COMPMID-1578: Add ShuffleNet graph example
Change-Id: I0e577568b74cf5b2285e6f8c51f8403c9d5f78d2
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/148747
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>
diff --git a/examples/graph_alexnet.cpp b/examples/graph_alexnet.cpp
index 124f672..97788d8 100644
--- a/examples/graph_alexnet.cpp
+++ b/examples/graph_alexnet.cpp
@@ -31,11 +31,7 @@
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
- */
+/** Example demonstrating how to implement AlexNet's network using the Compute Library's graph API */
class GraphAlexnetExample : public Example
{
public:
diff --git a/examples/graph_cl_shufflenet.cpp b/examples/graph_cl_shufflenet.cpp
new file mode 100644
index 0000000..70695ba
--- /dev/null
+++ b/examples/graph_cl_shufflenet.cpp
@@ -0,0 +1,251 @@
+/*
+ * 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/CommonGraphOptions.h"
+#include "utils/GraphUtils.h"
+#include "utils/Utils.h"
+
+using namespace arm_compute::utils;
+using namespace arm_compute::graph::frontend;
+using namespace arm_compute::graph_utils;
+
+/** Example demonstrating how to implement ShuffleNet network using the Compute Library's graph API */
+class ShuffleNetExample : public Example
+{
+public:
+ ShuffleNetExample()
+ : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "ShuffleNet")
+ {
+ }
+ bool do_setup(int argc, char **argv) override
+ {
+ // Parse arguments
+ cmd_parser.parse(argc, argv);
+
+ // Consume common parameters
+ common_params = consume_common_graph_parameters(common_opts);
+
+ // Return when help menu is requested
+ if(common_params.help)
+ {
+ cmd_parser.print_help(argv[0]);
+ return false;
+ }
+
+ // Set default layout if needed (Single kernel grouped convolution not yet supported int NHWC)
+ if(!common_opts.data_layout->is_set())
+ {
+ common_params.data_layout = DataLayout::NCHW;
+ }
+
+ // Checks
+ ARM_COMPUTE_EXIT_ON_MSG(arm_compute::is_data_type_quantized_asymmetric(common_params.data_type), "QASYMM8 not supported for this graph");
+ ARM_COMPUTE_EXIT_ON_MSG(common_params.target != Target::CL, "Only CL backend is supported for this network");
+
+ // Print parameter values
+ std::cout << common_params << std::endl;
+ std::cout << "Model: Shufflenet_1_g4" << std::endl;
+
+ // Create model path
+ std::string model_path = "/cnn_data/shufflenet_model/";
+
+ // Get trainable parameters data path
+ std::string data_path = common_params.data_path;
+
+ // Add model path to data path
+ if(!data_path.empty())
+ {
+ data_path += model_path;
+ }
+
+ // Create input descriptor
+ const TensorShape tensor_shape = permute_shape(TensorShape(224U, 224U, 3U, 1U), DataLayout::NCHW, common_params.data_layout);
+ TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(common_params.data_layout);
+
+ // Set weights trained layout
+ const DataLayout weights_layout = DataLayout::NCHW;
+
+ // Create preprocessor
+ std::unique_ptr<IPreprocessor> preprocessor = arm_compute::support::cpp14::make_unique<TFPreproccessor>(0);
+
+ graph << common_params.target
+ << common_params.fast_math_hint
+ << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor), false /* Do not convert to BGR */))
+ << ConvolutionLayer(
+ 3U, 3U, 24U,
+ get_weights_accessor(data_path, "conv3_0_w_0.npy", weights_layout),
+ get_weights_accessor(data_path, "conv3_0_b_0.npy", weights_layout),
+ PadStrideInfo(2, 2, 1, 1))
+ .set_name("Conv1/convolution")
+ << BatchNormalizationLayer(
+ get_weights_accessor(data_path, "conv3_0_bn_rm_0.npy"),
+ get_weights_accessor(data_path, "conv3_0_bn_riv_0.npy"),
+ get_weights_accessor(data_path, "conv3_0_bn_s_0.npy"),
+ get_weights_accessor(data_path, "conv3_0_bn_b_0.npy"),
+ 1e-5f)
+ .set_name("Conv1/BatchNorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv1/Relu")
+ << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 1, 1))).set_name("pool1/MaxPool");
+
+ // Stage 2
+ add_residual_block(data_path, DataLayout::NCHW, 0U /* unit */, 112U /* depth */, 2U /* stride */);
+ add_residual_block(data_path, DataLayout::NCHW, 1U /* unit */, 136U /* depth */, 1U /* stride */);
+ add_residual_block(data_path, DataLayout::NCHW, 2U /* unit */, 136U /* depth */, 1U /* stride */);
+ add_residual_block(data_path, DataLayout::NCHW, 3U /* unit */, 136U /* depth */, 1U /* stride */);
+
+ // Stage 3
+ add_residual_block(data_path, DataLayout::NCHW, 4U /* unit */, 136U /* depth */, 2U /* stride */);
+ add_residual_block(data_path, DataLayout::NCHW, 5U /* unit */, 272U /* depth */, 1U /* stride */);
+ add_residual_block(data_path, DataLayout::NCHW, 6U /* unit */, 272U /* depth */, 1U /* stride */);
+ add_residual_block(data_path, DataLayout::NCHW, 7U /* unit */, 272U /* depth */, 1U /* stride */);
+ add_residual_block(data_path, DataLayout::NCHW, 8U /* unit */, 272U /* depth */, 1U /* stride */);
+ add_residual_block(data_path, DataLayout::NCHW, 9U /* unit */, 272U /* depth */, 1U /* stride */);
+ add_residual_block(data_path, DataLayout::NCHW, 10U /* unit */, 272U /* depth */, 1U /* stride */);
+ add_residual_block(data_path, DataLayout::NCHW, 11U /* unit */, 272U /* depth */, 1U /* stride */);
+
+ // Stage 4
+ add_residual_block(data_path, DataLayout::NCHW, 12U /* unit */, 272U /* depth */, 2U /* stride */);
+ add_residual_block(data_path, DataLayout::NCHW, 13U /* unit */, 544U /* depth */, 1U /* stride */);
+ add_residual_block(data_path, DataLayout::NCHW, 14U /* unit */, 544U /* depth */, 1U /* stride */);
+ add_residual_block(data_path, DataLayout::NCHW, 15U /* unit */, 544U /* depth */, 1U /* stride */);
+
+ graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG)).set_name("predictions/AvgPool")
+ << FlattenLayer().set_name("predictions/Reshape")
+ << FullyConnectedLayer(
+ 1000U,
+ get_weights_accessor(data_path, "pred_w_0.npy", weights_layout),
+ get_weights_accessor(data_path, "pred_b_0.npy"))
+ .set_name("predictions/FC")
+ << SoftmaxLayer().set_name("predictions/Softmax")
+ << OutputLayer(get_output_accessor(common_params, 5));
+
+ // Finalize graph
+ GraphConfig config;
+ config.num_threads = common_params.threads;
+ config.use_tuner = common_params.enable_tuner;
+ config.tuner_file = common_params.tuner_file;
+
+ graph.finalize(common_params.target, config);
+
+ return true;
+ }
+
+ void do_run() override
+ {
+ // Run graph
+ graph.run();
+ }
+
+private:
+ CommandLineParser cmd_parser;
+ CommonGraphOptions common_opts;
+ CommonGraphParams common_params;
+ Stream graph;
+
+ void add_residual_block(const std::string &data_path, DataLayout weights_layout,
+ unsigned int unit, unsigned int depth, unsigned int stride)
+ {
+ PadStrideInfo dwc_info = PadStrideInfo(1, 1, 1, 1);
+ const unsigned int gconv_id = unit * 2;
+ const unsigned int num_groups = 4;
+ const std::string unit_id_name = arm_compute::support::cpp11::to_string(unit);
+ const std::string gconv_id_name = arm_compute::support::cpp11::to_string(gconv_id);
+ const std::string gconv_id_1_name = arm_compute::support::cpp11::to_string(gconv_id + 1);
+ const std::string unit_name = "unit" + unit_id_name;
+
+ SubStream left_ss(graph);
+ SubStream right_ss(graph);
+
+ if(stride == 2)
+ {
+ right_ss << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(2, 2, 1, 1))).set_name(unit_name + "/pool_1/AveragePool");
+ dwc_info = PadStrideInfo(2, 2, 1, 1);
+ }
+
+ left_ss << ConvolutionLayer(
+ 1U, 1U, depth,
+ get_weights_accessor(data_path, "gconv1_" + gconv_id_name + "_w_0.npy", weights_layout),
+ std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
+ PadStrideInfo(1, 1, 0, 0), num_groups)
+ .set_name(unit_name + "/gconv1_" + gconv_id_name + "/convolution")
+ << BatchNormalizationLayer(
+ get_weights_accessor(data_path, "gconv1_" + gconv_id_name + "_bn_rm_0.npy"),
+ get_weights_accessor(data_path, "gconv1_" + gconv_id_name + "_bn_riv_0.npy"),
+ get_weights_accessor(data_path, "gconv1_" + gconv_id_name + "_bn_s_0.npy"),
+ get_weights_accessor(data_path, "gconv1_" + gconv_id_name + "_bn_b_0.npy"),
+ 1e-5f)
+ .set_name(unit_name + "/gconv1_" + gconv_id_name + "/BatchNorm")
+ << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "/gconv1_" + gconv_id_name + "/Relu")
+ << ChannelShuffleLayer(num_groups).set_name(unit_name + "/shuffle_0/ChannelShufle")
+ << DepthwiseConvolutionLayer(
+ 3U, 3U,
+ get_weights_accessor(data_path, "gconv3_" + unit_id_name + "_w_0.npy", weights_layout),
+ std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
+ dwc_info)
+ .set_name(unit_name + "/gconv3_" + unit_id_name + "/depthwise")
+ << BatchNormalizationLayer(
+ get_weights_accessor(data_path, "gconv3_" + unit_id_name + "_bn_rm_0.npy"),
+ get_weights_accessor(data_path, "gconv3_" + unit_id_name + "_bn_riv_0.npy"),
+ get_weights_accessor(data_path, "gconv3_" + unit_id_name + "_bn_s_0.npy"),
+ get_weights_accessor(data_path, "gconv3_" + unit_id_name + "_bn_b_0.npy"),
+ 1e-5f)
+ .set_name(unit_name + "/gconv3_" + unit_id_name + "/BatchNorm")
+ << ConvolutionLayer(
+ 1U, 1U, depth,
+ get_weights_accessor(data_path, "gconv1_" + gconv_id_1_name + "_w_0.npy", weights_layout),
+ std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
+ PadStrideInfo(1, 1, 0, 0), num_groups)
+ .set_name(unit_name + "/gconv1_" + gconv_id_1_name + "/convolution")
+ << BatchNormalizationLayer(
+ get_weights_accessor(data_path, "gconv1_" + gconv_id_1_name + "_bn_rm_0.npy"),
+ get_weights_accessor(data_path, "gconv1_" + gconv_id_1_name + "_bn_riv_0.npy"),
+ get_weights_accessor(data_path, "gconv1_" + gconv_id_1_name + "_bn_s_0.npy"),
+ get_weights_accessor(data_path, "gconv1_" + gconv_id_1_name + "_bn_b_0.npy"),
+ 1e-5f)
+ .set_name(unit_name + "/gconv1_" + gconv_id_1_name + "/BatchNorm");
+
+ if(stride == 2)
+ {
+ graph << ConcatLayer(std::move(left_ss), std::move(right_ss)).set_name(unit_name + "/Concat");
+ }
+ else
+ {
+ graph << EltwiseLayer(std::move(left_ss), std::move(right_ss), EltwiseOperation::Add).set_name(unit_name + "/Add");
+ }
+ graph << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "/Relu");
+ }
+};
+
+/** Main program for ShuffleNet
+ *
+ * @note To list all the possible arguments execute the binary appended with the --help option
+ *
+ * @param[in] argc Number of arguments
+ * @param[in] argv Arguments
+ */
+int main(int argc, char **argv)
+{
+ return arm_compute::utils::run_example<ShuffleNetExample>(argc, argv);
+}
diff --git a/examples/graph_googlenet.cpp b/examples/graph_googlenet.cpp
index f6aad5d..c0954f7 100644
--- a/examples/graph_googlenet.cpp
+++ b/examples/graph_googlenet.cpp
@@ -31,11 +31,7 @@
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
- */
+/** Example demonstrating how to implement Googlenet's network using the Compute Library's graph API */
class GraphGooglenetExample : public Example
{
public:
diff --git a/examples/graph_inception_resnet_v2.cpp b/examples/graph_inception_resnet_v2.cpp
index 150de7d..7e9fd41 100644
--- a/examples/graph_inception_resnet_v2.cpp
+++ b/examples/graph_inception_resnet_v2.cpp
@@ -31,11 +31,7 @@
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
- */
+/** Example demonstrating how to implement InceptionV4's network using the Compute Library's graph API */
class InceptionResNetV2Example final : public Example
{
public:
diff --git a/examples/graph_inception_v3.cpp b/examples/graph_inception_v3.cpp
index 80e771b..d51be62 100644
--- a/examples/graph_inception_v3.cpp
+++ b/examples/graph_inception_v3.cpp
@@ -31,11 +31,7 @@
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
- */
+/** Example demonstrating how to implement InceptionV3's network using the Compute Library's graph API */
class InceptionV3Example : public Example
{
public:
diff --git a/examples/graph_inception_v4.cpp b/examples/graph_inception_v4.cpp
index 00a6c7d..b64f176 100644
--- a/examples/graph_inception_v4.cpp
+++ b/examples/graph_inception_v4.cpp
@@ -31,11 +31,7 @@
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
- */
+/** Example demonstrating how to implement InceptionV4's network using the Compute Library's graph API */
class InceptionV4Example final : public Example
{
public:
diff --git a/examples/graph_lenet.cpp b/examples/graph_lenet.cpp
index 6b9f302..79cf122 100644
--- a/examples/graph_lenet.cpp
+++ b/examples/graph_lenet.cpp
@@ -31,11 +31,7 @@
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, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) )
- */
+/** Example demonstrating how to implement LeNet's network using the Compute Library's graph API */
class GraphLenetExample : public Example
{
public:
diff --git a/examples/graph_mobilenet.cpp b/examples/graph_mobilenet.cpp
index ab6a4a8..7a6f981 100644
--- a/examples/graph_mobilenet.cpp
+++ b/examples/graph_mobilenet.cpp
@@ -32,11 +32,7 @@
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
- */
+/** Example demonstrating how to implement MobileNet's network using the Compute Library's graph API */
class GraphMobilenetExample : public Example
{
public:
diff --git a/examples/graph_resnet50.cpp b/examples/graph_resnet50.cpp
index 5b6a481..f13383e 100644
--- a/examples/graph_resnet50.cpp
+++ b/examples/graph_resnet50.cpp
@@ -31,11 +31,7 @@
using namespace arm_compute::graph::frontend;
using namespace arm_compute::graph_utils;
-/** Example demonstrating how to implement ResNetV1_50 network using the Compute Library's graph API
- *
- * @param[in] argc Number of arguments
- * @param[in] argv Arguments
- */
+/** Example demonstrating how to implement ResNetV1_50 network using the Compute Library's graph API */
class GraphResNetV1_50Example : public Example
{
public:
diff --git a/examples/graph_resnet_v2_50.cpp b/examples/graph_resnet_v2_50.cpp
index 03c21ac..b2c85e8 100644
--- a/examples/graph_resnet_v2_50.cpp
+++ b/examples/graph_resnet_v2_50.cpp
@@ -31,11 +31,7 @@
using namespace arm_compute::graph::frontend;
using namespace arm_compute::graph_utils;
-/** Example demonstrating how to implement ResNetV2_50 network using the Compute Library's graph API
- *
- * @param[in] argc Number of arguments
- * @param[in] argv Arguments
- */
+/** Example demonstrating how to implement ResNetV2_50 network using the Compute Library's graph API */
class GraphResNetV2_50Example : public Example
{
public:
diff --git a/examples/graph_resnext50.cpp b/examples/graph_resnext50.cpp
index c369fa9..a52e181 100644
--- a/examples/graph_resnext50.cpp
+++ b/examples/graph_resnext50.cpp
@@ -31,11 +31,7 @@
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
- */
+/** Example demonstrating how to implement ResNeXt50 network using the Compute Library's graph API */
class GraphResNeXt50Example : public Example
{
public:
diff --git a/examples/graph_squeezenet.cpp b/examples/graph_squeezenet.cpp
index cee0ffb..ddb4474 100644
--- a/examples/graph_squeezenet.cpp
+++ b/examples/graph_squeezenet.cpp
@@ -31,11 +31,7 @@
using namespace arm_compute::graph::frontend;
using namespace arm_compute::graph_utils;
-/** 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
- */
+/** Example demonstrating how to implement Squeezenet's network using the Compute Library's graph API */
class GraphSqueezenetExample : public Example
{
public:
diff --git a/examples/graph_squeezenet_v1_1.cpp b/examples/graph_squeezenet_v1_1.cpp
index 013664d..de3d4e8 100644
--- a/examples/graph_squeezenet_v1_1.cpp
+++ b/examples/graph_squeezenet_v1_1.cpp
@@ -31,11 +31,7 @@
using namespace arm_compute::graph::frontend;
using namespace arm_compute::graph_utils;
-/** 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
- */
+/** Example demonstrating how to implement Squeezenet's v1.1 network using the Compute Library's graph API */
class GraphSqueezenet_v1_1Example : public Example
{
public:
diff --git a/examples/graph_vgg16.cpp b/examples/graph_vgg16.cpp
index 69b3a9d..8e1e621 100644
--- a/examples/graph_vgg16.cpp
+++ b/examples/graph_vgg16.cpp
@@ -31,11 +31,7 @@
using namespace arm_compute::graph::frontend;
using namespace arm_compute::graph_utils;
-/** 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
- */
+/** Example demonstrating how to implement VGG16's network using the Compute Library's graph API */
class GraphVGG16Example : public Example
{
public:
diff --git a/examples/graph_vgg19.cpp b/examples/graph_vgg19.cpp
index 8d64c1c..6127c74 100644
--- a/examples/graph_vgg19.cpp
+++ b/examples/graph_vgg19.cpp
@@ -30,11 +30,7 @@
using namespace arm_compute::utils;
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
- */
+/** Example demonstrating how to implement VGG19's network using the Compute Library's graph API */
class GraphVGG19Example : public Example
{
public:
diff --git a/examples/graph_yolov3.cpp b/examples/graph_yolov3.cpp
index 24186d5..098855e 100644
--- a/examples/graph_yolov3.cpp
+++ b/examples/graph_yolov3.cpp
@@ -31,11 +31,7 @@
using namespace arm_compute::graph::frontend;
using namespace arm_compute::graph_utils;
-/** Example demonstrating how to implement YOLOv3 network using the Compute Library's graph API
- *
- * @param[in] argc Number of arguments
- * @param[in] argv Arguments
- */
+/** Example demonstrating how to implement YOLOv3 network using the Compute Library's graph API */
class GraphYOLOv3Example : public Example
{
public: