examples/graph_resnet_v2_50.cpp

/*
 * Copyright (c) 2018-2021 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 ResNetV2_50 network using the Compute Library's graph API */
class GraphResNetV2_50Example : public Example
{
public:
    GraphResNetV2_50Example()
        : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "ResNetV2_50")
    {
    }
    bool do_setup(int argc, char **argv) override
    {
        // Parse arguments
        cmd_parser.parse(argc, argv);
        cmd_parser.validate();

        // 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;
        }

        // Print parameter values
        std::cout << common_params << std::endl;

        // Get trainable parameters data path
        std::string data_path  = common_params.data_path;
        std::string model_path = "/cnn_data/resnet_v2_50_model/";
        if(!data_path.empty())
        {
            data_path += model_path;
        }

        // Create a preprocessor object
        std::unique_ptr<IPreprocessor> preprocessor = std::make_unique<TFPreproccessor>();

        // Create input descriptor
        const auto        operation_layout = common_params.data_layout;
        const TensorShape tensor_shape     = permute_shape(TensorShape(224U, 224U, 3U, common_params.batches), DataLayout::NCHW, operation_layout);
        TensorDescriptor  input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout);

        // Set weights trained layout
        const DataLayout weights_layout = DataLayout::NCHW;

        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(
                  7U, 7U, 64U,
                  get_weights_accessor(data_path, "conv1_weights.npy", weights_layout),
                  get_weights_accessor(data_path, "conv1_biases.npy", weights_layout),
                  PadStrideInfo(2, 2, 3, 3))
              .set_name("conv1/convolution")
              << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR))).set_name("pool1/MaxPool");

        add_residual_block(data_path, "block1", weights_layout, 64, 3, 2);
        add_residual_block(data_path, "block2", weights_layout, 128, 4, 2);
        add_residual_block(data_path, "block3", weights_layout, 256, 6, 2);
        add_residual_block(data_path, "block4", weights_layout, 512, 3, 1);

        graph << BatchNormalizationLayer(
                  get_weights_accessor(data_path, "postnorm_moving_mean.npy"),
                  get_weights_accessor(data_path, "postnorm_moving_variance.npy"),
                  get_weights_accessor(data_path, "postnorm_gamma.npy"),
                  get_weights_accessor(data_path, "postnorm_beta.npy"),
                  0.000009999999747378752f)
              .set_name("postnorm/BatchNorm")
              << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("postnorm/Relu")
              << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, operation_layout)).set_name("pool5")
              << ConvolutionLayer(
                  1U, 1U, 1001U,
                  get_weights_accessor(data_path, "logits_weights.npy", weights_layout),
                  get_weights_accessor(data_path, "logits_biases.npy"),
                  PadStrideInfo(1, 1, 0, 0))
              .set_name("logits/convolution")
              << FlattenLayer().set_name("predictions/Reshape")
              << 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_mode         = common_params.tuner_mode;
        config.tuner_file         = common_params.tuner_file;
        config.mlgo_file          = common_params.mlgo_file;
        config.use_synthetic_type = arm_compute::is_data_type_quantized(common_params.data_type);
        config.synthetic_type     = common_params.data_type;

        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, const std::string &name, DataLayout weights_layout,
                            unsigned int base_depth, unsigned int num_units, unsigned int stride)
    {
        for(unsigned int i = 0; i < num_units; ++i)
        {
            // Generate unit names
            std::stringstream unit_path_ss;
            unit_path_ss << name << "_unit_" << (i + 1) << "_bottleneck_v2_";
            std::stringstream unit_name_ss;
            unit_name_ss << name << "/unit" << (i + 1) << "/bottleneck_v2/";

            std::string unit_path = unit_path_ss.str();
            std::string unit_name = unit_name_ss.str();

            const TensorShape last_shape = graph.graph().node(graph.tail_node())->output(0)->desc().shape;
            unsigned int      depth_in   = last_shape[arm_compute::get_data_layout_dimension_index(common_params.data_layout, DataLayoutDimension::CHANNEL)];
            unsigned int      depth_out  = base_depth * 4;

            // All units have stride 1 apart from last one
            unsigned int middle_stride = (i == (num_units - 1)) ? stride : 1;

            // Preact
            SubStream preact(graph);
            preact << BatchNormalizationLayer(
                       get_weights_accessor(data_path, unit_path + "preact_moving_mean.npy"),
                       get_weights_accessor(data_path, unit_path + "preact_moving_variance.npy"),
                       get_weights_accessor(data_path, unit_path + "preact_gamma.npy"),
                       get_weights_accessor(data_path, unit_path + "preact_beta.npy"),
                       0.000009999999747378752f)
                   .set_name(unit_name + "preact/BatchNorm")
                   << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "preact/Relu");

            // Create bottleneck path
            SubStream shortcut(graph);
            if(depth_in == depth_out)
            {
                if(middle_stride != 1)
                {
                    shortcut << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 1, common_params.data_layout, PadStrideInfo(middle_stride, middle_stride, 0, 0), true)).set_name(unit_name + "shortcut/MaxPool");
                }
            }
            else
            {
                shortcut.forward_tail(preact.tail_node());
                shortcut << ConvolutionLayer(
                             1U, 1U, depth_out,
                             get_weights_accessor(data_path, unit_path + "shortcut_weights.npy", weights_layout),
                             get_weights_accessor(data_path, unit_path + "shortcut_biases.npy", weights_layout),
                             PadStrideInfo(1, 1, 0, 0))
                         .set_name(unit_name + "shortcut/convolution");
            }

            // Create residual path
            SubStream residual(preact);
            residual << ConvolutionLayer(
                         1U, 1U, base_depth,
                         get_weights_accessor(data_path, unit_path + "conv1_weights.npy", weights_layout),
                         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_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"),
                         0.000009999999747378752f)
                     .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_path + "conv2_weights.npy", weights_layout),
                         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_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"),
                         0.000009999999747378752f)
                     .set_name(unit_name + "conv2/BatchNorm")
                     << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "conv1/Relu")
                     << ConvolutionLayer(
                         1U, 1U, depth_out,
                         get_weights_accessor(data_path, unit_path + "conv3_weights.npy", weights_layout),
                         get_weights_accessor(data_path, unit_path + "conv3_biases.npy", weights_layout),
                         PadStrideInfo(1, 1, 0, 0))
                     .set_name(unit_name + "conv3/convolution");

            graph << EltwiseLayer(std::move(shortcut), std::move(residual), EltwiseOperation::Add).set_name(unit_name + "add");
        }
    }
};

/** Main program for ResNetV2_50
 *
 * Model is based on:
 *      https://arxiv.org/abs/1603.05027
 *      "Identity Mappings in Deep Residual Networks"
 *      Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun
 *
 * Provenance: download.tensorflow.org/models/resnet_v2_50_2017_04_14.tar.gz
 *
 * @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<GraphResNetV2_50Example>(argc, argv);
}