d", 112, std::make_tuple(144U, 288U), std::make_tuple(32U, 64U), 64U)
- << get_inception_node(data_path, "inception_4e", 256, std::make_tuple(160U, 320U), std::make_tuple(32U, 128U), 128U)
- << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
- << get_inception_node(data_path, "inception_5a", 256, std::make_tuple(160U, 320U), std::make_tuple(32U, 128U), 128U)
- << get_inception_node(data_path, "inception_5b", 384, std::make_tuple(192U, 384U), std::make_tuple(48U, 128U), 128U)
- << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 7, PadStrideInfo(1, 1, 0, 0, DimensionRoundingType::CEIL)))
+ << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)));
+ graph << get_inception_node(data_path, "inception_3a", 64, std::make_tuple(96U, 128U), std::make_tuple(16U, 32U), 32U);
+ graph << get_inception_node(data_path, "inception_3b", 128, std::make_tuple(128U, 192U), std::make_tuple(32U, 96U), 64U);
+ graph << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)));
+ graph << get_inception_node(data_path, "inception_4a", 192, std::make_tuple(96U, 208U), std::make_tuple(16U, 48U), 64U);
+ graph << get_inception_node(data_path, "inception_4b", 160, std::make_tuple(112U, 224U), std::make_tuple(24U, 64U), 64U);
+ graph << get_inception_node(data_path, "inception_4c", 128, std::make_tuple(128U, 256U), std::make_tuple(24U, 64U), 64U);
+ graph << get_inception_node(data_path, "inception_4d", 112, std::make_tuple(144U, 288U), std::make_tuple(32U, 64U), 64U);
+ graph << get_inception_node(data_path, "inception_4e", 256, std::make_tuple(160U, 320U), std::make_tuple(32U, 128U), 128U);
+ graph << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)));
+ graph << get_inception_node(data_path, "inception_5a", 256, std::make_tuple(160U, 320U), std::make_tuple(32U, 128U), 128U);
+ graph << get_inception_node(data_path, "inception_5b", 384, std::make_tuple(192U, 384U), std::make_tuple(48U, 128U), 128U);
+ graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 7, PadStrideInfo(1, 1, 0, 0, DimensionRoundingType::CEIL)))
<< FullyConnectedLayer(
1000U,
get_weights_accessor(data_path, "/cnn_data/googlenet_model/loss3/loss3_classifier_w.npy"),
get_weights_accessor(data_path, "/cnn_data/googlenet_model/loss3/loss3_classifier_b.npy"))
<< SoftmaxLayer()
- << Tensor(get_output_accessor(label, 5));
+ << OutputLayer(get_output_accessor(label, 5));
- // In order to enable the OpenCL tuner, graph_init() has to be called only when all nodes have been instantiated
- graph.graph_init(int_target_hint == 2);
+ // Finalize graph
+ GraphConfig config;
+ config.use_tuner = (target == 2);
+ graph.finalize(target_hint, config);
}
void do_run() override
{
@@ -145,7 +154,7 @@
}
private:
- Graph graph{};
+ Stream graph{ 0, "GoogleNet" };
BranchLayer get_inception_node(const std::string &data_path, std::string &¶m_path,
unsigned int a_filt,
@@ -154,7 +163,7 @@
unsigned int d_filt)
{
std::string total_path = "/cnn_data/googlenet_model/" + param_path + "/" + param_path + "_";
- SubGraph i_a;
+ SubStream i_a(graph);
i_a << ConvolutionLayer(
1U, 1U, a_filt,
get_weights_accessor(data_path, total_path + "1x1_w.npy"),
@@ -162,7 +171,7 @@
PadStrideInfo(1, 1, 0, 0))
<< ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
- SubGraph i_b;
+ SubStream i_b(graph);
i_b << ConvolutionLayer(
1U, 1U, std::get<0>(b_filters),
get_weights_accessor(data_path, total_path + "3x3_reduce_w.npy"),
@@ -176,7 +185,7 @@
PadStrideInfo(1, 1, 1, 1))
<< ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
- SubGraph i_c;
+ SubStream i_c(graph);
i_c << ConvolutionLayer(
1U, 1U, std::get<0>(c_filters),
get_weights_accessor(data_path, total_path + "5x5_reduce_w.npy"),
@@ -190,7 +199,7 @@
PadStrideInfo(1, 1, 2, 2))
<< ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
- SubGraph i_d;
+ SubStream i_d(graph);
i_d << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL)))
<< ConvolutionLayer(
1U, 1U, d_filt,
@@ -206,7 +215,7 @@
/** Main program for Googlenet
*
* @param[in] argc Number of arguments
- * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL), [optional] Path to the weights folder, [optional] image, [optional] labels )
+ * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) )
*/
int main(int argc, char **argv)
{
diff --git a/examples/graph_inception_v3.cpp b/examples/graph_inception_v3.cpp
index a55b34e..d1d6ab4 100644
--- a/examples/graph_inception_v3.cpp
+++ b/examples/graph_inception_v3.cpp
@@ -21,9 +21,7 @@
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
-#include "arm_compute/graph/Graph.h"
-#include "arm_compute/graph/Nodes.h"
-#include "arm_compute/graph/SubGraph.h"
+#include "arm_compute/graph.h"
#include "support/ToolchainSupport.h"
#include "utils/GraphUtils.h"
#include "utils/Utils.h"
@@ -32,15 +30,15 @@
#include <tuple>
using namespace arm_compute::utils;
-using namespace arm_compute::graph;
+using namespace arm_compute::graph::frontend;
using namespace arm_compute::graph_utils;
/** Example demonstrating how to implement InceptionV3's network using the Compute Library's graph API
*
* @param[in] argc Number of arguments
- * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels )
+ * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) )
*/
-class InceptionV3Example final : public Example
+class InceptionV3Example : public Example
{
public:
void do_setup(int argc, char **argv) override
@@ -53,140 +51,179 @@
std::unique_ptr<IPreprocessor> preprocessor = arm_compute::support::cpp14::make_unique<TFPreproccessor>();
// Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON
- const int int_target_hint = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
- TargetHint target_hint = set_target_hint(int_target_hint);
+ const int target = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
+ Target target_hint = set_target_hint(target);
+ FastMathHint fast_math_hint = FastMathHint::DISABLED;
// Parse arguments
if(argc < 2)
{
// Print help
- std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels] [fast_math_hint]\n\n";
std::cout << "No data folder provided: using random values\n\n";
}
else if(argc == 2)
{
- std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels] [fast_math_hint]\n\n";
std::cout << "No data folder provided: using random values\n\n";
}
else if(argc == 3)
{
data_path = argv[2];
- std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels] [fast_math_hint]\n\n";
std::cout << "No image provided: using random values\n\n";
}
else if(argc == 4)
{
data_path = argv[2];
image = argv[3];
- std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels] [fast_math_hint]\n\n";
std::cout << "No text file with labels provided: skipping output accessor\n\n";
}
- else
+ else if(argc == 5)
{
data_path = argv[2];
image = argv[3];
label = argv[4];
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " " << argv[4] << " [fast_math_hint]\n\n";
+ std::cout << "No fast math info provided: disabling fast math\n\n";
+ }
+ else
+ {
+ data_path = argv[2];
+ image = argv[3];
+ label = argv[4];
+ fast_math_hint = (std::strtol(argv[5], nullptr, 1) == 0) ? FastMathHint::DISABLED : FastMathHint::ENABLED;
}
- graph << target_hint << Tensor(TensorInfo(TensorShape(299U, 299U, 3U, 1U), 1, DataType::F32),
- get_input_accessor(image, std::move(preprocessor), false))
-
+ graph << target_hint
+ << fast_math_hint
+ << InputLayer(TensorDescriptor(TensorShape(299U, 299U, 3U, 1U), DataType::F32),
+ get_input_accessor(image, std::move(preprocessor), false))
<< ConvolutionLayer(3U, 3U, 32U,
get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_1a_3x3_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0))
+ .set_name("Conv2d_1a_3x3/convolution")
<< BatchNormalizationLayer(get_weights_accessor(data_path,
"/cnn_data/inceptionv3_model/Conv2d_1a_3x3_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path,
"/cnn_data/inceptionv3_model/Conv2d_1a_3x3_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f), get_weights_accessor(data_path,
"/cnn_data/inceptionv3_model/Conv2d_1a_3x3_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
-
+ 0.001f)
+ .set_name("Conv2d_1a_3x3/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_1a_3x3/Relu")
<< ConvolutionLayer(3U, 3U, 32U,
get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_2a_3x3_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
+ .set_name("Conv2d_2a_3x3/convolution")
<< BatchNormalizationLayer(get_weights_accessor(data_path,
"/cnn_data/inceptionv3_model/Conv2d_2a_3x3_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path,
"/cnn_data/inceptionv3_model/Conv2d_2a_3x3_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f), get_weights_accessor(data_path,
"/cnn_data/inceptionv3_model/Conv2d_2a_3x3_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ 0.001f)
+ .set_name("Conv2d_2a_3x3/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_2a_3x3/Relu")
<< ConvolutionLayer(3U, 3U, 64U,
get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_2b_3x3_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1))
+ .set_name("Conv2d_2b_3x3/convolution")
<< BatchNormalizationLayer(get_weights_accessor(data_path,
"/cnn_data/inceptionv3_model/Conv2d_2b_3x3_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path,
"/cnn_data/inceptionv3_model/Conv2d_2b_3x3_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f), get_weights_accessor(data_path,
"/cnn_data/inceptionv3_model/Conv2d_2b_3x3_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ 0.001f)
+ .set_name("Conv2d_2b_3x3/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_2b_3x3/Relu")
- << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
+ << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))).set_name("MaxPool_3a_3x3/MaxPool")
<< ConvolutionLayer(1U, 1U, 80U,
get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_3b_1x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
+ .set_name("Conv2d_3b_1x1/convolution")
<< BatchNormalizationLayer(get_weights_accessor(data_path,
"/cnn_data/inceptionv3_model/Conv2d_3b_1x1_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path,
"/cnn_data/inceptionv3_model/Conv2d_3b_1x1_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f), get_weights_accessor(data_path,
"/cnn_data/inceptionv3_model/Conv2d_3b_1x1_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ 0.001f)
+ .set_name("Conv2d_3b_1x1/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_3b_1x1/Relu")
<< ConvolutionLayer(3U, 3U, 192U,
get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_4a_3x3_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
+ .set_name("Conv2d_4a_3x3/convolution")
<< BatchNormalizationLayer(get_weights_accessor(data_path,
"/cnn_data/inceptionv3_model/Conv2d_4a_3x3_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path,
"/cnn_data/inceptionv3_model/Conv2d_4a_3x3_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f), get_weights_accessor(data_path,
"/cnn_data/inceptionv3_model/Conv2d_4a_3x3_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ 0.001f)
+ .set_name("Conv2d_4a_3x3/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_4a_3x3/Relu")
- << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
+ << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))).set_name("MaxPool_5a_3x3/MaxPool");
- << get_inception_node_A(data_path, "Mixed_5b", 64U, std::make_tuple(48U, 64U), std::make_tuple(64U, 96U, 96U),
+ graph << get_inception_node_A(data_path, "Mixed_5b", 64U, std::make_tuple(48U, 64U), std::make_tuple(64U, 96U, 96U),
32U)
- << get_inception_node_A(data_path, "Mixed_5c", 64U, std::make_tuple(48U, 64U), std::make_tuple(64U, 96U, 96U),
+ .set_name("Mixed_5b/concat");
+ graph << get_inception_node_A(data_path, "Mixed_5c", 64U, std::make_tuple(48U, 64U), std::make_tuple(64U, 96U, 96U),
64U, true)
- << get_inception_node_A(data_path, "Mixed_5d", 64U, std::make_tuple(48U, 64U), std::make_tuple(64U, 96U, 96U),
+ .set_name("Mixed_5c/concat");
+ graph << get_inception_node_A(data_path, "Mixed_5d", 64U, std::make_tuple(48U, 64U), std::make_tuple(64U, 96U, 96U),
64U)
+ .set_name("Mixed_5d/concat");
- << get_inception_node_B(data_path, "Mixed_6a", 384U, std::make_tuple(64U, 96U, 96U))
+ graph << get_inception_node_B(data_path, "Mixed_6a", 384U, std::make_tuple(64U, 96U, 96U)).set_name("Mixed_6a/concat");
- << get_inception_node_C(data_path, "Mixed_6b", 192U, std::make_tuple(128U, 128U, 192U),
+ graph << get_inception_node_C(data_path, "Mixed_6b", 192U, std::make_tuple(128U, 128U, 192U),
std::make_tuple(128U, 128U, 128U, 128U, 192U), 192U)
- << get_inception_node_C(data_path, "Mixed_6c", 192U, std::make_tuple(160U, 160U, 192U),
+ .set_name("Mixed_6b/concat");
+ graph << get_inception_node_C(data_path, "Mixed_6c", 192U, std::make_tuple(160U, 160U, 192U),
std::make_tuple(160U, 160U, 160U, 160U, 192U), 192U)
- << get_inception_node_C(data_path, "Mixed_6d", 192U, std::make_tuple(160U, 160U, 192U),
+ .set_name("Mixed_6c/concat");
+ graph << get_inception_node_C(data_path, "Mixed_6d", 192U, std::make_tuple(160U, 160U, 192U),
std::make_tuple(160U, 160U, 160U, 160U, 192U), 192U)
- << get_inception_node_C(data_path, "Mixed_6e", 192U, std::make_tuple(192U, 192U, 192U),
+ .set_name("Mixed_6d/concat");
+ graph << get_inception_node_C(data_path, "Mixed_6e", 192U, std::make_tuple(192U, 192U, 192U),
std::make_tuple(192U, 192U, 192U, 192U, 192U), 192U)
+ .set_name("Mixed_6e/concat");
- << get_inception_node_D(data_path, "Mixed_7a", std::make_tuple(192U, 320U),
+ graph << get_inception_node_D(data_path, "Mixed_7a", std::make_tuple(192U, 320U),
std::make_tuple(192U, 192U, 192U, 192U))
+ .set_name("Mixed_7a/concat");
- << get_inception_node_E(data_path, "Mixed_7b", 320U, std::make_tuple(384U, 384U, 384U),
+ graph << get_inception_node_E(data_path, "Mixed_7b", 320U, std::make_tuple(384U, 384U, 384U),
std::make_tuple(448U, 384U, 384U, 384U), 192U)
- << get_inception_node_E(data_path, "Mixed_7c", 320U, std::make_tuple(384U, 384U, 384U),
+ .set_name("Mixed_7b/concat");
+ graph << get_inception_node_E(data_path, "Mixed_7c", 320U, std::make_tuple(384U, 384U, 384U),
std::make_tuple(448U, 384U, 384U, 384U), 192U, true)
+ .set_name("Mixed_7c/concat");
- << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 8, PadStrideInfo(1, 1, 0, 0, DimensionRoundingType::CEIL)))
+ graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 8, PadStrideInfo(1, 1, 0, 0, DimensionRoundingType::CEIL))).set_name("Logits/AvgPool_1a_8x8/AvgPool")
<< ConvolutionLayer(1U, 1U, 1001U, get_weights_accessor(data_path,
"/cnn_data/inceptionv3_model/Logits_Conv2d_1c_1x1_weights.npy"),
get_weights_accessor(data_path,
"/cnn_data/inceptionv3_model/Logits_Conv2d_1c_1x1_biases.npy"),
PadStrideInfo(1, 1, 0, 0))
- << ReshapeLayer(TensorShape(1001U)) << SoftmaxLayer()
- << Tensor(get_output_accessor(label, 5));
+ .set_name("Logits/Conv2d_1c_1x1/convolution")
+ << ReshapeLayer(TensorShape(1001U)).set_name("Predictions/Reshape")
+ << SoftmaxLayer().set_name("Predictions/Softmax")
+ << OutputLayer(get_output_accessor(label, 5));
- // In order to enable the OpenCL tuner, graph_init() has to be called only when all nodes have been instantiated
- graph.graph_init(int_target_hint == 2);
+ // Finalize graph
+ GraphConfig config;
+ config.use_tuner = (target == 2);
+ graph.finalize(target_hint, config);
}
void do_run() override
@@ -195,7 +232,7 @@
}
private:
- Graph graph{};
+ Stream graph{ 0, "InceptionV3" };
private:
BranchLayer get_inception_node_A(const std::string &data_path, std::string &¶m_path,
@@ -216,91 +253,112 @@
conv_id1 = "_1_0c_";
}
- SubGraph i_a;
+ SubStream i_a(graph);
i_a << ConvolutionLayer(
1U, 1U, a_filt,
get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(1, 1, 0, 0))
+ .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/convolution")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f),
get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+ 0.001f)
+ .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_0/Conv2d_0a_1x1/Relu");
- SubGraph i_b;
+ SubStream i_b(graph);
i_b << ConvolutionLayer(
1U, 1U, std::get<0>(b_filters),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id0 + "1x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(1, 1, 0, 0))
+ .set_name(param_path + "/Branch_1/Conv2d" + conv_id0 + "1x1/convolution")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id0 + "1x1_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id0 + "1x1_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id0 + "1x1_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ 0.001f)
+ .set_name(param_path + "/Branch_1/Conv2d" + conv_id0 + "1x1/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d" + conv_id0 + "1x1/Relu")
<< ConvolutionLayer(
5U, 5U, std::get<1>(b_filters),
get_weights_accessor(data_path, total_path + "Branch_1_Conv" + conv_id1 + "5x5_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(1, 1, 2, 2))
+ .set_name(param_path + "/Branch_1/Conv2d" + conv_id1 + "5x5/convolution")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, total_path + "Branch_1_Conv" + conv_id1 + "5x5_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, total_path + "Branch_1_Conv" + conv_id1 + "5x5_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f),
get_weights_accessor(data_path, total_path + "Branch_1_Conv" + conv_id1 + "5x5_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+ 0.001f)
+ .set_name(param_path + "/Branch_1/Conv2d" + conv_id1 + "5x5/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d" + conv_id1 + "5x5/Relu");
- SubGraph i_c;
+ SubStream i_c(graph);
i_c << ConvolutionLayer(
1U, 1U, std::get<0>(c_filters),
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(1, 1, 0, 0))
+ .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/convolution")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f),
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ 0.001f)
+ .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0a_1x1/Relu")
<< ConvolutionLayer(
3U, 3U, std::get<1>(c_filters),
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(1, 1, 1, 1))
+ .set_name(param_path + "/Branch_2/Conv2d_0b_3x3/convolution")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f),
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ 0.001f)
+ .set_name(param_path + "/Branch_2/Conv2d_0b_3x3/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0b_3x3/Relu")
<< ConvolutionLayer(
3U, 3U, std::get<2>(c_filters),
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(1, 1, 1, 1))
+ .set_name(param_path + "/Branch_2/Conv2d_0c_3x3/convolution")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f),
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+ 0.001f)
+ .set_name(param_path + "/Branch_2/Conv2d_0c_3x3/BatchNorm/batcnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0c_3x3/Relu");
- SubGraph i_d;
- i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true))
+ SubStream i_d(graph);
+ i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true)).set_name(param_path + "/Branch_3/AvgPool_0a_3x3/AvgPool")
<< ConvolutionLayer(
1U, 1U, d_filt,
get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(1, 1, 0, 0))
+ .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/convolution")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f),
get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+ 0.001f)
+ .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_3/Conv2d_0b_1x1/Relu");
return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d));
}
@@ -310,57 +368,68 @@
std::tuple<unsigned int, unsigned int, unsigned int> b_filters)
{
std::string total_path = "/cnn_data/inceptionv3_model/" + param_path + "_";
- SubGraph i_a;
+ SubStream i_a(graph);
i_a << ConvolutionLayer(
3U, 3U, a_filt,
get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_1x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(2, 2, 0, 0))
+ .set_name(param_path + "/Branch_0/Conv2d_1a_1x1/convolution")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_1x1_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_1x1_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f),
get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_1x1_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+ 0.001f)
+ .set_name(param_path + "/Branch_0/Conv2d_1a_1x1/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_0/Conv2d_1a_1x1/Relu");
- SubGraph i_b;
+ SubStream i_b(graph);
i_b << ConvolutionLayer(
1U, 1U, std::get<0>(b_filters),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(1, 1, 0, 0))
+ .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/convolution")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ 0.001f)
+ .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0a_1x1/Relu")
<< ConvolutionLayer(
3U, 3U, std::get<1>(b_filters),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(1, 1, 1, 1))
+ .set_name(param_path + "/Branch_1/Conv2d_0b_3x3/convolution")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ 0.001f)
+ .set_name(param_path + "/Branch_1/Conv2d_0b_3x3/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0b_3x3/Relu")
<< ConvolutionLayer(
3U, 3U, std::get<2>(b_filters),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_1x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(2, 2, 0, 0))
+ .set_name(param_path + "/Branch_1/Conv2d_1a_1x1/convolution")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_1x1_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_1x1_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_1x1_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+ 0.001f)
+ .set_name(param_path + "/Branch_1/Conv2d_1a_1x1/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_1a_1x1/Relu");
- SubGraph i_c;
- i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 1.f, 0.f));
+ SubStream i_c(graph);
+ i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))).set_name(param_path + "/Branch_2/MaxPool_1a_3x3/MaxPool");
return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c));
}
@@ -372,124 +441,154 @@
unsigned int d_filt)
{
std::string total_path = "/cnn_data/inceptionv3_model/" + param_path + "_";
- SubGraph i_a;
+ SubStream i_a(graph);
i_a << ConvolutionLayer(
1U, 1U, a_filt,
get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(1, 1, 0, 0))
+ .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/convolution")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f),
get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+ 0.001f)
+ .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_0/Conv2d_0a_1x1/Relu");
- SubGraph i_b;
+ SubStream i_b(graph);
i_b << ConvolutionLayer(
1U, 1U, std::get<0>(b_filters),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(1, 1, 0, 0))
+ .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/convolution")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ 0.001f)
+ .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0a_1x1/Relu")
<< ConvolutionLayer(
7U, 1U, std::get<1>(b_filters),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(1, 1, 3, 0))
+ .set_name(param_path + "/Branch_1/Conv2d_0b_1x7/convolution")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ 0.001f)
+ .set_name(param_path + "/Branch_1/Conv2d_0b_1x7/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0b_1x7/Relu")
<< ConvolutionLayer(
1U, 7U, std::get<2>(b_filters),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(1, 1, 0, 3))
+ .set_name(param_path + "/Branch_1/Conv2d_0c_7x1/convolution")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+ 0.001f)
+ .set_name(param_path + "/Branch_1/Conv2d_0c_7x1/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_0/Conv2d_0c_7x1/Relu");
- SubGraph i_c;
+ SubStream i_c(graph);
i_c << ConvolutionLayer(
1U, 1U, std::get<0>(c_filters),
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(1, 1, 0, 0))
+ .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/convolution")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f),
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ 0.001f)
+ .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0a_1x1/Relu")
<< ConvolutionLayer(
1U, 7U, std::get<1>(c_filters),
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(1, 1, 0, 3))
+ .set_name(param_path + "/Branch_2/Conv2d_0b_7x1/convolution")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f),
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ 0.001f)
+ .set_name(param_path + "/Branch_2/Conv2d_0b_7x1/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0b_7x1/Relu")
<< ConvolutionLayer(
7U, 1U, std::get<2>(c_filters),
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(1, 1, 3, 0))
+ .set_name(param_path + "/Branch_2/Conv2d_0c_1x7/convolution")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f),
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ 0.001f)
+ .set_name(param_path + "/Branch_2/Conv2d_0c_1x7/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0c_1x7/Relu")
<< ConvolutionLayer(
1U, 7U, std::get<3>(c_filters),
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(1, 1, 0, 3))
+ .set_name(param_path + "/Branch_2/Conv2d_0d_7x1/convolution")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f),
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ 0.001f)
+ .set_name(param_path + "/Branch_2/Conv2d_0d_7x1/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0d_7x1/Relu")
<< ConvolutionLayer(
7U, 1U, std::get<4>(c_filters),
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(1, 1, 3, 0))
+ .set_name(param_path + "/Branch_2/Conv2d_0e_1x7/convolution")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f),
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+ 0.001f)
+ .set_name(param_path + "/Branch_2/Conv2d_0e_1x7/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0e_1x7/Relu");
- SubGraph i_d;
- i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true))
+ SubStream i_d(graph);
+ i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true)).set_name(param_path + "/Branch_3/AvgPool_0a_3x3/AvgPool")
<< ConvolutionLayer(
1U, 1U, d_filt,
get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(1, 1, 0, 0))
+ .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/convolution")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f),
get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+ 0.001f)
+ .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_3/Conv2d_0b_1x1/Relu");
return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d));
}
@@ -499,79 +598,96 @@
std::tuple<unsigned int, unsigned int, unsigned int, unsigned int> b_filters)
{
std::string total_path = "/cnn_data/inceptionv3_model/" + param_path + "_";
- SubGraph i_a;
+ SubStream i_a(graph);
i_a << ConvolutionLayer(
1U, 1U, std::get<0>(a_filters),
get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(1, 1, 0, 0))
+ .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/convolution")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f),
get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ 0.001f)
+ .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_0/Conv2d_0a_1x1/Relu")
<< ConvolutionLayer(
3U, 3U, std::get<1>(a_filters),
get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(2, 2, 0, 0))
+ .set_name(param_path + "/Branch_0/Conv2d_1a_3x3/convolution")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f),
get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+ 0.001f)
+ .set_name(param_path + "/Branch_0/Conv2d_1a_3x3/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_0/Conv2d_1a_3x3/Relu");
- SubGraph i_b;
+ SubStream i_b(graph);
i_b << ConvolutionLayer(
1U, 1U, std::get<0>(b_filters),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(1, 1, 0, 0))
+ .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/convolution")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ 0.001f)
+ .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0a_1x1/Relu")
<< ConvolutionLayer(
7U, 1U, std::get<1>(b_filters),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(1, 1, 3, 0))
+ .set_name(param_path + "/Branch_1/Conv2d_0b_1x7/convolution")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ 0.001f)
+ .set_name(param_path + "/Branch_1/Conv2d_0b_1x7/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0b_1x7/Relu")
<< ConvolutionLayer(
1U, 7U, std::get<2>(b_filters),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(1, 1, 0, 3))
+ .set_name(param_path + "/Branch_1/Conv2d_0c_7x1/convolution")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ 0.001f)
+ .set_name(param_path + "/Branch_1/Conv2d_0c_7x1/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0c_7x1/Relu")
<< ConvolutionLayer(
3U, 3U, std::get<3>(b_filters),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(2, 2, 0, 0))
+ .set_name(param_path + "/Branch_1/Conv2d_1a_3x3/convolution")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+ 0.001f)
+ .set_name(param_path + "/Branch_1/Conv2d_1a_3x3/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_1a_3x3/Relu");
- SubGraph i_c;
- i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 1.f, 0.f));
+ SubStream i_c(graph);
+ i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))).set_name(param_path + "/Branch_2/MaxPool_1a_3x3/MaxPool");
return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c));
}
@@ -591,123 +707,154 @@
}
std::string total_path = "/cnn_data/inceptionv3_model/" + param_path + "_";
- SubGraph i_a;
+ SubStream i_a(graph);
i_a << ConvolutionLayer(
1U, 1U, a_filt,
get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(1, 1, 0, 0))
+ .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/convolution")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f),
get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+ 0.001f)
+ .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_0/Conv2d_0a_1x1/Relu");
- SubGraph i_b1;
- i_b1 << ConvolutionLayer(
- 3U, 1U, std::get<1>(b_filters),
- get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_weights.npy"),
- std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
- PadStrideInfo(1, 1, 1, 0))
- << BatchNormalizationLayer(
- get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_moving_mean.npy"),
- get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_moving_variance.npy"),
- get_random_accessor(1.f, 1.f),
- get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
-
- SubGraph i_b2;
- i_b2 << ConvolutionLayer(
- 1U, 3U, std::get<2>(b_filters),
- get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id + "3x1_weights.npy"),
- std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
- PadStrideInfo(1, 1, 0, 1))
- << BatchNormalizationLayer(
- get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id + "3x1_BatchNorm_moving_mean.npy"),
- get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id + "3x1_BatchNorm_moving_variance.npy"),
- get_random_accessor(1.f, 1.f),
- get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id + "3x1_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
-
- SubGraph i_b;
+ SubStream i_b(graph);
i_b << ConvolutionLayer(
1U, 1U, std::get<0>(b_filters),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(1, 1, 0, 0))
+ .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/convolution")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f),
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
- << BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_b1), std::move(i_b2));
+ 0.001f)
+ .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0a_1x1/Relu");
- SubGraph i_c1;
- i_c1 << ConvolutionLayer(
- 3U, 1U, std::get<2>(c_filters),
- get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_weights.npy"),
+ SubStream i_b1(static_cast<IStream &>(i_b));
+ i_b1 << ConvolutionLayer(
+ 3U, 1U, std::get<1>(b_filters),
+ get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(1, 1, 1, 0))
+ .set_name(param_path + "/Branch_1/Conv2d_0b_1x3/convolution")
<< BatchNormalizationLayer(
- get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_BatchNorm_moving_mean.npy"),
- get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_BatchNorm_moving_variance.npy"),
+ get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_moving_mean.npy"),
+ get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f),
- get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+ get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_beta.npy"),
+ 0.001f)
+ .set_name(param_path + "/Branch_1/Conv2d_0b_1x3/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0b_1x3/Relu");
- SubGraph i_c2;
- i_c2 << ConvolutionLayer(
- 1U, 3U, std::get<3>(c_filters),
- get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_3x1_weights.npy"),
+ SubStream i_b2(static_cast<IStream &>(i_b));
+ i_b2 << ConvolutionLayer(
+ 1U, 3U, std::get<2>(b_filters),
+ get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id + "3x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(1, 1, 0, 1))
+ .set_name(param_path + "/Branch_1/Conv2d" + conv_id + "3x1/convolution")
<< BatchNormalizationLayer(
- get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_3x1_BatchNorm_moving_mean.npy"),
- get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_3x1_BatchNorm_moving_variance.npy"),
+ get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id + "3x1_BatchNorm_moving_mean.npy"),
+ get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id + "3x1_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f),
- get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_3x1_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+ get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id + "3x1_BatchNorm_beta.npy"),
+ 0.001f)
+ .set_name(param_path + "/Branch_1/Conv2d" + conv_id + "3x1/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d" + conv_id + "3x1/Relu");
- SubGraph i_c;
+ // Merge b1 and b2
+ i_b << BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_b1), std::move(i_b2)).set_name(param_path + "/Branch_1/concat");
+
+ SubStream i_c(graph);
i_c << ConvolutionLayer(
1U, 1U, std::get<0>(c_filters),
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(1, 1, 0, 0))
+ .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/convolution")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f),
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ 0.001f)
+ .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0a_1x1/Relu")
<< ConvolutionLayer(
3U, 3U, std::get<1>(c_filters),
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(1, 1, 1, 1))
+ .set_name(param_path + "/Branch_2/Conv2d_0b_3x3/convolution")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f),
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
- << BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_c1), std::move(i_c2));
+ 0.001f)
+ .set_name(param_path + "/Branch_2/Conv2d_0b_3x3/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0b_3x3/Relu");
- SubGraph i_d;
- i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true))
+ SubStream i_c1(static_cast<IStream &>(i_c));
+ i_c1 << ConvolutionLayer(
+ 3U, 1U, std::get<2>(c_filters),
+ get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_weights.npy"),
+ std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
+ PadStrideInfo(1, 1, 1, 0))
+ .set_name(param_path + "/Branch_2/Conv2d_0c_1x3/convolution")
+ << BatchNormalizationLayer(
+ get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_BatchNorm_moving_mean.npy"),
+ get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_BatchNorm_moving_variance.npy"),
+ get_random_accessor(1.f, 1.f),
+ get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_BatchNorm_beta.npy"),
+ 0.001f)
+ .set_name(param_path + "/Branch_2/Conv2d_0c_1x3/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0c_1x3/Relu");
+
+ SubStream i_c2(static_cast<IStream &>(i_c));
+ i_c2 << ConvolutionLayer(
+ 1U, 3U, std::get<3>(c_filters),
+ get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_3x1_weights.npy"),
+ std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
+ PadStrideInfo(1, 1, 0, 1))
+ .set_name(param_path + "/Branch_2/Conv2d_0d_3x1/convolution")
+ << BatchNormalizationLayer(
+ get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_3x1_BatchNorm_moving_mean.npy"),
+ get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_3x1_BatchNorm_moving_variance.npy"),
+ get_random_accessor(1.f, 1.f),
+ get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_3x1_BatchNorm_beta.npy"),
+ 0.001f)
+ .set_name(param_path + "/Branch_2/Conv2d_0d_3x1/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0d_3x1/Relu");
+
+ // Merge i_c1 and i_c2
+ i_c << BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_c1), std::move(i_c2)).set_name(param_path + "/Branch_2/concat");
+
+ SubStream i_d(graph);
+ i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true)).set_name(param_path + "/Branch_3/AvgPool_0a_3x3/AvgPool")
<< ConvolutionLayer(
1U, 1U, d_filt,
get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(1, 1, 0, 0))
+ .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/convolution")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_variance.npy"),
get_random_accessor(1.f, 1.f),
get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+ 0.001f)
+ .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/BatchNorm/batchnorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_3/Conv2d_0b_1x1/Relu");
return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d));
}
@@ -716,7 +863,7 @@
/** Main program for Inception V3
*
* @param[in] argc Number of arguments
- * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels )
+ * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) )
*/
int main(int argc, char **argv)
{
diff --git a/examples/graph_inception_v4.cpp b/examples/graph_inception_v4.cpp
index f004b41..ed95baa 100644
--- a/examples/graph_inception_v4.cpp
+++ b/examples/graph_inception_v4.cpp
@@ -21,9 +21,7 @@
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
-#include "arm_compute/graph/Graph.h"
-#include "arm_compute/graph/Nodes.h"
-#include "arm_compute/graph/SubGraph.h"
+#include "arm_compute/graph.h"
#include "support/ToolchainSupport.h"
#include "utils/GraphUtils.h"
#include "utils/Utils.h"
@@ -32,19 +30,22 @@
#include <tuple>
using namespace arm_compute::utils;
-using namespace arm_compute::graph;
+using namespace arm_compute::graph::frontend;
using namespace arm_compute::graph_utils;
/** Example demonstrating how to implement InceptionV4's network using the Compute Library's graph API
*
* @param[in] argc Number of arguments
- * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels )
+ * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) )
*/
class InceptionV4Example final : public Example
{
public:
void do_setup(int argc, char **argv) override
{
+ // Disabled the test for now because the process gets killed on Linux Firefly 32 bit even when using ConvolutionMethodHint::DIRECT.
+ // Needs to review/rework to run the code below.
+#if __aarch64__
std::string data_path; /* Path to the trainable data */
std::string image; /* Image data */
std::string label; /* Label data */
@@ -52,45 +53,56 @@
// Create a preprocessor object
std::unique_ptr<IPreprocessor> preprocessor = arm_compute::support::cpp14::make_unique<TFPreproccessor>();
- // Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON
- const int int_target_hint = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
- TargetHint target_hint = set_target_hint(int_target_hint);
+ // Set target. 0 (NEON), 1 (OpenCL). By default it is NEON
+ const int target = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
+ Target target_hint = set_target_hint(target);
+ FastMathHint fast_math_hint = FastMathHint::DISABLED;
// Parse arguments
if(argc < 2)
{
// Print help
- std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels] [fast_math_hint]\n\n";
std::cout << "No data folder provided: using random values\n\n";
}
else if(argc == 2)
{
- std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels] [fast_math_hint]\n\n";
std::cout << "No data folder provided: using random values\n\n";
}
else if(argc == 3)
{
data_path = argv[2];
- std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels] [fast_math_hint]\n\n";
std::cout << "No image provided: using random values\n\n";
}
else if(argc == 4)
{
data_path = argv[2];
image = argv[3];
- std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels] [fast_math_hint]\n\n";
std::cout << "No text file with labels provided: skipping output accessor\n\n";
}
- else
+ else if(argc == 5)
{
data_path = argv[2];
image = argv[3];
label = argv[4];
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " " << argv[4] << " [fast_math_hint]\n\n";
+ std::cout << "No fast math info provided: disabling fast math\n\n";
+ }
+ else
+ {
+ data_path = argv[2];
+ image = argv[3];
+ label = argv[4];
+ fast_math_hint = (std::strtol(argv[5], nullptr, 1) == 0) ? FastMathHint::DISABLED : FastMathHint::ENABLED;
}
- graph << target_hint << Tensor(TensorInfo(TensorShape(299U, 299U, 3U, 1U), 1, DataType::F32),
- get_input_accessor(image, std::move(preprocessor), false))
-
+ graph << target_hint
+ << fast_math_hint
+ << InputLayer(TensorDescriptor(TensorShape(299U, 299U, 3U, 1U), DataType::F32),
+ get_input_accessor(image, std::move(preprocessor), false))
// Conv2d_1a_3x3
<< ConvolutionLayer(3U, 3U, 32U,
get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_1a_3x3_weights.npy"),
@@ -120,64 +132,71 @@
get_random_accessor(1.f, 1.f),
get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2b_3x3_BatchNorm_beta.npy"),
0.001f)
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
- << get_mixed_3a(data_path)
- << get_mixed_4a(data_path)
- << get_mixed_5a(data_path)
- // 4 inception A blocks
- << get_inceptionA_block(data_path, "Mixed_5b")
- << get_inceptionA_block(data_path, "Mixed_5c")
- << get_inceptionA_block(data_path, "Mixed_5d")
- << get_inceptionA_block(data_path, "Mixed_5e")
- // reduction A block
- << get_reductionA_block(data_path)
- // 7 inception B blocks
- << get_inceptionB_block(data_path, "Mixed_6b")
- << get_inceptionB_block(data_path, "Mixed_6c")
- << get_inceptionB_block(data_path, "Mixed_6d")
- << get_inceptionB_block(data_path, "Mixed_6e")
- << get_inceptionB_block(data_path, "Mixed_6f")
- << get_inceptionB_block(data_path, "Mixed_6g")
- << get_inceptionB_block(data_path, "Mixed_6h")
- // reduction B block
- << get_reductionB_block(data_path)
- // 3 inception C blocks
- << get_inceptionC_block(data_path, "Mixed_7b")
- << get_inceptionC_block(data_path, "Mixed_7c")
- << get_inceptionC_block(data_path, "Mixed_7d")
- << PoolingLayer(PoolingLayerInfo(PoolingType::AVG))
+ graph << get_mixed_3a(data_path);
+ graph << get_mixed_4a(data_path);
+ graph << get_mixed_5a(data_path);
+ // 4 inception A blocks
+ graph << get_inceptionA_block(data_path, "Mixed_5b");
+ graph << get_inceptionA_block(data_path, "Mixed_5c");
+ graph << get_inceptionA_block(data_path, "Mixed_5d");
+ graph << get_inceptionA_block(data_path, "Mixed_5e");
+ // reduction A block
+ graph << get_reductionA_block(data_path);
+ // 7 inception B blocks
+ graph << get_inceptionB_block(data_path, "Mixed_6b");
+ graph << get_inceptionB_block(data_path, "Mixed_6c");
+ graph << get_inceptionB_block(data_path, "Mixed_6d");
+ graph << get_inceptionB_block(data_path, "Mixed_6e");
+ graph << get_inceptionB_block(data_path, "Mixed_6f");
+ graph << get_inceptionB_block(data_path, "Mixed_6g");
+ graph << get_inceptionB_block(data_path, "Mixed_6h");
+ // reduction B block
+ graph << get_reductionB_block(data_path);
+ // 3 inception C blocks
+ graph << get_inceptionC_block(data_path, "Mixed_7b");
+ graph << get_inceptionC_block(data_path, "Mixed_7c");
+ graph << get_inceptionC_block(data_path, "Mixed_7d");
+ graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG))
<< FlattenLayer()
<< FullyConnectedLayer(
1001U,
get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Logits_Logits_weights.npy"),
get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Logits_Logits_biases.npy"))
<< SoftmaxLayer()
- << Tensor(get_output_accessor(label, 5));
+ << OutputLayer(get_output_accessor(label, 5));
- // In order to enable the OpenCL tuner, graph_init() has to be called only when all nodes have been instantiated
- graph.graph_init(int_target_hint == 2);
+ // Finalize graph
+ GraphConfig config;
+ config.use_tuner = (target == 2);
+ graph.finalize(target_hint, config);
+#else /* __aarch64__ */
+ using namespace arm_compute;
+ ARM_COMPUTE_UNUSED(argc);
+ ARM_COMPUTE_UNUSED(argv);
+#endif /* __aarch64__ */
}
void do_run() override
{
+#if __aarch64__
graph.run();
+#endif /* __aarch64__ */
}
private:
- Graph graph{};
+ Stream graph{ 0, "InceptionV4" };
private:
BranchLayer get_mixed_3a(const std::string &data_path)
{
std::string total_path = "/cnn_data/inceptionv4_model/Mixed_3a_";
- SubGraph i_a;
- i_a << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true))
- // TODO (geopin01) : Remove once we understand why a single node graph does not run in CL
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 1.f, 0.f));
+ SubStream i_a(graph);
+ i_a << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true));
- SubGraph i_b;
+ SubStream i_b(graph);
i_b << ConvolutionLayer(3U, 3U, 96U,
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_3x3_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0))
@@ -195,7 +214,7 @@
{
std::string total_path = "/cnn_data/inceptionv4_model/Mixed_4a_";
- SubGraph i_a;
+ SubStream i_a(graph);
i_a << ConvolutionLayer(1U, 1U, 64U,
get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
@@ -215,7 +234,7 @@
0.001f)
<< ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
- SubGraph i_b;
+ SubStream i_b(graph);
i_b << ConvolutionLayer(1U, 1U, 64U,
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
@@ -260,7 +279,7 @@
{
std::string total_path = "/cnn_data/inceptionv4_model/Mixed_5a_";
- SubGraph i_a;
+ SubStream i_a(graph);
i_a << ConvolutionLayer(3U, 3U, 192U,
get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0))
@@ -271,10 +290,8 @@
0.001f)
<< ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
- SubGraph i_b;
- i_b << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true))
- // TODO (geopin01) : Remove once we understand why a single node graph does not run in CL
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 1.f, 0.f));
+ SubStream i_b(graph);
+ i_b << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true));
return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b));
}
@@ -283,7 +300,7 @@
{
std::string total_path = "/cnn_data/inceptionv4_model/" + param_path + "_";
- SubGraph i_a;
+ SubStream i_a(graph);
i_a << ConvolutionLayer(1U, 1U, 96U,
get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
@@ -294,7 +311,7 @@
0.001f)
<< ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
- SubGraph i_b;
+ SubStream i_b(graph);
i_b << ConvolutionLayer(1U, 1U, 64U,
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
@@ -314,7 +331,7 @@
0.001f)
<< ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
- SubGraph i_c;
+ SubStream i_c(graph);
i_c << ConvolutionLayer(1U, 1U, 64U,
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
@@ -343,7 +360,7 @@
0.001f)
<< ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
- SubGraph i_d;
+ SubStream i_d(graph);
i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true))
<< ConvolutionLayer(1U, 1U, 96U,
get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy"),
@@ -362,7 +379,7 @@
{
std::string total_path = "/cnn_data/inceptionv4_model/Mixed_6a_";
- SubGraph i_a;
+ SubStream i_a(graph);
i_a << ConvolutionLayer(3U, 3U, 384U,
get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0))
@@ -373,7 +390,7 @@
0.001f)
<< ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
- SubGraph i_b;
+ SubStream i_b(graph);
i_b << ConvolutionLayer(1U, 1U, 192U,
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
@@ -402,10 +419,9 @@
0.001f)
<< ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
- SubGraph i_c;
- i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true))
- // TODO (geopin01) : Remove once we understand why a single node graph does not run in CL
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 1.f, 0.f));
+ SubStream i_c(graph);
+ i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true));
+
return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c));
}
@@ -413,7 +429,7 @@
{
std::string total_path = "/cnn_data/inceptionv4_model/" + param_path + "_";
- SubGraph i_a;
+ SubStream i_a(graph);
i_a << ConvolutionLayer(1U, 1U, 384U,
get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
@@ -424,7 +440,7 @@
0.001f)
<< ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
- SubGraph i_b;
+ SubStream i_b(graph);
i_b << ConvolutionLayer(1U, 1U, 192U,
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
@@ -453,7 +469,7 @@
0.001f)
<< ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
- SubGraph i_c;
+ SubStream i_c(graph);
i_c << ConvolutionLayer(1U, 1U, 192U,
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
@@ -500,7 +516,7 @@
0.001f)
<< ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
- SubGraph i_d;
+ SubStream i_d(graph);
i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true))
<< ConvolutionLayer(1U, 1U, 128U,
get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy"),
@@ -519,7 +535,7 @@
{
std::string total_path = "/cnn_data/inceptionv4_model/Mixed_7a_";
- SubGraph i_a;
+ SubStream i_a(graph);
i_a << ConvolutionLayer(1U, 1U, 192U,
get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
@@ -539,7 +555,7 @@
0.001f)
<< ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
- SubGraph i_b;
+ SubStream i_b(graph);
i_b << ConvolutionLayer(1U, 1U, 256U,
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
@@ -577,10 +593,9 @@
0.001f)
<< ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
- SubGraph i_c;
- i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true))
- // TODO (geopin01) : Remove once we understand why a single node graph does not run in CL
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 1.f, 0.f));
+ SubStream i_c(graph);
+ i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true));
+
return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c));
}
@@ -588,7 +603,7 @@
{
std::string total_path = "/cnn_data/inceptionv4_model/" + param_path + "_";
- SubGraph i_a;
+ SubStream i_a(graph);
i_a << ConvolutionLayer(1U, 1U, 256U,
get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
@@ -599,7 +614,21 @@
0.001f)
<< ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
- SubGraph i_b1;
+ SubStream i_b(graph);
+ i_b << ConvolutionLayer(
+ 1U, 1U, 384U,
+ get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"),
+ std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
+ PadStrideInfo(1, 1, 0, 0))
+ << BatchNormalizationLayer(
+ get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
+ get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
+ get_random_accessor(1.f, 1.f),
+ get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"),
+ 0.001f)
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+
+ SubStream i_b1(static_cast<IStream &>(i_b));
i_b1 << ConvolutionLayer(
3U, 1U, 256U,
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_weights.npy"),
@@ -613,7 +642,7 @@
0.001f)
<< ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
- SubGraph i_b2;
+ SubStream i_b2(static_cast<IStream &>(i_b));
i_b2 << ConvolutionLayer(
1U, 3U, 256U,
get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_3x1_weights.npy"),
@@ -627,50 +656,10 @@
0.001f)
<< ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
- SubGraph i_b;
- i_b << ConvolutionLayer(
- 1U, 1U, 384U,
- get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"),
- std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
- PadStrideInfo(1, 1, 0, 0))
- << BatchNormalizationLayer(
- get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
- get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
- get_random_accessor(1.f, 1.f),
- get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"),
- 0.001f)
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
- << BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_b1), std::move(i_b2));
+ // Merge b1 and b2
+ i_b << BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_b1), std::move(i_b2));
- SubGraph i_c1;
- i_c1 << ConvolutionLayer(
- 3U, 1U, 256U,
- get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_1x3_weights.npy"),
- std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
- PadStrideInfo(1, 1, 1, 0))
- << BatchNormalizationLayer(
- get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_1x3_BatchNorm_moving_mean.npy"),
- get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_1x3_BatchNorm_moving_variance.npy"),
- get_random_accessor(1.f, 1.f),
- get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_1x3_BatchNorm_beta.npy"),
- 0.001f)
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
-
- SubGraph i_c2;
- i_c2 << ConvolutionLayer(
- 1U, 3U, 256U,
- get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_3x1_weights.npy"),
- std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
- PadStrideInfo(1, 1, 0, 1))
- << BatchNormalizationLayer(
- get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_3x1_BatchNorm_moving_mean.npy"),
- get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_3x1_BatchNorm_moving_variance.npy"),
- get_random_accessor(1.f, 1.f),
- get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_3x1_BatchNorm_beta.npy"),
- 0.001f)
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
-
- SubGraph i_c;
+ SubStream i_c(graph);
i_c << ConvolutionLayer(
1U, 1U, 384U,
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy"),
@@ -706,10 +695,40 @@
get_random_accessor(1.f, 1.f),
get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_BatchNorm_beta.npy"),
0.001f)
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
- << BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_c1), std::move(i_c2));
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
- SubGraph i_d;
+ SubStream i_c1(static_cast<IStream &>(i_c));
+ i_c1 << ConvolutionLayer(
+ 3U, 1U, 256U,
+ get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_1x3_weights.npy"),
+ std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
+ PadStrideInfo(1, 1, 1, 0))
+ << BatchNormalizationLayer(
+ get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_1x3_BatchNorm_moving_mean.npy"),
+ get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_1x3_BatchNorm_moving_variance.npy"),
+ get_random_accessor(1.f, 1.f),
+ get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_1x3_BatchNorm_beta.npy"),
+ 0.001f)
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+
+ SubStream i_c2(static_cast<IStream &>(i_c));
+ i_c2 << ConvolutionLayer(
+ 1U, 3U, 256U,
+ get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_3x1_weights.npy"),
+ std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
+ PadStrideInfo(1, 1, 0, 1))
+ << BatchNormalizationLayer(
+ get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_3x1_BatchNorm_moving_mean.npy"),
+ get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_3x1_BatchNorm_moving_variance.npy"),
+ get_random_accessor(1.f, 1.f),
+ get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_3x1_BatchNorm_beta.npy"),
+ 0.001f)
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+
+ // Merge i_c1 and i_c2
+ i_c << BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_c1), std::move(i_c2));
+
+ SubStream i_d(graph);
i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true))
<< ConvolutionLayer(1U, 1U, 256U,
get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy"),
@@ -728,7 +747,7 @@
/** Main program for Inception V4
*
* @param[in] argc Number of arguments
- * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels )
+ * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) )
*/
int main(int argc, char **argv)
{
diff --git a/examples/graph_lenet.cpp b/examples/graph_lenet.cpp
index 61bc7bd..32c7582 100644
--- a/examples/graph_lenet.cpp
+++ b/examples/graph_lenet.cpp
@@ -21,8 +21,8 @@
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
-#include "arm_compute/graph/Graph.h"
-#include "arm_compute/graph/Nodes.h"
+#include "arm_compute/graph.h"
+
#include "support/ToolchainSupport.h"
#include "utils/GraphUtils.h"
#include "utils/Utils.h"
@@ -30,13 +30,13 @@
#include <cstdlib>
using namespace arm_compute::utils;
-using namespace arm_compute::graph;
+using namespace arm_compute::graph::frontend;
using namespace arm_compute::graph_utils;
/** Example demonstrating how to implement LeNet's network using the Compute Library's graph API
*
* @param[in] argc Number of arguments
- * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL), [optional] Path to the weights folder, [optional] batches )
+ * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL), [optional] Path to the weights folder, [optional] batches, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) )
*/
class GraphLenetExample : public Example
{
@@ -47,64 +47,81 @@
unsigned int batches = 4; /** Number of batches */
// Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON
- const int int_target_hint = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
- TargetHint target_hint = set_target_hint(int_target_hint);
+ const int target = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
+ Target target_hint = set_target_hint(target);
+
+ FastMathHint fast_math_hint = FastMathHint::DISABLED;
// Parse arguments
if(argc < 2)
{
// Print help
- std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [batches]\n\n";
+ std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [batches] [fast_math_hint]\n\n";
std::cout << "No data folder provided: using random values\n\n";
}
else if(argc == 2)
{
- std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [batches]\n\n";
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [batches] [fast_math_hint]\n\n";
std::cout << "No data folder provided: using random values\n\n";
}
else if(argc == 3)
{
//Do something with argv[1]
data_path = argv[2];
- std::cout << "Usage: " << argv[0] << " [path_to_data] [batches]\n\n";
+ std::cout << "Usage: " << argv[0] << " [path_to_data] [batches] [fast_math_hint]\n\n";
std::cout << "No number of batches where specified, thus will use the default : " << batches << "\n\n";
}
+ else if(argc == 4)
+ {
+ data_path = argv[2];
+ batches = std::strtol(argv[3], nullptr, 0);
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [fast_math_hint]\n\n";
+ std::cout << "No fast math info provided: disabling fast math\n\n";
+ }
else
{
//Do something with argv[1] and argv[2]
- data_path = argv[2];
- batches = std::strtol(argv[3], nullptr, 0);
+ data_path = argv[2];
+ batches = std::strtol(argv[3], nullptr, 0);
+ fast_math_hint = (std::strtol(argv[4], nullptr, 1) == 0) ? FastMathHint::DISABLED : FastMathHint::ENABLED;
}
//conv1 << pool1 << conv2 << pool2 << fc1 << act1 << fc2 << smx
graph << target_hint
- << Tensor(TensorInfo(TensorShape(28U, 28U, 1U, batches), 1, DataType::F32), DummyAccessor())
+ << fast_math_hint
+ << InputLayer(TensorDescriptor(TensorShape(28U, 28U, 1U, batches), DataType::F32), get_input_accessor(""))
<< ConvolutionLayer(
5U, 5U, 20U,
get_weights_accessor(data_path, "/cnn_data/lenet_model/conv1_w.npy"),
get_weights_accessor(data_path, "/cnn_data/lenet_model/conv1_b.npy"),
PadStrideInfo(1, 1, 0, 0))
- << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0)))
+ .set_name("conv1")
+ << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0))).set_name("pool1")
<< ConvolutionLayer(
5U, 5U, 50U,
get_weights_accessor(data_path, "/cnn_data/lenet_model/conv2_w.npy"),
get_weights_accessor(data_path, "/cnn_data/lenet_model/conv2_b.npy"),
PadStrideInfo(1, 1, 0, 0))
- << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0)))
+ .set_name("conv2")
+ << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0))).set_name("pool2")
<< FullyConnectedLayer(
500U,
get_weights_accessor(data_path, "/cnn_data/lenet_model/ip1_w.npy"),
get_weights_accessor(data_path, "/cnn_data/lenet_model/ip1_b.npy"))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ .set_name("ip1")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu")
<< FullyConnectedLayer(
10U,
get_weights_accessor(data_path, "/cnn_data/lenet_model/ip2_w.npy"),
get_weights_accessor(data_path, "/cnn_data/lenet_model/ip2_b.npy"))
- << SoftmaxLayer()
- << Tensor(DummyAccessor(0));
+ .set_name("ip2")
+ << SoftmaxLayer().set_name("prob")
+ << OutputLayer(get_output_accessor(""));
- // In order to enable the OpenCL tuner, graph_init() has to be called only when all nodes have been instantiated
- graph.graph_init(int_target_hint == 2);
+ // Finalize graph
+ GraphConfig config;
+ config.use_tuner = (target == 2);
+ graph.finalize(target_hint, config);
}
void do_run() override
{
@@ -113,13 +130,13 @@
}
private:
- Graph graph{};
+ Stream graph{ 0, "LeNet" };
};
/** Main program for LeNet
*
* @param[in] argc Number of arguments
- * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL), [optional] Path to the weights folder, [optional] batches )
+ * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] batches, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) )
*/
int main(int argc, char **argv)
{
diff --git a/examples/graph_mobilenet.cpp b/examples/graph_mobilenet.cpp
index 1a930dd..50dc024 100644
--- a/examples/graph_mobilenet.cpp
+++ b/examples/graph_mobilenet.cpp
@@ -21,8 +21,7 @@
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
-#include "arm_compute/graph/Graph.h"
-#include "arm_compute/graph/Nodes.h"
+#include "arm_compute/graph.h"
#include "support/ToolchainSupport.h"
#include "utils/GraphUtils.h"
#include "utils/Utils.h"
@@ -30,13 +29,13 @@
#include <cstdlib>
using namespace arm_compute::utils;
-using namespace arm_compute::graph;
+using namespace arm_compute::graph::frontend;
using namespace arm_compute::graph_utils;
/** Example demonstrating how to implement MobileNet's network using the Compute Library's graph API
*
* @param[in] argc Number of arguments
- * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels )
+ * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels, [optional] data layout, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) )
*/
class GraphMobilenetExample : public Example
{
@@ -51,54 +50,80 @@
std::unique_ptr<IPreprocessor> preprocessor = arm_compute::support::cpp14::make_unique<TFPreproccessor>();
// Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON
- const int int_target_hint = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
- TargetHint target_hint = set_target_hint(int_target_hint);
- ConvolutionMethodHint convolution_hint = ConvolutionMethodHint::GEMM;
+ const int target = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
+ Target target_hint = set_target_hint(target);
+ ConvolutionMethod convolution_hint = ConvolutionMethod::GEMM;
+ DepthwiseConvolutionMethod depthwise_convolution_hint = DepthwiseConvolutionMethod::OPTIMIZED_3x3;
+ FastMathHint fast_math_hint = FastMathHint::DISABLED;
// Set model to execute. 0 (MobileNetV1_1.0_224), 1 (MobileNetV1_0.75_160)
int model_id = (argc > 2) ? std::strtol(argv[2], nullptr, 10) : 0;
ARM_COMPUTE_ERROR_ON_MSG(model_id > 1, "Invalid model ID. Model must be 0 (MobileNetV1_1.0_224) or 1 (MobileNetV1_0.75_160)");
- float depth_scale = (model_id == 0) ? 1.f : 0.75;
- unsigned int spatial_size = (model_id == 0) ? 224 : 160;
- std::string model_path = (model_id == 0) ? "/cnn_data/mobilenet_v1_1_224_model/" : "/cnn_data/mobilenet_v1_075_160_model/";
+ int layout_id = (argc > 3) ? std::strtol(argv[3], nullptr, 10) : 0;
+ ARM_COMPUTE_ERROR_ON_MSG(layout_id > 1, "Invalid layout ID. Layout must be 0 (NCHW) or 1 (NHWC)");
+
+ float depth_scale = (model_id == 0) ? 1.f : 0.75;
+ unsigned int spatial_size = (model_id == 0) ? 224 : 160;
+ std::string model_path = (model_id == 0) ? "/cnn_data/mobilenet_v1_1_224_model/" : "/cnn_data/mobilenet_v1_075_160_model/";
+ TensorDescriptor input_descriptor_nchw = TensorDescriptor(TensorShape(spatial_size, spatial_size, 3U, 1U), DataType::F32);
+ TensorDescriptor input_descriptor_nhwc = TensorDescriptor(TensorShape(3U, spatial_size, spatial_size, 1U), DataType::F32).set_layout(DataLayout::NHWC);
+ TensorDescriptor input_descriptor = (layout_id == 0) ? input_descriptor_nchw : input_descriptor_nhwc;
// Parse arguments
if(argc < 2)
{
// Print help
- std::cout << "Usage: " << argv[0] << " [target] [model] [path_to_data] [image] [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " [target] [model] [layout] [path_to_data] [image] [labels] [fast_math_hint]\n\n";
std::cout << "No model ID provided: using MobileNetV1_1.0_224\n\n";
+ std::cout << "No data layout provided: using NCHW\n\n";
std::cout << "No data folder provided: using random values\n\n";
}
else if(argc == 2)
{
- std::cout << "Usage: " << argv[0] << " " << argv[1] << " [model] [path_to_data] [image] [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " [model] [layout] [path_to_data] [image] [labels] [fast_math_hint]\n\n";
std::cout << "No model ID provided: using MobileNetV1_1.0_224\n\n";
+ std::cout << "No data layout provided: using NCHW\n\n";
std::cout << "No data folder provided: using random values\n\n";
}
else if(argc == 3)
{
- std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [path_to_data] [image] [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [layout] [path_to_data] [image] [labels] [fast_math_hint]\n\n";
+ std::cout << "No data layout provided: using NCHW\n\n";
std::cout << "No data folder provided: using random values\n\n";
}
else if(argc == 4)
{
- data_path = argv[3];
- std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [image] [labels]\n\n";
- std::cout << "No image provided: using random values\n\n";
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [path_to_data] [image] [labels] [fast_math_hint]\n\n";
+ std::cout << "No data folder provided: using random values\n\n";
}
else if(argc == 5)
{
- data_path = argv[3];
- image = argv[4];
- std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels]\n\n";
+ data_path = argv[4];
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " " << argv[4] << " [image] [labels] [fast_math_hint]\n\n";
+ std::cout << "No image provided: using random values\n\n";
std::cout << "No text file with labels provided: skipping output accessor\n\n";
}
+ else if(argc == 6)
+ {
+ data_path = argv[4];
+ image = argv[5];
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels] [fast_math_hint]\n\n";
+ std::cout << "No text file with labels provided: skipping output accessor\n\n";
+ }
+ else if(argc == 7)
+ {
+ data_path = argv[4];
+ image = argv[5];
+ label = argv[6];
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " " << argv[4] << " [fast_math_hint]\n\n";
+ std::cout << "No fast math info provided: disabling fast math\n\n";
+ }
else
{
- data_path = argv[3];
- image = argv[4];
- label = argv[5];
+ data_path = argv[4];
+ image = argv[5];
+ label = argv[6];
+ fast_math_hint = (std::strtol(argv[7], nullptr, 1) == 0) ? FastMathHint::DISABLED : FastMathHint::ENABLED;
}
// Add model path to data path
@@ -109,44 +134,52 @@
graph << target_hint
<< convolution_hint
- << Tensor(TensorInfo(TensorShape(spatial_size, spatial_size, 3U, 1U), 1, DataType::F32),
- get_input_accessor(image, std::move(preprocessor), false))
+ << depthwise_convolution_hint
+ << fast_math_hint
+ << InputLayer(input_descriptor,
+ get_input_accessor(image, std::move(preprocessor), false))
<< ConvolutionLayer(
3U, 3U, 32U * depth_scale,
- get_weights_accessor(data_path, "Conv2d_0_weights.npy"),
+ get_weights_accessor(data_path, "Conv2d_0_weights.npy", DataLayout::NCHW),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR))
+ .set_name("Conv2d_0")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, "Conv2d_0_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, "Conv2d_0_BatchNorm_moving_variance.npy"),
get_weights_accessor(data_path, "Conv2d_0_BatchNorm_gamma.npy"),
get_weights_accessor(data_path, "Conv2d_0_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f))
- << get_dwsc_node(data_path, "Conv2d_1", 64 * depth_scale, PadStrideInfo(1, 1, 1, 1), PadStrideInfo(1, 1, 0, 0))
- << get_dwsc_node(data_path, "Conv2d_2", 128 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0))
- << get_dwsc_node(data_path, "Conv2d_3", 128 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0))
- << get_dwsc_node(data_path, "Conv2d_4", 256 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0))
- << get_dwsc_node(data_path, "Conv2d_5", 256 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0))
- << get_dwsc_node(data_path, "Conv2d_6", 512 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0))
- << get_dwsc_node(data_path, "Conv2d_7", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0))
- << get_dwsc_node(data_path, "Conv2d_8", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0))
- << get_dwsc_node(data_path, "Conv2d_9", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0))
- << get_dwsc_node(data_path, "Conv2d_10", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0))
- << get_dwsc_node(data_path, "Conv2d_11", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0))
- << get_dwsc_node(data_path, "Conv2d_12", 1024 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0))
- << get_dwsc_node(data_path, "Conv2d_13", 1024 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0))
- << PoolingLayer(PoolingLayerInfo(PoolingType::AVG))
+ 0.001f)
+ .set_name("Conv2d_0/BatchNorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)).set_name("Conv2d_0/Relu6");
+ graph << get_dwsc_node(data_path, "Conv2d_1", 64 * depth_scale, PadStrideInfo(1, 1, 1, 1), PadStrideInfo(1, 1, 0, 0));
+ graph << get_dwsc_node(data_path, "Conv2d_2", 128 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
+ graph << get_dwsc_node(data_path, "Conv2d_3", 128 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
+ graph << get_dwsc_node(data_path, "Conv2d_4", 256 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
+ graph << get_dwsc_node(data_path, "Conv2d_5", 256 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
+ graph << get_dwsc_node(data_path, "Conv2d_6", 512 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
+ graph << get_dwsc_node(data_path, "Conv2d_7", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
+ graph << get_dwsc_node(data_path, "Conv2d_8", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
+ graph << get_dwsc_node(data_path, "Conv2d_9", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
+ graph << get_dwsc_node(data_path, "Conv2d_10", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
+ graph << get_dwsc_node(data_path, "Conv2d_11", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
+ graph << get_dwsc_node(data_path, "Conv2d_12", 1024 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
+ graph << get_dwsc_node(data_path, "Conv2d_13", 1024 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
+ graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG)).set_name("Logits/AvgPool_1a")
<< ConvolutionLayer(
1U, 1U, 1001U,
- get_weights_accessor(data_path, "Logits_Conv2d_1c_1x1_weights.npy"),
+ get_weights_accessor(data_path, "Logits_Conv2d_1c_1x1_weights.npy", DataLayout::NCHW),
get_weights_accessor(data_path, "Logits_Conv2d_1c_1x1_biases.npy"),
PadStrideInfo(1, 1, 0, 0))
- << ReshapeLayer(TensorShape(1001U))
- << SoftmaxLayer()
- << Tensor(get_output_accessor(label, 5));
+ .set_name("Logits/Conv2d_1c_1x1")
+ << ReshapeLayer(TensorShape(1001U)).set_name("Reshape")
+ << SoftmaxLayer().set_name("Softmax")
+ << OutputLayer(get_output_accessor(label, 5));
- // In order to enable the OpenCL tuner, graph_init() has to be called only when all nodes have been instantiated
- graph.graph_init(int_target_hint == 2);
+ // Finalize graph
+ GraphConfig config;
+ config.use_tuner = (target == 2);
+ graph.finalize(target_hint, config);
}
void do_run() override
{
@@ -155,37 +188,42 @@
}
private:
- Graph graph{};
+ Stream graph{ 0, "MobileNetV1" };
BranchLayer get_dwsc_node(const std::string &data_path, std::string &¶m_path,
unsigned int conv_filt,
PadStrideInfo dwc_pad_stride_info, PadStrideInfo conv_pad_stride_info)
{
std::string total_path = param_path + "_";
- SubGraph sg;
+ SubStream sg(graph);
sg << DepthwiseConvolutionLayer(
3U, 3U,
- get_weights_accessor(data_path, total_path + "depthwise_depthwise_weights.npy"),
+ get_weights_accessor(data_path, total_path + "depthwise_depthwise_weights.npy", DataLayout::NCHW),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
- dwc_pad_stride_info,
- true)
+ dwc_pad_stride_info)
+ .set_name(total_path + "depthwise/depthwise")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_moving_variance.npy"),
get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_gamma.npy"),
get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f))
+ 0.001f)
+ .set_name(total_path + "depthwise/BatchNorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)).set_name(total_path + "depthwise/Relu6")
<< ConvolutionLayer(
1U, 1U, conv_filt,
- get_weights_accessor(data_path, total_path + "pointwise_weights.npy"),
+ get_weights_accessor(data_path, total_path + "pointwise_weights.npy", DataLayout::NCHW),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
conv_pad_stride_info)
+ .set_name(total_path + "pointwise/Conv2D")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, total_path + "pointwise_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, total_path + "pointwise_BatchNorm_moving_variance.npy"),
get_weights_accessor(data_path, total_path + "pointwise_BatchNorm_gamma.npy"),
get_weights_accessor(data_path, total_path + "pointwise_BatchNorm_beta.npy"),
- 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f));
+ 0.001f)
+ .set_name(total_path + "pointwise/BatchNorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)).set_name(total_path + "pointwise/Relu6");
return BranchLayer(std::move(sg));
}
@@ -194,11 +232,13 @@
/** Main program for MobileNetV1
*
* @param[in] argc Number of arguments
- * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL),
+ * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner),
* [optional] Model ID (0 = MobileNetV1_1.0_224, 1 = MobileNetV1_0.75_160),
* [optional] Path to the weights folder,
* [optional] image,
- * [optional] labels )
+ * [optional] labels,
+ * [optional] data layout,
+ * [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) )
*/
int main(int argc, char **argv)
{
diff --git a/examples/graph_mobilenet_qasymm8.cpp b/examples/graph_mobilenet_qasymm8.cpp
index 76b13dd..6cf0e48 100644
--- a/examples/graph_mobilenet_qasymm8.cpp
+++ b/examples/graph_mobilenet_qasymm8.cpp
@@ -21,22 +21,24 @@
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
-#include "arm_compute/graph/Graph.h"
-#include "arm_compute/graph/Nodes.h"
+#include "arm_compute/graph.h"
#include "support/ToolchainSupport.h"
#include "utils/GraphUtils.h"
#include "utils/Utils.h"
+#include <cstdlib>
+
using namespace arm_compute;
-using namespace arm_compute::graph;
+using namespace arm_compute::utils;
+using namespace arm_compute::graph::frontend;
using namespace arm_compute::graph_utils;
/** Example demonstrating how to implement QASYMM8 MobileNet's network using the Compute Library's graph API
*
* @param[in] argc Number of arguments
- * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] npy_input, [optional] labels )
+ * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] npy_input, [optional] labels, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) )
*/
-class GraphMobileNetQASYMM8Example : public utils::Example
+class GraphMobileNetQASYMM8Example : public Example
{
public:
void do_setup(int argc, char **argv) override
@@ -90,84 +92,96 @@
};
// Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON
- const int int_target_hint = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
- TargetHint target_hint = set_target_hint(int_target_hint);
+ const int target = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
+ Target target_hint = set_target_hint(target);
+ FastMathHint fast_math_hint = FastMathHint::DISABLED;
// Parse arguments
if(argc < 2)
{
// Print help
- std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [npy_input] [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [npy_input] [labels] [fast_math_hint]\n\n";
std::cout << "No data folder provided: using random values\n\n";
}
else if(argc == 2)
{
- std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [npy_input] [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [npy_input] [labels] [fast_math_hint]\n\n";
std::cout << "No input provided: using random values\n\n";
}
else if(argc == 4)
{
data_path = argv[2];
input = argv[3];
- std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels] [fast_math_hint]\n\n";
std::cout << "No text file with labels provided: skipping output accessor\n\n";
}
- else
+ else if(argc == 5)
{
data_path = argv[2];
input = argv[3];
label = argv[4];
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " " << argv[4] << " [fast_math_hint]\n\n";
+ std::cout << "No fast math info provided: disabling fast math\n\n";
+ }
+ else
+ {
+ data_path = argv[2];
+ input = argv[3];
+ label = argv[4];
+ fast_math_hint = (std::strtol(argv[5], nullptr, 1) == 0) ? FastMathHint::DISABLED : FastMathHint::ENABLED;
}
graph << target_hint
- << arm_compute::graph::Tensor(TensorInfo(TensorShape(224U, 224U, 3U, 1U), 1, DataType::QASYMM8, in_quant_info),
- get_weights_accessor(data_path, "/cnn_data/mobilenet_qasymm8_model/" + input))
+ << DepthwiseConvolutionMethod::OPTIMIZED_3x3
+ << fast_math_hint
+ << InputLayer(TensorDescriptor(TensorShape(224U, 224U, 3U, 1U), DataType::QASYMM8, in_quant_info),
+ get_weights_accessor(data_path, "/cnn_data/mobilenet_qasymm8_model/" + input))
<< ConvolutionLayer(
3U, 3U, 32U,
get_weights_accessor(data_path, "/cnn_data/mobilenet_qasymm8_model/Conv2d_0_weights.npy"),
get_weights_accessor(data_path, "/cnn_data/mobilenet_qasymm8_model/Conv2d_0_bias.npy"),
PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR),
- 1, WeightsInfo(),
- conv_weights_quant_info.at(0),
- mid_quant_info)
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f))
- << get_dwsc_node(data_path, "Conv2d_1", 64U, PadStrideInfo(1U, 1U, 1U, 1U), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(0), point_weights_quant_info.at(0))
- << get_dwsc_node(data_path, "Conv2d_2", 128U, PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(1),
- point_weights_quant_info.at(1))
- << get_dwsc_node(data_path, "Conv2d_3", 128U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(2),
- point_weights_quant_info.at(2))
- << get_dwsc_node(data_path, "Conv2d_4", 256U, PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(3),
- point_weights_quant_info.at(3))
- << get_dwsc_node(data_path, "Conv2d_5", 256U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(4),
- point_weights_quant_info.at(4))
- << get_dwsc_node(data_path, "Conv2d_6", 512U, PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(5),
- point_weights_quant_info.at(5))
- << get_dwsc_node(data_path, "Conv2d_7", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(6),
- point_weights_quant_info.at(6))
- << get_dwsc_node(data_path, "Conv2d_8", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(7),
- point_weights_quant_info.at(7))
- << get_dwsc_node(data_path, "Conv2d_9", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(8),
- point_weights_quant_info.at(8))
- << get_dwsc_node(data_path, "Conv2d_10", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(9),
- point_weights_quant_info.at(9))
- << get_dwsc_node(data_path, "Conv2d_11", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(10),
- point_weights_quant_info.at(10))
- << get_dwsc_node(data_path, "Conv2d_12", 1024U, PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(11),
- point_weights_quant_info.at(11))
- << get_dwsc_node(data_path, "Conv2d_13", 1024U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(12),
+ 1, conv_weights_quant_info.at(0), mid_quant_info)
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f));
+ graph << get_dwsc_node(data_path, "Conv2d_1", 64U, PadStrideInfo(1U, 1U, 1U, 1U), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(0), point_weights_quant_info.at(0));
+ graph << get_dwsc_node(data_path, "Conv2d_2", 128U, PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(1),
+ point_weights_quant_info.at(1));
+ graph << get_dwsc_node(data_path, "Conv2d_3", 128U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(2),
+ point_weights_quant_info.at(2));
+ graph << get_dwsc_node(data_path, "Conv2d_4", 256U, PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(3),
+ point_weights_quant_info.at(3));
+ graph << get_dwsc_node(data_path, "Conv2d_5", 256U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(4),
+ point_weights_quant_info.at(4));
+ graph << get_dwsc_node(data_path, "Conv2d_6", 512U, PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(5),
+ point_weights_quant_info.at(5));
+ graph << get_dwsc_node(data_path, "Conv2d_7", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(6),
+ point_weights_quant_info.at(6));
+ graph << get_dwsc_node(data_path, "Conv2d_8", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(7),
+ point_weights_quant_info.at(7));
+ graph << get_dwsc_node(data_path, "Conv2d_9", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(8),
+ point_weights_quant_info.at(8));
+ graph << get_dwsc_node(data_path, "Conv2d_10", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(9),
+ point_weights_quant_info.at(9));
+ graph << get_dwsc_node(data_path, "Conv2d_11", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(10),
+ point_weights_quant_info.at(10));
+ graph << get_dwsc_node(data_path, "Conv2d_12", 1024U, PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(11),
+ point_weights_quant_info.at(11));
+ graph << get_dwsc_node(data_path, "Conv2d_13", 1024U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(12),
point_weights_quant_info.at(12))
<< PoolingLayer(PoolingLayerInfo(PoolingType::AVG))
<< ConvolutionLayer(
1U, 1U, 1001U,
get_weights_accessor(data_path, "/cnn_data/mobilenet_qasymm8_model/Logits_Conv2d_1c_1x1_weights.npy"),
get_weights_accessor(data_path, "/cnn_data/mobilenet_qasymm8_model/Logits_Conv2d_1c_1x1_bias.npy"),
- PadStrideInfo(1U, 1U, 0U, 0U), 1, WeightsInfo(), conv_weights_quant_info.at(1))
+ PadStrideInfo(1U, 1U, 0U, 0U), 1, conv_weights_quant_info.at(1))
<< ReshapeLayer(TensorShape(1001U))
<< SoftmaxLayer()
- << arm_compute::graph::Tensor(get_output_accessor(label, 5));
+ << OutputLayer(get_output_accessor(label, 5));
- // In order to enable the OpenCL tuner, graph_init() has to be called only when all nodes have been instantiated
- graph.graph_init(int_target_hint == 2);
+ // Finalize graph
+ GraphConfig config;
+ config.use_tuner = (target == 2);
+ graph.finalize(target_hint, config);
}
void do_run() override
{
@@ -176,7 +190,7 @@
}
private:
- Graph graph{};
+ Stream graph{ 0, "MobileNetV1_QASYMM8" };
/** This function produces a depthwise separable convolution node (i.e. depthwise + pointwise layers) with ReLU6 activation after each layer.
*
@@ -196,23 +210,19 @@
QuantizationInfo depth_weights_quant_info, QuantizationInfo point_weights_quant_info)
{
std::string total_path = "/cnn_data/mobilenet_qasymm8_model/" + param_path + "_";
- SubGraph sg;
+ SubStream sg(graph);
sg << DepthwiseConvolutionLayer(
3U, 3U,
get_weights_accessor(data_path, total_path + "depthwise_weights.npy"),
get_weights_accessor(data_path, total_path + "depthwise_bias.npy"),
- dwc_pad_stride_info,
- true,
- depth_weights_quant_info)
+ dwc_pad_stride_info, depth_weights_quant_info)
<< ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f))
<< ConvolutionLayer(
1U, 1U, conv_filt,
get_weights_accessor(data_path, total_path + "pointwise_weights.npy"),
get_weights_accessor(data_path, total_path + "pointwise_bias.npy"),
- conv_pad_stride_info,
- 1, WeightsInfo(),
- point_weights_quant_info)
+ conv_pad_stride_info, 1, point_weights_quant_info)
<< ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f));
return BranchLayer(std::move(sg));
@@ -221,9 +231,9 @@
/** Main program for MobileNetQASYMM8
*
* @param[in] argc Number of arguments
- * @param[in] argv Arguments ( [optional] Path to the weights folder, [optional] npy_input, [optional] labels )
+ * @param[in] argv Arguments ( [optional] Path to the weights folder, [optional] npy_input, [optional] labels, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) )
*/
int main(int argc, char **argv)
{
- return utils::run_example<GraphMobileNetQASYMM8Example>(argc, argv);
+ return arm_compute::utils::run_example<GraphMobileNetQASYMM8Example>(argc, argv);
}
diff --git a/examples/graph_resnet50.cpp b/examples/graph_resnet50.cpp
index 1bd76f0..bafa9a5 100644
--- a/examples/graph_resnet50.cpp
+++ b/examples/graph_resnet50.cpp
@@ -21,8 +21,7 @@
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
-#include "arm_compute/graph/Graph.h"
-#include "arm_compute/graph/Nodes.h"
+#include "arm_compute/graph.h"
#include "support/ToolchainSupport.h"
#include "utils/GraphUtils.h"
#include "utils/Utils.h"
@@ -30,13 +29,13 @@
#include <cstdlib>
using namespace arm_compute::utils;
-using namespace arm_compute::graph;
+using namespace arm_compute::graph::frontend;
using namespace arm_compute::graph_utils;
-/** Example demonstrating how to implement Microsoft's ResNet50 network using the Compute Library's graph API
+/** Example demonstrating how to implement ResNet50 network using the Compute Library's graph API
*
* @param[in] argc Number of arguments
- * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL), [optional] Path to the weights folder, [optional] image, [optional] labels )
+ * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) )
*/
class GraphResNet50Example : public Example
{
@@ -53,76 +52,93 @@
false /* Do not convert to BGR */);
// Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON
- const int int_target_hint = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
- TargetHint target_hint = set_target_hint(int_target_hint);
+ const int target = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
+ Target target_hint = set_target_hint(target);
+ FastMathHint fast_math_hint = FastMathHint::DISABLED;
// Parse arguments
if(argc < 2)
{
// Print help
- std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels] [fast_math_hint]\n\n";
std::cout << "No data folder provided: using random values\n\n";
}
else if(argc == 2)
{
- std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels] [fast_math_hint]\n\n";
std::cout << "No data folder provided: using random values\n\n";
}
else if(argc == 3)
{
data_path = argv[2];
- std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels] [fast_math_hint]\n\n";
std::cout << "No image provided: using random values\n\n";
}
else if(argc == 4)
{
data_path = argv[2];
image = argv[3];
- std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels] [fast_math_hint]\n\n";
std::cout << "No text file with labels provided: skipping output accessor\n\n";
}
- else
+ else if(argc == 5)
{
data_path = argv[2];
image = argv[3];
label = argv[4];
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " " << argv[4] << " [fast_math_hint]\n\n";
+ std::cout << "No fast math info provided: disabling fast math\n\n";
+ }
+ else
+ {
+ data_path = argv[2];
+ image = argv[3];
+ label = argv[4];
+ fast_math_hint = (std::strtol(argv[5], nullptr, 1) == 0) ? FastMathHint::DISABLED : FastMathHint::ENABLED;
}
graph << target_hint
- << Tensor(TensorInfo(TensorShape(224U, 224U, 3U, 1U), 1, DataType::F32),
- get_input_accessor(image, std::move(preprocessor), false /* Do not convert to BGR */))
+ << fast_math_hint
+ << InputLayer(TensorDescriptor(TensorShape(224U, 224U, 3U, 1U), DataType::F32),
+ get_input_accessor(image, std::move(preprocessor), false /* Do not convert to BGR */))
<< ConvolutionLayer(
7U, 7U, 64U,
get_weights_accessor(data_path, "/cnn_data/resnet50_model/conv1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(2, 2, 3, 3))
+ .set_name("conv1/convolution")
<< BatchNormalizationLayer(
get_weights_accessor(data_path, "/cnn_data/resnet50_model/conv1_BatchNorm_moving_mean.npy"),
get_weights_accessor(data_path, "/cnn_data/resnet50_model/conv1_BatchNorm_moving_variance.npy"),
get_weights_accessor(data_path, "/cnn_data/resnet50_model/conv1_BatchNorm_gamma.npy"),
get_weights_accessor(data_path, "/cnn_data/resnet50_model/conv1_BatchNorm_beta.npy"),
0.0000100099996416f)
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
- << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR)));
+ .set_name("conv1/BatchNorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv1/Relu")
+ << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR))).set_name("pool1/MaxPool");
add_residual_block(data_path, "block1", 64, 3, 2);
add_residual_block(data_path, "block2", 128, 4, 2);
add_residual_block(data_path, "block3", 256, 6, 2);
add_residual_block(data_path, "block4", 512, 3, 1);
- graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG))
+ graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG)).set_name("pool5")
<< ConvolutionLayer(
1U, 1U, 1000U,
get_weights_accessor(data_path, "/cnn_data/resnet50_model/logits_weights.npy"),
get_weights_accessor(data_path, "/cnn_data/resnet50_model/logits_biases.npy"),
PadStrideInfo(1, 1, 0, 0))
- << FlattenLayer()
- << SoftmaxLayer()
- << Tensor(get_output_accessor(label, 5));
+ .set_name("logits/convolution")
+ << FlattenLayer().set_name("predictions/Reshape")
+ << SoftmaxLayer().set_name("predictions/Softmax")
+ << OutputLayer(get_output_accessor(label, 5));
- // In order to enable the OpenCL tuner, graph_init() has to be called only when all nodes have been instantiated
- graph.graph_init(int_target_hint == 2);
+ // Finalize graph
+ GraphConfig config;
+ config.use_tuner = (target == 2);
+ graph.finalize(target_hint, config);
}
+
void do_run() override
{
// Run graph
@@ -130,15 +146,19 @@
}
private:
- Graph graph{};
+ Stream graph{ 0, "ResNet50" };
void add_residual_block(const std::string &data_path, const std::string &name, unsigned int base_depth, unsigned int num_units, unsigned int stride)
{
for(unsigned int i = 0; i < num_units; ++i)
{
- std::stringstream unit;
- unit << "/cnn_data/resnet50_model/" << name << "_unit_" << (i + 1) << "_bottleneck_v1_";
- std::string unit_name = unit.str();
+ std::stringstream unit_path_ss;
+ unit_path_ss << "/cnn_data/resnet50_model/" << name << "_unit_" << (i + 1) << "_bottleneck_v1_";
+ std::stringstream unit_name_ss;
+ unit_name_ss << name << "/unit" << (i + 1) << "/bottleneck_v1/";
+
+ std::string unit_path = unit_path_ss.str();
+ std::string unit_name = unit_name_ss.str();
unsigned int middle_stride = 1;
@@ -147,76 +167,84 @@
middle_stride = stride;
}
- SubGraph right;
+ SubStream right(graph);
right << ConvolutionLayer(
1U, 1U, base_depth,
- get_weights_accessor(data_path, unit_name + "conv1_weights.npy"),
+ get_weights_accessor(data_path, unit_path + "conv1_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(1, 1, 0, 0))
+ .set_name(unit_name + "conv1/convolution")
<< BatchNormalizationLayer(
- get_weights_accessor(data_path, unit_name + "conv1_BatchNorm_moving_mean.npy"),
- get_weights_accessor(data_path, unit_name + "conv1_BatchNorm_moving_variance.npy"),
- get_weights_accessor(data_path, unit_name + "conv1_BatchNorm_gamma.npy"),
- get_weights_accessor(data_path, unit_name + "conv1_BatchNorm_beta.npy"),
+ get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_moving_mean.npy"),
+ get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_moving_variance.npy"),
+ get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_gamma.npy"),
+ get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_beta.npy"),
0.0000100099996416f)
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ .set_name(unit_name + "conv1/BatchNorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "conv1/Relu")
<< ConvolutionLayer(
3U, 3U, base_depth,
- get_weights_accessor(data_path, unit_name + "conv2_weights.npy"),
+ get_weights_accessor(data_path, unit_path + "conv2_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(middle_stride, middle_stride, 1, 1))
+ .set_name(unit_name + "conv2/convolution")
<< BatchNormalizationLayer(
- get_weights_accessor(data_path, unit_name + "conv2_BatchNorm_moving_mean.npy"),
- get_weights_accessor(data_path, unit_name + "conv2_BatchNorm_moving_variance.npy"),
- get_weights_accessor(data_path, unit_name + "conv2_BatchNorm_gamma.npy"),
- get_weights_accessor(data_path, unit_name + "conv2_BatchNorm_beta.npy"),
+ get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_moving_mean.npy"),
+ get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_moving_variance.npy"),
+ get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_gamma.npy"),
+ get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_beta.npy"),
0.0000100099996416f)
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ .set_name(unit_name + "conv2/BatchNorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "conv1/Relu")
<< ConvolutionLayer(
1U, 1U, base_depth * 4,
- get_weights_accessor(data_path, unit_name + "conv3_weights.npy"),
+ get_weights_accessor(data_path, unit_path + "conv3_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(1, 1, 0, 0))
+ .set_name(unit_name + "conv3/convolution")
<< BatchNormalizationLayer(
- get_weights_accessor(data_path, unit_name + "conv3_BatchNorm_moving_mean.npy"),
- get_weights_accessor(data_path, unit_name + "conv3_BatchNorm_moving_variance.npy"),
- get_weights_accessor(data_path, unit_name + "conv3_BatchNorm_gamma.npy"),
- get_weights_accessor(data_path, unit_name + "conv3_BatchNorm_beta.npy"),
- 0.0000100099996416f);
+ get_weights_accessor(data_path, unit_path + "conv3_BatchNorm_moving_mean.npy"),
+ get_weights_accessor(data_path, unit_path + "conv3_BatchNorm_moving_variance.npy"),
+ get_weights_accessor(data_path, unit_path + "conv3_BatchNorm_gamma.npy"),
+ get_weights_accessor(data_path, unit_path + "conv3_BatchNorm_beta.npy"),
+ 0.0000100099996416f)
+ .set_name(unit_name + "conv2/BatchNorm");
if(i == 0)
{
- SubGraph left;
+ SubStream left(graph);
left << ConvolutionLayer(
1U, 1U, base_depth * 4,
- get_weights_accessor(data_path, unit_name + "shortcut_weights.npy"),
+ get_weights_accessor(data_path, unit_path + "shortcut_weights.npy"),
std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
PadStrideInfo(1, 1, 0, 0))
+ .set_name(unit_name + "shortcut/convolution")
<< BatchNormalizationLayer(
- get_weights_accessor(data_path, unit_name + "shortcut_BatchNorm_moving_mean.npy"),
- get_weights_accessor(data_path, unit_name + "shortcut_BatchNorm_moving_variance.npy"),
- get_weights_accessor(data_path, unit_name + "shortcut_BatchNorm_gamma.npy"),
- get_weights_accessor(data_path, unit_name + "shortcut_BatchNorm_beta.npy"),
- 0.0000100099996416f);
+ get_weights_accessor(data_path, unit_path + "shortcut_BatchNorm_moving_mean.npy"),
+ get_weights_accessor(data_path, unit_path + "shortcut_BatchNorm_moving_variance.npy"),
+ get_weights_accessor(data_path, unit_path + "shortcut_BatchNorm_gamma.npy"),
+ get_weights_accessor(data_path, unit_path + "shortcut_BatchNorm_beta.npy"),
+ 0.0000100099996416f)
+ .set_name(unit_name + "shortcut/BatchNorm");
- graph << ResidualLayer(std::move(left), std::move(right));
+ graph << BranchLayer(BranchMergeMethod::ADD, std::move(left), std::move(right)).set_name(unit_name + "add");
}
else if(middle_stride > 1)
{
- SubGraph left;
- left << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 1, PadStrideInfo(middle_stride, middle_stride, 0, 0), true))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 1.f, 0.f));
+ SubStream left(graph);
+ left << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 1, PadStrideInfo(middle_stride, middle_stride, 0, 0), true)).set_name(unit_name + "shortcut/MaxPool");
- graph << ResidualLayer(std::move(left), std::move(right));
+ graph << BranchLayer(BranchMergeMethod::ADD, std::move(left), std::move(right)).set_name(unit_name + "add");
}
else
{
- graph << ResidualLayer(std::move(right));
+ SubStream left(graph);
+ graph << BranchLayer(BranchMergeMethod::ADD, std::move(left), std::move(right)).set_name(unit_name + "add");
}
- graph << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+ graph << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Relu");
}
}
};
@@ -224,7 +252,7 @@
/** Main program for ResNet50
*
* @param[in] argc Number of arguments
- * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL), [optional] Path to the weights folder, [optional] image, [optional] labels )
+ * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) )
*/
int main(int argc, char **argv)
{
diff --git a/examples/graph_resnext50.cpp b/examples/graph_resnext50.cpp
new file mode 100644
index 0000000..f96a02e
--- /dev/null
+++ b/examples/graph_resnext50.cpp
@@ -0,0 +1,208 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/graph.h"
+#include "support/ToolchainSupport.h"
+#include "utils/GraphUtils.h"
+#include "utils/Utils.h"
+
+#include <cstdlib>
+
+using namespace arm_compute::utils;
+using namespace arm_compute::graph::frontend;
+using namespace arm_compute::graph_utils;
+
+/** Example demonstrating how to implement ResNeXt50 network using the Compute Library's graph API
+ *
+ * @param[in] argc Number of arguments
+ * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] npy_in, [optional] npy_out, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) )
+ */
+class GraphResNeXt50Example : public Example
+{
+public:
+ void do_setup(int argc, char **argv) override
+ {
+ std::string data_path; /* Path to the trainable data */
+ std::string npy_in; /* Input npy data */
+ std::string npy_out; /* Output npy data */
+
+ // Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON
+ const int target = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
+ Target target_hint = set_target_hint(target);
+ FastMathHint fast_math_hint = FastMathHint::DISABLED;
+
+ // Parse arguments
+ if(argc < 2)
+ {
+ // Print help
+ std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [npy_in] [npy_out] [fast_math_hint]\n\n";
+ std::cout << "No data folder provided: using random values\n\n";
+ }
+ else if(argc == 2)
+ {
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [npy_in] [npy_out] [fast_math_hint]\n\n";
+ std::cout << "No data folder provided: using random values\n\n";
+ }
+ else if(argc == 3)
+ {
+ data_path = argv[2];
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [npy_in] [npy_out] [fast_math_hint]\n\n";
+ std::cout << "No input npy file provided: using random values\n\n";
+ }
+ else if(argc == 4)
+ {
+ data_path = argv[2];
+ npy_in = argv[3];
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [npy_out] [fast_math_hint]\n\n";
+ std::cout << "No output npy file provided: skipping output accessor\n\n";
+ }
+ else if(argc == 5)
+ {
+ data_path = argv[2];
+ npy_in = argv[3];
+ npy_out = argv[4];
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " " << argv[4] << " [fast_math_hint]\n\n";
+ std::cout << "No fast math info provided: disabling fast math\n\n";
+ }
+ else
+ {
+ data_path = argv[2];
+ npy_in = argv[3];
+ npy_out = argv[4];
+ fast_math_hint = (std::strtol(argv[5], nullptr, 1) == 0) ? FastMathHint::DISABLED : FastMathHint::ENABLED;
+ }
+
+ graph << target_hint
+ << fast_math_hint
+ << InputLayer(TensorDescriptor(TensorShape(224U, 224U, 3U, 1U), DataType::F32),
+ get_input_accessor(npy_in))
+ << ScaleLayer(get_weights_accessor(data_path, "/cnn_data/resnext50_model/bn_data_mul.npy"),
+ get_weights_accessor(data_path, "/cnn_data/resnext50_model/bn_data_add.npy"))
+ .set_name("bn_data/Scale")
+ << ConvolutionLayer(
+ 7U, 7U, 64U,
+ get_weights_accessor(data_path, "/cnn_data/resnext50_model/conv0_weights.npy"),
+ get_weights_accessor(data_path, "/cnn_data/resnext50_model/conv0_biases.npy"),
+ PadStrideInfo(2, 2, 2, 3, 2, 3, DimensionRoundingType::FLOOR))
+ .set_name("conv0/Convolution")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv0/Relu")
+ << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR))).set_name("pool0");
+
+ add_residual_block(data_path, /*ofm*/ 256, /*stage*/ 1, /*num_unit*/ 3, /*stride_conv_unit1*/ 1);
+ add_residual_block(data_path, 512, 2, 4, 2);
+ add_residual_block(data_path, 1024, 3, 6, 2);
+ add_residual_block(data_path, 2048, 4, 3, 2);
+
+ graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG)).set_name("pool1")
+ << FlattenLayer().set_name("predictions/Reshape")
+ << OutputLayer(get_npy_output_accessor(npy_out, TensorShape(2048U), DataType::F32));
+
+ // Finalize graph
+ GraphConfig config;
+ config.use_tuner = (target == 2);
+ graph.finalize(target_hint, config);
+ }
+
+ void do_run() override
+ {
+ // Run graph
+ graph.run();
+ }
+
+private:
+ Stream graph{ 0, "ResNeXt50" };
+
+ void add_residual_block(const std::string &data_path, unsigned int base_depth, unsigned int stage, unsigned int num_units, unsigned int stride_conv_unit1)
+ {
+ for(unsigned int i = 0; i < num_units; ++i)
+ {
+ std::stringstream unit_path_ss;
+ unit_path_ss << "/cnn_data/resnext50_model/stage" << stage << "_unit" << (i + 1) << "_";
+ std::string unit_path = unit_path_ss.str();
+
+ std::stringstream unit_name_ss;
+ unit_name_ss << "stage" << stage << "/unit" << (i + 1) << "/";
+ std::string unit_name = unit_name_ss.str();
+
+ PadStrideInfo pad_grouped_conv(1, 1, 1, 1);
+ if(i == 0)
+ {
+ pad_grouped_conv = (stage == 1) ? PadStrideInfo(stride_conv_unit1, stride_conv_unit1, 1, 1) : PadStrideInfo(stride_conv_unit1, stride_conv_unit1, 0, 1, 0, 1, DimensionRoundingType::FLOOR);
+ }
+
+ SubStream right(graph);
+ right << ConvolutionLayer(
+ 1U, 1U, base_depth / 2,
+ get_weights_accessor(data_path, unit_path + "conv1_weights.npy"),
+ get_weights_accessor(data_path, unit_path + "conv1_biases.npy"),
+ PadStrideInfo(1, 1, 0, 0))
+ .set_name(unit_name + "conv1/convolution")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "conv1/Relu")
+
+ << ConvolutionLayer(
+ 3U, 3U, base_depth / 2,
+ get_weights_accessor(data_path, unit_path + "conv2_weights.npy"),
+ std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
+ pad_grouped_conv, 32)
+ .set_name(unit_name + "conv2/convolution")
+ << ScaleLayer(get_weights_accessor(data_path, unit_path + "bn2_mul.npy"),
+ get_weights_accessor(data_path, unit_path + "bn2_add.npy"))
+ .set_name(unit_name + "conv1/Scale")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "conv2/Relu")
+
+ << ConvolutionLayer(
+ 1U, 1U, base_depth,
+ get_weights_accessor(data_path, unit_path + "conv3_weights.npy"),
+ get_weights_accessor(data_path, unit_path + "conv3_biases.npy"),
+ PadStrideInfo(1, 1, 0, 0))
+ .set_name(unit_name + "conv3/convolution");
+
+ SubStream left(graph);
+ if(i == 0)
+ {
+ left << ConvolutionLayer(
+ 1U, 1U, base_depth,
+ get_weights_accessor(data_path, unit_path + "sc_weights.npy"),
+ std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
+ PadStrideInfo(stride_conv_unit1, stride_conv_unit1, 0, 0))
+ .set_name(unit_name + "sc/convolution")
+ << ScaleLayer(get_weights_accessor(data_path, unit_path + "sc_bn_mul.npy"),
+ get_weights_accessor(data_path, unit_path + "sc_bn_add.npy"))
+ .set_name(unit_name + "sc/scale");
+ }
+
+ graph << BranchLayer(BranchMergeMethod::ADD, std::move(left), std::move(right)).set_name(unit_name + "add");
+ graph << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Relu");
+ }
+ }
+};
+
+/** Main program for ResNeXt50
+ *
+ * @param[in] argc Number of arguments
+ * @param[in] argv Arguments ( [[optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] npy_in, [optional] npy_out )
+ */
+int main(int argc, char **argv)
+{
+ return arm_compute::utils::run_example<GraphResNeXt50Example>(argc, argv);
+}
diff --git a/examples/graph_squeezenet.cpp b/examples/graph_squeezenet.cpp
index d0c823a..b632688 100644
--- a/examples/graph_squeezenet.cpp
+++ b/examples/graph_squeezenet.cpp
@@ -21,9 +21,7 @@
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
-#include "arm_compute/graph/Graph.h"
-#include "arm_compute/graph/Nodes.h"
-#include "arm_compute/graph/SubGraph.h"
+#include "arm_compute/graph.h"
#include "support/ToolchainSupport.h"
#include "utils/GraphUtils.h"
#include "utils/Utils.h"
@@ -32,18 +30,14 @@
#include <tuple>
using namespace arm_compute::utils;
-using namespace arm_compute::graph;
+using namespace arm_compute::graph::frontend;
using namespace arm_compute::graph_utils;
using namespace arm_compute::logging;
-namespace
-{
-} // namespace
-
/** Example demonstrating how to implement Squeezenet's network using the Compute Library's graph API
*
* @param[in] argc Number of arguments
- * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL), [optional] Path to the weights folder, [optional] image, [optional] labels )
+ * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) )
*/
class GraphSqueezenetExample : public Example
{
@@ -59,44 +53,55 @@
std::unique_ptr<IPreprocessor> preprocessor = arm_compute::support::cpp14::make_unique<CaffePreproccessor>(mean_rgb);
// Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON
- const int int_target_hint = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
- TargetHint target_hint = set_target_hint(int_target_hint);
+ const int target = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
+ Target target_hint = set_target_hint(target);
+ FastMathHint fast_math_hint = FastMathHint::DISABLED;
// Parse arguments
if(argc < 2)
{
// Print help
- std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels] [fast_math_hint]\n\n";
std::cout << "No data folder provided: using random values\n\n";
}
else if(argc == 2)
{
- std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels] [fast_math_hint]\n\n";
std::cout << "No data folder provided: using random values\n\n";
}
else if(argc == 3)
{
data_path = argv[2];
- std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels] [fast_math_hint]\n\n";
std::cout << "No image provided: using random values\n\n";
}
else if(argc == 4)
{
data_path = argv[2];
image = argv[3];
- std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels] [fast_math_hint]\n\n";
std::cout << "No text file with labels provided: skipping output accessor\n\n";
}
- else
+ else if(argc == 5)
{
data_path = argv[2];
image = argv[3];
label = argv[4];
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " " << argv[4] << " [fast_math_hint]\n\n";
+ std::cout << "No fast math info provided: disabling fast math\n\n";
+ }
+ else
+ {
+ data_path = argv[2];
+ image = argv[3];
+ label = argv[4];
+ fast_math_hint = (std::strtol(argv[5], nullptr, 1) == 0) ? FastMathHint::DISABLED : FastMathHint::ENABLED;
}
graph << target_hint
- << Tensor(TensorInfo(TensorShape(224U, 224U, 3U, 1U), 1, DataType::F32),
- get_input_accessor(image, std::move(preprocessor)))
+ << fast_math_hint
+ << InputLayer(TensorDescriptor(TensorShape(224U, 224U, 3U, 1U), DataType::F32),
+ get_input_accessor(image, std::move(preprocessor)))
<< ConvolutionLayer(
7U, 7U, 96U,
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/conv1_w.npy"),
@@ -109,60 +114,60 @@
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire2_squeeze1x1_w.npy"),
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire2_squeeze1x1_b.npy"),
PadStrideInfo(1, 1, 0, 0))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
- << get_expand_fire_node(data_path, "fire2", 64U, 64U)
- << ConvolutionLayer(
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+ graph << get_expand_fire_node(data_path, "fire2", 64U, 64U);
+ graph << ConvolutionLayer(
1U, 1U, 16U,
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire3_squeeze1x1_w.npy"),
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire3_squeeze1x1_b.npy"),
PadStrideInfo(1, 1, 0, 0))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
- << get_expand_fire_node(data_path, "fire3", 64U, 64U)
- << ConvolutionLayer(
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+ graph << get_expand_fire_node(data_path, "fire3", 64U, 64U);
+ graph << ConvolutionLayer(
1U, 1U, 32U,
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire4_squeeze1x1_w.npy"),
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire4_squeeze1x1_b.npy"),
PadStrideInfo(1, 1, 0, 0))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
- << get_expand_fire_node(data_path, "fire4", 128U, 128U)
- << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+ graph << get_expand_fire_node(data_path, "fire4", 128U, 128U);
+ graph << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
<< ConvolutionLayer(
1U, 1U, 32U,
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire5_squeeze1x1_w.npy"),
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire5_squeeze1x1_b.npy"),
PadStrideInfo(1, 1, 0, 0))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
- << get_expand_fire_node(data_path, "fire5", 128U, 128U)
- << ConvolutionLayer(
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+ graph << get_expand_fire_node(data_path, "fire5", 128U, 128U);
+ graph << ConvolutionLayer(
1U, 1U, 48U,
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire6_squeeze1x1_w.npy"),
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire6_squeeze1x1_b.npy"),
PadStrideInfo(1, 1, 0, 0))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
- << get_expand_fire_node(data_path, "fire6", 192U, 192U)
- << ConvolutionLayer(
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+ graph << get_expand_fire_node(data_path, "fire6", 192U, 192U);
+ graph << ConvolutionLayer(
1U, 1U, 48U,
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire7_squeeze1x1_w.npy"),
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire7_squeeze1x1_b.npy"),
PadStrideInfo(1, 1, 0, 0))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
- << get_expand_fire_node(data_path, "fire7", 192U, 192U)
- << ConvolutionLayer(
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+ graph << get_expand_fire_node(data_path, "fire7", 192U, 192U);
+ graph << ConvolutionLayer(
1U, 1U, 64U,
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire8_squeeze1x1_w.npy"),
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire8_squeeze1x1_b.npy"),
PadStrideInfo(1, 1, 0, 0))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
- << get_expand_fire_node(data_path, "fire8", 256U, 256U)
- << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+ graph << get_expand_fire_node(data_path, "fire8", 256U, 256U);
+ graph << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
<< ConvolutionLayer(
1U, 1U, 64U,
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire9_squeeze1x1_w.npy"),
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire9_squeeze1x1_b.npy"),
PadStrideInfo(1, 1, 0, 0))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
- << get_expand_fire_node(data_path, "fire9", 256U, 256U)
- << ConvolutionLayer(
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+ graph << get_expand_fire_node(data_path, "fire9", 256U, 256U);
+ graph << ConvolutionLayer(
1U, 1U, 1000U,
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/conv10_w.npy"),
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/conv10_b.npy"),
@@ -171,10 +176,12 @@
<< PoolingLayer(PoolingLayerInfo(PoolingType::AVG))
<< FlattenLayer()
<< SoftmaxLayer()
- << Tensor(get_output_accessor(label, 5));
+ << OutputLayer(get_output_accessor(label, 5));
- // In order to enable the OpenCL tuner, graph_init() has to be called only when all nodes have been instantiated
- graph.graph_init(int_target_hint == 2);
+ // Finalize graph
+ GraphConfig config;
+ config.use_tuner = (target == 2);
+ graph.finalize(target_hint, config);
}
void do_run() override
{
@@ -183,12 +190,12 @@
}
private:
- Graph graph{};
+ Stream graph{ 0, "SqueezeNetV1" };
BranchLayer get_expand_fire_node(const std::string &data_path, std::string &¶m_path, unsigned int expand1_filt, unsigned int expand3_filt)
{
std::string total_path = "/cnn_data/squeezenet_v1.0_model/" + param_path + "_";
- SubGraph i_a;
+ SubStream i_a(graph);
i_a << ConvolutionLayer(
1U, 1U, expand1_filt,
get_weights_accessor(data_path, total_path + "expand1x1_w.npy"),
@@ -196,7 +203,7 @@
PadStrideInfo(1, 1, 0, 0))
<< ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
- SubGraph i_b;
+ SubStream i_b(graph);
i_b << ConvolutionLayer(
3U, 3U, expand3_filt,
get_weights_accessor(data_path, total_path + "expand3x3_w.npy"),
@@ -211,7 +218,7 @@
/** Main program for Squeezenet v1.0
*
* @param[in] argc Number of arguments
- * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL), [optional] Path to the weights folder, [optional] image, [optional] labels )
+ * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) )
*/
int main(int argc, char **argv)
{
diff --git a/examples/graph_squeezenet_v1_1.cpp b/examples/graph_squeezenet_v1_1.cpp
index 189cc02..9e3466b 100644
--- a/examples/graph_squeezenet_v1_1.cpp
+++ b/examples/graph_squeezenet_v1_1.cpp
@@ -21,9 +21,7 @@
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
-#include "arm_compute/graph/Graph.h"
-#include "arm_compute/graph/Nodes.h"
-#include "arm_compute/graph/SubGraph.h"
+#include "arm_compute/graph.h"
#include "support/ToolchainSupport.h"
#include "utils/GraphUtils.h"
#include "utils/Utils.h"
@@ -32,9 +30,8 @@
#include <tuple>
using namespace arm_compute::utils;
-using namespace arm_compute::graph;
+using namespace arm_compute::graph::frontend;
using namespace arm_compute::graph_utils;
-using namespace arm_compute::logging;
namespace
{
@@ -43,7 +40,7 @@
/** Example demonstrating how to implement Squeezenet's v1.1 network using the Compute Library's graph API
*
* @param[in] argc Number of arguments
- * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL), [optional] Path to the weights folder, [optional] image, [optional] labels )
+ * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) )
*/
class GraphSqueezenet_v1_1Example : public Example
{
@@ -59,44 +56,56 @@
std::unique_ptr<IPreprocessor> preprocessor = arm_compute::support::cpp14::make_unique<CaffePreproccessor>(mean_rgb);
// Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON
- const int int_target_hint = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
- TargetHint target_hint = set_target_hint(int_target_hint);
+ const int target = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
+ Target target_hint = set_target_hint(target);
+ FastMathHint fast_math_hint = FastMathHint::DISABLED;
// Parse arguments
if(argc < 2)
{
// Print help
- std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels] [fast_math_hint]\n\n";
std::cout << "No data folder provided: using random values\n\n";
}
else if(argc == 2)
{
- std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels] [fast_math_hint]\n\n";
std::cout << "No data folder provided: using random values\n\n";
}
else if(argc == 3)
{
data_path = argv[2];
- std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels] [fast_math_hint]\n\n";
std::cout << "No image provided: using random values\n\n";
}
else if(argc == 4)
{
data_path = argv[2];
image = argv[3];
- std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels] [fast_math_hint]\n\n";
std::cout << "No text file with labels provided: skipping output accessor\n\n";
}
- else
+ else if(argc == 5)
{
data_path = argv[2];
image = argv[3];
label = argv[4];
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " " << argv[4] << " [fast_math_hint]\n\n";
+ std::cout << "No fast math info provided: disabling fast math\n\n";
+ }
+ else
+ {
+ data_path = argv[2];
+ image = argv[3];
+ label = argv[4];
+ fast_math_hint = (std::strtol(argv[5], nullptr, 1) == 0) ? FastMathHint::DISABLED : FastMathHint::ENABLED;
}
graph << target_hint
- << Tensor(TensorInfo(TensorShape(227U, 227U, 3U, 1U), 1, DataType::F32),
- get_input_accessor(image, std::move(preprocessor)))
+ << fast_math_hint
+ << InputLayer(TensorDescriptor(TensorShape(227U, 227U, 3U, 1U), DataType::F32),
+ get_input_accessor(image, std::move(preprocessor)))
+ << ConvolutionMethod::DIRECT
<< ConvolutionLayer(
3U, 3U, 64U,
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/conv1_w.npy"),
@@ -104,65 +113,66 @@
PadStrideInfo(2, 2, 0, 0))
<< ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
<< PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
+ << ConvolutionMethod::DEFAULT
<< ConvolutionLayer(
1U, 1U, 16U,
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire2_squeeze1x1_w.npy"),
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire2_squeeze1x1_b.npy"),
PadStrideInfo(1, 1, 0, 0))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
- << get_expand_fire_node(data_path, "fire2", 64U, 64U)
- << ConvolutionLayer(
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+ graph << get_expand_fire_node(data_path, "fire2", 64U, 64U);
+ graph << ConvolutionLayer(
1U, 1U, 16U,
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire3_squeeze1x1_w.npy"),
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire3_squeeze1x1_b.npy"),
PadStrideInfo(1, 1, 0, 0))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
- << get_expand_fire_node(data_path, "fire3", 64U, 64U)
- << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+ graph << get_expand_fire_node(data_path, "fire3", 64U, 64U);
+ graph << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
<< ConvolutionLayer(
1U, 1U, 32U,
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire4_squeeze1x1_w.npy"),
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire4_squeeze1x1_b.npy"),
PadStrideInfo(1, 1, 0, 0))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
- << get_expand_fire_node(data_path, "fire4", 128U, 128U)
- << ConvolutionLayer(
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+ graph << get_expand_fire_node(data_path, "fire4", 128U, 128U);
+ graph << ConvolutionLayer(
1U, 1U, 32U,
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire5_squeeze1x1_w.npy"),
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire5_squeeze1x1_b.npy"),
PadStrideInfo(1, 1, 0, 0))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
- << get_expand_fire_node(data_path, "fire5", 128U, 128U)
- << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+ graph << get_expand_fire_node(data_path, "fire5", 128U, 128U);
+ graph << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
<< ConvolutionLayer(
1U, 1U, 48U,
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire6_squeeze1x1_w.npy"),
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire6_squeeze1x1_b.npy"),
PadStrideInfo(1, 1, 0, 0))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
- << get_expand_fire_node(data_path, "fire6", 192U, 192U)
- << ConvolutionLayer(
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+ graph << get_expand_fire_node(data_path, "fire6", 192U, 192U);
+ graph << ConvolutionLayer(
1U, 1U, 48U,
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire7_squeeze1x1_w.npy"),
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire7_squeeze1x1_b.npy"),
PadStrideInfo(1, 1, 0, 0))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
- << get_expand_fire_node(data_path, "fire7", 192U, 192U)
- << ConvolutionLayer(
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+ graph << get_expand_fire_node(data_path, "fire7", 192U, 192U);
+ graph << ConvolutionLayer(
1U, 1U, 64U,
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire8_squeeze1x1_w.npy"),
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire8_squeeze1x1_b.npy"),
PadStrideInfo(1, 1, 0, 0))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
- << get_expand_fire_node(data_path, "fire8", 256U, 256U)
- << ConvolutionLayer(
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+ graph << get_expand_fire_node(data_path, "fire8", 256U, 256U);
+ graph << ConvolutionLayer(
1U, 1U, 64U,
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire9_squeeze1x1_w.npy"),
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire9_squeeze1x1_b.npy"),
PadStrideInfo(1, 1, 0, 0))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
- << get_expand_fire_node(data_path, "fire9", 256U, 256U)
- << ConvolutionLayer(
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
+ graph << get_expand_fire_node(data_path, "fire9", 256U, 256U);
+ graph << ConvolutionLayer(
1U, 1U, 1000U,
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/conv10_w.npy"),
get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/conv10_b.npy"),
@@ -171,10 +181,12 @@
<< PoolingLayer(PoolingLayerInfo(PoolingType::AVG))
<< FlattenLayer()
<< SoftmaxLayer()
- << Tensor(get_output_accessor(label, 5));
+ << OutputLayer(get_output_accessor(label, 5));
- // In order to enable the OpenCL tuner, graph_init() has to be called only when all nodes have been instantiated
- graph.graph_init(int_target_hint == 2);
+ // Finalize graph
+ GraphConfig config;
+ config.use_tuner = (target == 2);
+ graph.finalize(target_hint, config);
}
void do_run() override
{
@@ -183,12 +195,12 @@
}
private:
- Graph graph{};
+ Stream graph{ 0, "SqueezeNetV1.1" };
BranchLayer get_expand_fire_node(const std::string &data_path, std::string &¶m_path, unsigned int expand1_filt, unsigned int expand3_filt)
{
std::string total_path = "/cnn_data/squeezenet_v1_1_model/" + param_path + "_";
- SubGraph i_a;
+ SubStream i_a(graph);
i_a << ConvolutionLayer(
1U, 1U, expand1_filt,
get_weights_accessor(data_path, total_path + "expand1x1_w.npy"),
@@ -196,7 +208,7 @@
PadStrideInfo(1, 1, 0, 0))
<< ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
- SubGraph i_b;
+ SubStream i_b(graph);
i_b << ConvolutionLayer(
3U, 3U, expand3_filt,
get_weights_accessor(data_path, total_path + "expand3x3_w.npy"),
@@ -211,7 +223,7 @@
/** Main program for Squeezenet v1.1
*
* @param[in] argc Number of arguments
- * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL), [optional] Path to the weights folder, [optional] image, [optional] labels )
+ * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) )
*/
int main(int argc, char **argv)
{
diff --git a/examples/graph_vgg16.cpp b/examples/graph_vgg16.cpp
index c8cc5b2..72e7240 100644
--- a/examples/graph_vgg16.cpp
+++ b/examples/graph_vgg16.cpp
@@ -21,8 +21,7 @@
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
-#include "arm_compute/graph/Graph.h"
-#include "arm_compute/graph/Nodes.h"
+#include "arm_compute/graph.h"
#include "support/ToolchainSupport.h"
#include "utils/GraphUtils.h"
#include "utils/Utils.h"
@@ -30,27 +29,13 @@
#include <cstdlib>
using namespace arm_compute::utils;
-using namespace arm_compute::graph;
+using namespace arm_compute::graph::frontend;
using namespace arm_compute::graph_utils;
-namespace
-{
-/** This function checks if we can use GEMM-based convolution trying to allocate a memory of size "size_in_bytes"
- *
- * @param[in] size_in_bytes Memory size in bytes needed for VGG-16
- *
- * @return The convolution layer hint
- */
-ConvolutionMethodHint convolution_hint_vgg16(size_t size_in_bytes)
-{
- return ((get_mem_free_from_meminfo() * 1024) >= size_in_bytes) ? ConvolutionMethodHint::GEMM : ConvolutionMethodHint::DIRECT;
-}
-} // namespace
-
/** Example demonstrating how to implement VGG16's network using the Compute Library's graph API
*
* @param[in] argc Number of arguments
- * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL), [optional] Path to the weights folder, [optional] image, [optional] labels )
+ * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) )
*/
class GraphVGG16Example : public Example
{
@@ -66,168 +51,198 @@
std::unique_ptr<IPreprocessor> preprocessor = arm_compute::support::cpp14::make_unique<CaffePreproccessor>(mean_rgb);
// Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON
- const int int_target_hint = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
- TargetHint target_hint = set_target_hint(int_target_hint);
+ const int target = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
+ Target target_hint = set_target_hint(target);
+ const bool is_opencl = target_hint == Target::CL;
- // Check if we can use GEMM-based convolutions evaluating if the platform has at least 1.8 GB of available memory
- const size_t memory_required = 1932735283L;
- ConvolutionMethodHint convolution_hint = convolution_hint_vgg16(memory_required);
+ ConvolutionMethod first_convolution3x3_hint = is_opencl ? ConvolutionMethod::DIRECT : ConvolutionMethod::GEMM;
+ ConvolutionMethod convolution3x3_hint = ConvolutionMethod::DEFAULT;
+ FastMathHint fast_math_hint = FastMathHint::DISABLED;
// Parse arguments
if(argc < 2)
{
// Print help
- std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels] [fast_math_hint]\n\n";
std::cout << "No data folder provided: using random values\n\n";
}
else if(argc == 2)
{
- std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels] [fast_math_hint]\n\n";
std::cout << "No data folder provided: using random values\n\n";
}
else if(argc == 3)
{
data_path = argv[2];
- std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels] [fast_math_hint]\n\n";
std::cout << "No image provided: using random values\n\n";
}
else if(argc == 4)
{
data_path = argv[2];
image = argv[3];
- std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels] [fast_math_hint]\n\n";
std::cout << "No text file with labels provided: skipping output accessor\n\n";
}
- else
+ else if(argc == 5)
{
data_path = argv[2];
image = argv[3];
label = argv[4];
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " " << argv[4] << " [fast_math_hint]\n\n";
+ std::cout << "No fast math info provided: disabling fast math\n\n";
+ }
+ else
+ {
+ data_path = argv[2];
+ image = argv[3];
+ label = argv[4];
+ fast_math_hint = (std::strtol(argv[5], nullptr, 1) == 0) ? FastMathHint::DISABLED : FastMathHint::ENABLED;
}
graph << target_hint
- << convolution_hint
- << Tensor(TensorInfo(TensorShape(224U, 224U, 3U, 1U), 1, DataType::F32),
- get_input_accessor(image, std::move(preprocessor)))
+ << fast_math_hint
+ << first_convolution3x3_hint
+ << InputLayer(TensorDescriptor(TensorShape(224U, 224U, 3U, 1U), DataType::F32),
+ get_input_accessor(image, std::move(preprocessor)))
// Layer 1
<< ConvolutionLayer(
3U, 3U, 64U,
get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv1_1_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv1_1_b.npy"),
PadStrideInfo(1, 1, 1, 1))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ .set_name("conv1_1")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv1_1/Relu")
+ << convolution3x3_hint
// Layer 2
<< ConvolutionLayer(
3U, 3U, 64U,
get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv1_2_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv1_2_b.npy"),
PadStrideInfo(1, 1, 1, 1))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
- << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0)))
+ .set_name("conv1_2")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv1_2/Relu")
+ << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0))).set_name("pool1")
// Layer 3
<< ConvolutionLayer(
3U, 3U, 128U,
get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv2_1_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv2_1_b.npy"),
PadStrideInfo(1, 1, 1, 1))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ .set_name("conv2_1")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv2_1/Relu")
// Layer 4
<< ConvolutionLayer(
3U, 3U, 128U,
get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv2_2_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv2_2_b.npy"),
PadStrideInfo(1, 1, 1, 1))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
- << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0)))
+ .set_name("conv2_2")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv2_2/Relu")
+ << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0))).set_name("pool2")
// Layer 5
<< ConvolutionLayer(
3U, 3U, 256U,
get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv3_1_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv3_1_b.npy"),
PadStrideInfo(1, 1, 1, 1))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ .set_name("conv3_1")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv3_1/Relu")
// Layer 6
<< ConvolutionLayer(
3U, 3U, 256U,
get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv3_2_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv3_2_b.npy"),
PadStrideInfo(1, 1, 1, 1))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ .set_name("conv3_2")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv3_2/Relu")
// Layer 7
<< ConvolutionLayer(
3U, 3U, 256U,
get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv3_3_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv3_3_b.npy"),
PadStrideInfo(1, 1, 1, 1))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
- << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0)))
+ .set_name("conv3_3")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv3_3/Relu")
+ << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0))).set_name("pool3")
// Layer 8
<< ConvolutionLayer(
3U, 3U, 512U,
get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv4_1_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv4_1_b.npy"),
PadStrideInfo(1, 1, 1, 1))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ .set_name("conv4_1")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv4_1/Relu")
// Layer 9
<< ConvolutionLayer(
3U, 3U, 512U,
get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv4_2_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv4_2_b.npy"),
PadStrideInfo(1, 1, 1, 1))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ .set_name("conv4_2")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv4_2/Relu")
// Layer 10
<< ConvolutionLayer(
3U, 3U, 512U,
get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv4_3_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv4_3_b.npy"),
PadStrideInfo(1, 1, 1, 1))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
- << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0)))
+ .set_name("conv4_3")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv4_3/Relu")
+ << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0))).set_name("pool4")
// Layer 11
<< ConvolutionLayer(
3U, 3U, 512U,
get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv5_1_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv5_1_b.npy"),
PadStrideInfo(1, 1, 1, 1))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ .set_name("conv5_1")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv5_1/Relu")
// Layer 12
<< ConvolutionLayer(
3U, 3U, 512U,
get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv5_2_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv5_2_b.npy"),
PadStrideInfo(1, 1, 1, 1))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ .set_name("conv5_2")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv5_2/Relu")
// Layer 13
<< ConvolutionLayer(
3U, 3U, 512U,
get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv5_3_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv5_3_b.npy"),
PadStrideInfo(1, 1, 1, 1))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
- << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0)))
+ .set_name("conv5_3")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv5_3/Relu")
+ << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0))).set_name("pool5")
// Layer 14
<< FullyConnectedLayer(
4096U,
get_weights_accessor(data_path, "/cnn_data/vgg16_model/fc6_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg16_model/fc6_b.npy"))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ .set_name("fc6")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Relu")
// Layer 15
<< FullyConnectedLayer(
4096U,
get_weights_accessor(data_path, "/cnn_data/vgg16_model/fc7_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg16_model/fc7_b.npy"))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ .set_name("fc7")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Relu_1")
// Layer 16
<< FullyConnectedLayer(
1000U,
get_weights_accessor(data_path, "/cnn_data/vgg16_model/fc8_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg16_model/fc8_b.npy"))
+ .set_name("fc8")
// Softmax
- << SoftmaxLayer()
- << Tensor(get_output_accessor(label, 5));
+ << SoftmaxLayer().set_name("prob")
+ << OutputLayer(get_output_accessor(label, 5));
- // In order to enable the OpenCL tuner, graph_init() has to be called only when all nodes have been instantiated
- graph.graph_init(int_target_hint == 2);
+ // Finalize graph
+ GraphConfig config;
+ config.use_tuner = (target == 2);
+ graph.finalize(target_hint, config);
}
void do_run() override
{
@@ -236,13 +251,13 @@
}
private:
- Graph graph{};
+ Stream graph{ 0, "VGG16" };
};
/** Main program for VGG16
*
* @param[in] argc Number of arguments
- * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL), [optional] Path to the weights folder, [optional] image, [optional] labels )
+ * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) )
*/
int main(int argc, char **argv)
{
diff --git a/examples/graph_vgg19.cpp b/examples/graph_vgg19.cpp
index 69ae23d..b15c3f2 100644
--- a/examples/graph_vgg19.cpp
+++ b/examples/graph_vgg19.cpp
@@ -21,8 +21,7 @@
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
-#include "arm_compute/graph/Graph.h"
-#include "arm_compute/graph/Nodes.h"
+#include "arm_compute/graph.h"
#include "support/ToolchainSupport.h"
#include "utils/GraphUtils.h"
#include "utils/Utils.h"
@@ -30,13 +29,13 @@
#include <cstdlib>
using namespace arm_compute::utils;
-using namespace arm_compute::graph;
+using namespace arm_compute::graph::frontend;
using namespace arm_compute::graph_utils;
/** Example demonstrating how to implement VGG19's network using the Compute Library's graph API
*
* @param[in] argc Number of arguments
- * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL), [optional] Path to the weights folder, [optional] image, [optional] labels )
+ * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) )
*/
class GraphVGG19Example : public Example
{
@@ -52,175 +51,211 @@
std::unique_ptr<IPreprocessor> preprocessor = arm_compute::support::cpp14::make_unique<CaffePreproccessor>(mean_rgb);
// Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON
- const int int_target_hint = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
- TargetHint target_hint = set_target_hint(int_target_hint);
- ConvolutionMethodHint convolution_hint = ConvolutionMethodHint::DIRECT;
+ const int target = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
+ Target target_hint = set_target_hint(target);
+ FastMathHint fast_math_hint = FastMathHint::DISABLED;
+ const bool is_opencl = target_hint == Target::CL;
+
+ ConvolutionMethod first_convolution3x3_hint = is_opencl ? ConvolutionMethod::DIRECT : ConvolutionMethod::GEMM;
+ ConvolutionMethod convolution3x3_hint = ConvolutionMethod::DEFAULT;
// Parse arguments
if(argc < 2)
{
// Print help
- std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels] [fast_math_hint]\n\n";
std::cout << "No data folder provided: using random values\n\n";
}
else if(argc == 2)
{
- std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels] [fast_math_hint]\n\n";
std::cout << "No data folder provided: using random values\n\n";
}
else if(argc == 3)
{
data_path = argv[2];
- std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels] [fast_math_hint]\n\n";
std::cout << "No image provided: using random values\n\n";
}
else if(argc == 4)
{
data_path = argv[2];
image = argv[3];
- std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels]\n\n";
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels] [fast_math_hint]\n\n";
std::cout << "No text file with labels provided: skipping output accessor\n\n";
}
- else
+ else if(argc == 5)
{
data_path = argv[2];
image = argv[3];
label = argv[4];
+ std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " " << argv[4] << " [fast_math_hint]\n\n";
+ std::cout << "No fast math info provided: disabling fast math\n\n";
+ }
+ else
+ {
+ data_path = argv[2];
+ image = argv[3];
+ label = argv[4];
+ fast_math_hint = (std::strtol(argv[5], nullptr, 1) == 0) ? FastMathHint::DISABLED : FastMathHint::ENABLED;
}
graph << target_hint
- << convolution_hint
- << Tensor(TensorInfo(TensorShape(224U, 224U, 3U, 1U), 1, DataType::F32),
- get_input_accessor(image, std::move(preprocessor)))
+ << first_convolution3x3_hint
+ << fast_math_hint
+ << InputLayer(TensorDescriptor(TensorShape(224U, 224U, 3U, 1U), DataType::F32),
+ get_input_accessor(image, std::move(preprocessor)))
// Layer 1
<< ConvolutionLayer(
3U, 3U, 64U,
get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv1_1_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv1_1_b.npy"),
PadStrideInfo(1, 1, 1, 1))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ .set_name("conv1_1")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv1_1/Relu")
+ << convolution3x3_hint
<< ConvolutionLayer(
3U, 3U, 64U,
get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv1_2_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv1_2_b.npy"),
PadStrideInfo(1, 1, 1, 1))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
- << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0)))
+ .set_name("conv1_2")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv1_2/Relu")
+ << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0))).set_name("pool1")
// Layer 2
<< ConvolutionLayer(
3U, 3U, 128U,
get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv2_1_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv2_1_b.npy"),
PadStrideInfo(1, 1, 1, 1))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ .set_name("conv2_1")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv2_1/Relu")
<< ConvolutionLayer(
3U, 3U, 128U,
get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv2_2_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv2_2_b.npy"),
PadStrideInfo(1, 1, 1, 1))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
- << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0)))
+ .set_name("conv2_2")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv2_2/Relu")
+ << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0))).set_name("pool2")
// Layer 3
<< ConvolutionLayer(
3U, 3U, 256U,
get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_1_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_1_b.npy"),
PadStrideInfo(1, 1, 1, 1))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ .set_name("conv3_1")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv3_1/Relu")
<< ConvolutionLayer(
3U, 3U, 256U,
get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_2_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_2_b.npy"),
PadStrideInfo(1, 1, 1, 1))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ .set_name("conv3_2")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv3_2/Relu")
<< ConvolutionLayer(
3U, 3U, 256U,
get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_3_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_3_b.npy"),
PadStrideInfo(1, 1, 1, 1))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ .set_name("conv3_3")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv3_3/Relu")
<< ConvolutionLayer(
3U, 3U, 256U,
get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_4_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_4_b.npy"),
PadStrideInfo(1, 1, 1, 1))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
- << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0)))
+ .set_name("conv3_4")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv3_4/Relu")
+ << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0))).set_name("pool3")
// Layer 4
<< ConvolutionLayer(
3U, 3U, 512U,
get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_1_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_1_b.npy"),
PadStrideInfo(1, 1, 1, 1))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ .set_name("conv4_1")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv4_1/Relu")
<< ConvolutionLayer(
3U, 3U, 512U,
get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_2_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_2_b.npy"),
PadStrideInfo(1, 1, 1, 1))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ .set_name("conv4_2")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv4_2/Relu")
<< ConvolutionLayer(
3U, 3U, 512U,
get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_3_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_3_b.npy"),
PadStrideInfo(1, 1, 1, 1))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ .set_name("conv4_3")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv4_3/Relu")
<< ConvolutionLayer(
3U, 3U, 512U,
get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_4_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_4_b.npy"),
PadStrideInfo(1, 1, 1, 1))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
- << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0)))
+ .set_name("conv4_4")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv4_4/Relu")
+ << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0))).set_name("pool4")
// Layer 5
<< ConvolutionLayer(
3U, 3U, 512U,
get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_1_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_1_b.npy"),
PadStrideInfo(1, 1, 1, 1))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ .set_name("conv5_1")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv5_1/Relu")
<< ConvolutionLayer(
3U, 3U, 512U,
get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_2_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_2_b.npy"),
PadStrideInfo(1, 1, 1, 1))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ .set_name("conv5_2")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv5_2/Relu")
<< ConvolutionLayer(
3U, 3U, 512U,
get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_3_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_3_b.npy"),
PadStrideInfo(1, 1, 1, 1))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ .set_name("conv5_3")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv5_3/Relu")
<< ConvolutionLayer(
3U, 3U, 512U,
get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_4_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_4_b.npy"),
PadStrideInfo(1, 1, 1, 1))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
- << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0)))
+ .set_name("conv5_4")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv5_4/Relu")
+ << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0))).set_name("pool5")
// Layer 6
<< FullyConnectedLayer(
4096U,
get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc6_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc6_b.npy"))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ .set_name("fc6")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Relu")
// Layer 7
<< FullyConnectedLayer(
4096U,
get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc7_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc7_b.npy"))
- << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
+ .set_name("fc7")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Relu_1")
// Layer 8
<< FullyConnectedLayer(
1000U,
get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc8_w.npy"),
get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc8_b.npy"))
+ .set_name("fc8")
// Softmax
- << SoftmaxLayer()
- << Tensor(get_output_accessor(label, 5));
+ << SoftmaxLayer().set_name("prob")
+ << OutputLayer(get_output_accessor(label, 5));
- // In order to enable the OpenCL tuner, graph_init() has to be called only when all nodes have been instantiated
- graph.graph_init(int_target_hint == 2);
+ // Finalize graph
+ GraphConfig config;
+ config.use_tuner = (target == 2);
+ graph.finalize(target_hint, config);
}
void do_run() override
{
@@ -229,13 +264,13 @@
}
private:
- Graph graph{};
+ Stream graph{ 0, "VGG19" };
};
/** Main program for VGG19
*
* @param[in] argc Number of arguments
- * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL), [optional] Path to the weights folder, [optional] image, [optional] labels )
+ * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) )
*/
int main(int argc, char **argv)
{