/*
 * Copyright (c) 2017-2018 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/graph2.h"
#include "support/ToolchainSupport.h"
#include "utils/GraphUtils.h"
#include "utils/Utils.h"

#include <cstdlib>
#include <tuple>

using namespace arm_compute::utils;
using namespace arm_compute::graph2::frontend;
using namespace arm_compute::graph_utils;
using namespace arm_compute::logging;

/** 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 )
 */
class GraphSqueezenetExample : public Example
{
public:
    void do_setup(int argc, char **argv) override
    {
        std::string data_path; /* Path to the trainable data */
        std::string image;     /* Image data */
        std::string label;     /* Label data */

        // Create a preprocessor object
        const std::array<float, 3> mean_rgb{ { 122.68f, 116.67f, 104.01f } };
        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 target                   = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
        Target    target_hint              = set_target_hint2(target);
        bool      enable_tuning            = (target == 2);
        bool      enable_memory_management = true;

        ConvolutionMethod convolution_hint = (target_hint == Target::CL) ? ConvolutionMethod::WINOGRAD : ConvolutionMethod::GEMM;

        // Parse arguments
        if(argc < 2)
        {
            // Print help
            std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels]\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 << "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 << "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 << "No text file with labels provided: skipping output accessor\n\n";
        }
        else
        {
            data_path = argv[2];
            image     = argv[3];
            label     = argv[4];
        }

        graph << target_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"),
                  get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/conv1_b.npy"),
                  PadStrideInfo(2, 2, 0, 0))
              << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
              << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
              << convolution_hint
              << ConvolutionLayer(
                  1U, 1U, 16U,
                  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(
                  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(
                  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)))
              << 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(
                  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(
                  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(
                  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)))
              << 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(
                  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"),
                  PadStrideInfo(1, 1, 0, 0))
              << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
              << PoolingLayer(PoolingLayerInfo(PoolingType::AVG))
              << FlattenLayer()
              << SoftmaxLayer()
              << OutputLayer(get_output_accessor(label, 5));

        // Finalize graph
        graph.finalize(target_hint, enable_tuning, enable_memory_management);
    }
    void do_run() override
    {
        // Run graph
        graph.run();
    }

private:
    Stream graph{ 0, "SqueezeNetV1" };

    BranchLayer get_expand_fire_node(const std::string &data_path, std::string &&param_path, unsigned int expand1_filt, unsigned int expand3_filt)
    {
        std::string total_path = "/cnn_data/squeezenet_v1.0_model/" + param_path + "_";
        SubStream   i_a(graph);
        i_a << ConvolutionLayer(
                1U, 1U, expand1_filt,
                get_weights_accessor(data_path, total_path + "expand1x1_w.npy"),
                get_weights_accessor(data_path, total_path + "expand1x1_b.npy"),
                PadStrideInfo(1, 1, 0, 0))
            << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));

        SubStream i_b(graph);
        i_b << ConvolutionLayer(
                3U, 3U, expand3_filt,
                get_weights_accessor(data_path, total_path + "expand3x3_w.npy"),
                get_weights_accessor(data_path, total_path + "expand3x3_b.npy"),
                PadStrideInfo(1, 1, 1, 1))
            << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));

        return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b));
    }
};

/** 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 )
 */
int main(int argc, char **argv)
{
    return arm_compute::utils::run_example<GraphSqueezenetExample>(argc, argv);
}