ncnn/tools/ncnnoptimize.cpp

// Tencent is pleased to support the open source community by making ncnn available.
//
// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
//
// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
// in compliance with the License. You may obtain a copy of the License at
//
// https://opensource.org/licenses/BSD-3-Clause
//
// Unless required by applicable law or agreed to in writing, software distributed
// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
// CONDITIONS OF ANY KIND, either express or implied. See the License for the
// specific language governing permissions and limitations under the License.

#include <algorithm>
#include <set>
#include <vector>

// ncnn public header
#include "datareader.h"
#include "net.h"
#include "layer.h"

// ncnn private header
#include "layer/batchnorm.h"
#include "layer/bias.h"
#include "layer/binaryop.h"
#include "layer/clip.h"
#include "layer/concat.h"
#include "layer/convolution.h"
#include "layer/convolutiondepthwise.h"
#include "layer/crop.h"
#include "layer/deconvolution.h"
#include "layer/deconvolutiondepthwise.h"
#include "layer/detectionoutput.h"
#include "layer/dropout.h"
#include "layer/eltwise.h"
#include "layer/elu.h"
#include "layer/exp.h"
#include "layer/expanddims.h"
#include "layer/flatten.h"
#include "layer/hardsigmoid.h"
#include "layer/hardswish.h"
#include "layer/innerproduct.h"
#include "layer/input.h"
#include "layer/instancenorm.h"
#include "layer/interp.h"
#include "layer/log.h"
#include "layer/lrn.h"
#include "layer/memorydata.h"
#include "layer/mvn.h"
#include "layer/normalize.h"
#include "layer/padding.h"
#include "layer/permute.h"
#include "layer/pooling.h"
#include "layer/power.h"
#include "layer/prelu.h"
#include "layer/priorbox.h"
#include "layer/proposal.h"
#include "layer/psroipooling.h"
#include "layer/quantize.h"
#include "layer/reduction.h"
#include "layer/relu.h"
#include "layer/reorg.h"
#include "layer/requantize.h"
#include "layer/reshape.h"
#include "layer/roialign.h"
#include "layer/roipooling.h"
#include "layer/scale.h"
#include "layer/slice.h"
#include "layer/shufflechannel.h"
#include "layer/softmax.h"
#include "layer/squeeze.h"
#include "layer/threshold.h"
#include "layer/unaryop.h"
#include "layer/yolodetectionoutput.h"
#include "layer/yolov3detectionoutput.h"

#if defined(__aarch64__) && defined(LINUX)
#include <locale>
#include <chrono>
#include <random>
#include <limits>
#include <cassert>

#define TEXT_GREEN  "\033[32m"
#define TEXT_YELLOW "\033[33m"
#define TEXT_RED    "\033[31m"
#define CLR         "\033[0m"

#endif // defined(__aarch64__) && defined(LINUX)

class DataReaderFromEmpty : public ncnn::DataReader
{
public:
    virtual int scan(const char* format, void* p) const { return 0; }
    virtual size_t read(void* /*buf*/, size_t size) const { return size; }
};

class NetOptimize : public ncnn::Net
{
public:
    // 0=fp32 1=fp16
    int storage_type;

public:
    int fuse_batchnorm_scale();
    int fuse_convolution_batchnorm();
    int fuse_convolutiondepthwise_batchnorm();
    int fuse_deconvolution_batchnorm();
    int fuse_deconvolutiondepthwise_batchnorm();
    int fuse_innerproduct_batchnorm();
    int fuse_innerproduct_dropout();
    int fuse_convolution_activation();
    int fuse_convolutiondepthwise_activation();
    int fuse_deconvolution_activation();
    int fuse_deconvolutiondepthwise_activation();
    int fuse_innerproduct_activation();

    int eliminate_dropout();
    int eliminate_pooling1x1();
    int eliminate_noop();
    int eliminate_orphaned_memorydata();
    int eliminate_flatten_after_global_pooling();
    int eliminate_reshape_after_global_pooling();
    int eliminate_flatten_after_innerproduct();
    int eliminate_reshape_before_binaryop();

    int replace_convolution_with_innerproduct_after_global_pooling();
    int replace_convolution_with_innerproduct_after_innerproduct();

public:
    int fprintf_param_int_array(int id, const ncnn::Mat& m, FILE* pp);
    int fprintf_param_float_array(int id, const ncnn::Mat& m, FILE* pp);

    int fwrite_weight_tag_data(int tag, const ncnn::Mat& data, FILE* bp);
    int fwrite_weight_data(const ncnn::Mat& data, FILE* bp);

    int save(const char* parampath, const char* binpath);

#if defined(__aarch64__) && defined(LINUX)
    void gauss_random(ncnn::Mat &m);
    void find_fastest_fp32_conv(const char* name, int w, int h, int c);
    int support_fp32_conv_type(const ncnn::Convolution* op, const ncnn::Mat& mat, const int type);
#endif
};

#if defined(__aarch64__) && defined(LINUX)
void NetOptimize::gauss_random(ncnn::Mat &m)
{
    std::random_device rd;
    std::mt19937 gen(rd());
    std::normal_distribution<float> d(1.0f, 1.0f);
    int size = m.total();
    for (int i = 0; i < size; ++i)
    {
        m[i] = d(gen);
    }
}

void NetOptimize::find_fastest_fp32_conv(const char* dataname, int w, int h, int c)
{
    ncnn::PoolAllocator allocator;
    allocator.clear();

    ncnn::Option opt;
    // embeded system generally use single thread
    opt.num_threads = 1;

    const size_t layer_count = layers.size();
    ncnn::Extractor ex = create_extractor();
    
    ncnn::Mat input(w, h, c);
    if (ex.input(dataname, input) < 0)
    {
        fprintf(stderr, "set input failed, check dataname.\n");
        return;
    }

    const char* IMPL_NAME[6] = {"baseline", "winograd", "pointwise", "im2col", "direct", "conv3x3s2"};
    for (int i = 0; i < layer_count; ++i)
    {
        ncnn::Layer* layer = layers[i];
        if (layer->type == "Convolution")
        {
            ncnn::Convolution* op = (ncnn::Convolution*)layer;

            ncnn::Mat bottom_blob;
            ncnn::Mat top_blob;

            ex.extract(layer->bottoms[0], bottom_blob);
            ex.extract(layer->tops[0], top_blob);

            if (bottom_blob.empty() || top_blob.empty())
            {
                continue;
            }

            ncnn::Mat weight_blob(op->kernel_w, op->kernel_h, bottom_blob.c * top_blob.c);

            fprintf(stdout, TEXT_GREEN "Input  [w h nc]: %d %d %d\n" CLR, bottom_blob.w, bottom_blob.h, bottom_blob.c);
            fprintf(stdout, TEXT_GREEN "Kernel [w h nc]: %d %d %d\n" CLR, op->kernel_w, op->kernel_h, bottom_blob.c * top_blob.c);
            fprintf(stdout, TEXT_GREEN "Output [w h nc]: %d %d %d\n" CLR, top_blob.w, top_blob.h, top_blob.c);

            // randomize input and kernel
            gauss_random(bottom_blob);
            
            // try every implementation
            double min_cost = std::numeric_limits<double>::max(); 
            int best_type = 0;

            // how much conv implementation type ncnn has ?
            for (int type = 1; type <= 5; ++type)
            {
                int support = support_fp32_conv_type(op, bottom_blob, type);
                if (support < 1)
                {
                    // implementation type mismatch convolution configuration, skip
                    continue;
                }

                op->impl_type = type;

                auto start = std::chrono::high_resolution_clock::now();
                const int NREPEATS = 20;
                op->create_pipeline(opt);
                for (int repeat = 0; repeat < NREPEATS; ++repeat)
                {
                    op->forward(top_blob, bottom_blob, opt);
                }
                op->destroy_pipeline(opt);

                auto stop = std::chrono::high_resolution_clock::now();
                double cur_cost = std::chrono::duration<double, std::micro>(stop-start).count() / NREPEATS;
                fprintf(stdout, TEXT_GREEN "%s cost %0.3lfms \n" CLR, IMPL_NAME[type], cur_cost/1000);
                if (cur_cost < min_cost)
                {
                    min_cost = cur_cost;
                    best_type = type;
                }
            }
            op->impl_type = best_type;
            
            fprintf(stdout, TEXT_YELLOW "%d: %s use %s \n\n" CLR, i, layer->name.c_str(), IMPL_NAME[op->impl_type]);
        }
    }
}

int NetOptimize::support_fp32_conv_type(const ncnn::Convolution* op, const ncnn::Mat& bottom, const int type)
{
    // not baseline, then k_h == k_w and s_h == s_w
    // no dilation conv shall be allowed
    if (op->kernel_w != op->kernel_h ||
        op->stride_w != op->stride_h ||
        op->dilation_w != op->dilation_h ||
        op->dilation_h != 1)
    {
        return -1;
    }

    // (kernel, stride) in {(1, 1), (1, 2), (2, 1), (3, 1), (3, 2), (4, 4), (5, 1), (5, 2), (7, 1), (7, 2)}
    const int support_table[7][4] = 
    {
        {1, 1, 0, 0},
        {1, 0, 0, 0},
        {1, 1, 0, 0},
        {0, 0, 0, 1},
        {1, 1, 0, 0},
        {0, 0, 0, 0},
        {1, 1, 0, 0}
    };
    // kernel_size x stride
    const int kernel = op->kernel_h,
              stride = op->stride_h;
    // if match prequisation
    switch(type)
    {
        case 1:
            // winograd
            if (kernel != 3 || stride != 1){
                return -1;
            }
            break;
        case 2:
            // pointwise
            // input_h == 1, input_w == 1, dilation == 1, stride == 1
            if (bottom.h != 1 || bottom.w != 1 || stride != 1)
            {
                return -1;
            }
            break;
        case 3:
            // im2col
            break;
        case 4:
            // direct conv 
            if (support_table[kernel-1][stride-1] == 0)
            {
                return -1;
            }
            break;
        case 5:
            // conv3x3s2
            // kernel == 3 and stride == 2
            if (kernel != 3 || stride != 2)
            {
                return -1;
            }
            break;
        default:
            fprintf(stderr, TEXT_RED "unrecognize convolution impl type: %d" CLR, type);
            break;
    }

    return 1;
}
#endif // defined(__aarch64__) && defined(LINUX)

int NetOptimize::fuse_batchnorm_scale()
{
    const size_t layer_count = layers.size();
    for (int i=0; i<layer_count; i++)
    {
        if (layers[i]->type != "BatchNorm")
            continue;

        // BatchNorm - Scale
        int top_blob_index = layers[i]->tops[0];

        int j = i + 1;
        for (; j<layer_count; j++)
        {
            if (layers[j]->type != "Scale")
                continue;

            if (layers[j]->bottoms.size() != 1)
                continue;

            if (layers[j]->bottoms[0] == top_blob_index)
                break;
        }

        if (j == layer_count)
            continue;

        // fuse BatchNorm - Scale to BatchNorm
        ncnn::BatchNorm* batchnorm = (ncnn::BatchNorm*)layers[i];
        ncnn::Scale* scale = (ncnn::Scale*)layers[j];

        fprintf(stderr, "fuse_batchnorm_scale %s %s\n", batchnorm->name.c_str(), scale->name.c_str());

        {
//             v = ((v - mean) / sqrt(var + eps) * slope + bias) * s + b
//               =  (v - mean) / sqrt(var + eps) * (slope * s) + (bias * s + b)

            int channels = batchnorm->channels;

            float* slope = batchnorm->slope_data;
            float* bias = batchnorm->bias_data;

            for (int q=0; q<channels; q++)
            {
                slope[q] = slope[q] * scale->scale_data[q];
                if (scale->bias_term)
                    bias[q] = bias[q] * scale->scale_data[q] + scale->bias_data[q];
                else
                    bias[q] = bias[q] * scale->scale_data[q];
            }
        }

        int top_blob_index_final = scale->tops[0];
        batchnorm->tops[0] = top_blob_index_final;
        blobs[top_blob_index_final].producer = i;
        scale->type = "ncnnfused";
    }

    return 0;
}

int NetOptimize::fuse_convolution_batchnorm()
{
    const size_t layer_count = layers.size();
    for (int i=0; i<layer_count; i++)
    {
        if (layers[i]->type != "Convolution")
            continue;

        // Convolution - BatchNorm
        int top_blob_index = layers[i]->tops[0];

        int j = i + 1;
        for (; j<layer_count; j++)
        {
            if (layers[j]->type != "BatchNorm")
                continue;

            if (layers[j]->bottoms.size() != 1)
                continue;

            if (layers[j]->bottoms[0] == top_blob_index)
                break;
        }

        if (j == layer_count)
            continue;

        // fuse Convolution - BatchNorm to Convolution
        ncnn::Convolution* convolution = (ncnn::Convolution*)layers[i];
        ncnn::BatchNorm* batchnorm = (ncnn::BatchNorm*)layers[j];

        fprintf(stderr, "fuse_convolution_batchnorm %s %s\n", convolution->name.c_str(), batchnorm->name.c_str());

        {
            int channels = batchnorm->channels;
            float eps = batchnorm->eps;

            // a = bias - slope * mean / sqrt(var + eps)
            // b = slope / sqrt(var + eps)
            // value = value * b + a

            std::vector<float> a(channels);
            std::vector<float> b(channels);
            for (int i=0; i<channels; i++)
            {
                float sqrt_var = static_cast<float>(sqrt(batchnorm->var_data[i] + eps));
                a[i] = batchnorm->bias_data[i] - batchnorm->slope_data[i] * batchnorm->mean_data[i] / sqrt_var;
                b[i] = batchnorm->slope_data[i] / sqrt_var;
            }

            if (convolution->bias_term == 0)
            {
                // init bias as zero
                convolution->bias_term = 1;
                convolution->bias_data = ncnn::Mat(channels);
                convolution->bias_data.fill(0.f);
            }

            const int weight_per_outch = convolution->weight_data_size / channels;

            float* weight = convolution->weight_data;
            float* bias = convolution->bias_data;
            for (int i=0; i<channels; i++)
            {
                float* conv_weight_outch = weight + weight_per_outch * i;
                for (int j=0; j<weight_per_outch; j++)
                {
                    conv_weight_outch[j] *= b[i];
                }

                bias[i] = bias[i] * b[i] + a[i];
            }
        }

        int top_blob_index_final = batchnorm->tops[0];
        convolution->tops[0] = top_blob_index_final;
        blobs[top_blob_index_final].producer = i;
        batchnorm->type = "ncnnfused";
    }

    return 0;
}

int NetOptimize::fuse_convolutiondepthwise_batchnorm()
{
    const size_t layer_count = layers.size();
    for (int i=0; i<layer_count; i++)
    {
        if (layers[i]->type != "ConvolutionDepthWise")
            continue;

        // ConvolutionDepthWise - BatchNorm
        int top_blob_index = layers[i]->tops[0];

        int j = i + 1;
        for (; j<layer_count; j++)
        {
            if (layers[j]->type != "BatchNorm")
                continue;

            if (layers[j]->bottoms.size() != 1)
                continue;

            if (layers[j]->bottoms[0] == top_blob_index)
                break;
        }

        if (j == layer_count)
            continue;

        // fuse ConvolutionDepthWise - BatchNorm to ConvolutionDepthWise
        ncnn::ConvolutionDepthWise* convolutiondepthwise = (ncnn::ConvolutionDepthWise*)layers[i];
        ncnn::BatchNorm* batchnorm = (ncnn::BatchNorm*)layers[j];

        fprintf(stderr, "fuse_convolutiondepthwise_batchnorm %s %s\n", convolutiondepthwise->name.c_str(), batchnorm->name.c_str());

        {
            int channels = batchnorm->channels;
            float eps = batchnorm->eps;

            // a = bias - slope * mean / sqrt(var + eps)
            // b = slope / sqrt(var + eps)
            // value = value * b + a

            std::vector<float> a(channels);
            std::vector<float> b(channels);
            for (int i=0; i<channels; i++)
            {
                float sqrt_var = static_cast<float>(sqrt(batchnorm->var_data[i] + eps));
                a[i] = batchnorm->bias_data[i] - batchnorm->slope_data[i] * batchnorm->mean_data[i] / sqrt_var;
                b[i] = batchnorm->slope_data[i] / sqrt_var;
            }

            if (convolutiondepthwise->bias_term == 0)
            {
                // init bias as zero
                convolutiondepthwise->bias_term = 1;
                convolutiondepthwise->bias_data = ncnn::Mat(channels);
                convolutiondepthwise->bias_data.fill(0.f);
            }

            const int weight_per_outch = convolutiondepthwise->weight_data_size / channels;

            float* weight = convolutiondepthwise->weight_data;
            float* bias = convolutiondepthwise->bias_data;
            for (int i=0; i<channels; i++)
            {
                float* conv_weight_outch = weight + weight_per_outch * i;
                for (int j=0; j<weight_per_outch; j++)
                {
                    conv_weight_outch[j] *= b[i];
                }

                bias[i] = bias[i] * b[i] + a[i];
            }
        }

        int top_blob_index_final = batchnorm->tops[0];
        convolutiondepthwise->tops[0] = top_blob_index_final;
        blobs[top_blob_index_final].producer = i;
        batchnorm->type = "ncnnfused";
    }

    return 0;
}

int NetOptimize::fuse_deconvolution_batchnorm()
{
    const size_t layer_count = layers.size();
    for (int i=0; i<layer_count; i++)
    {
        if (layers[i]->type != "Deconvolution")
            continue;

        // Deconvolution - BatchNorm
        int top_blob_index = layers[i]->tops[0];

        int j = i + 1;
        for (; j<layer_count; j++)
        {
            if (layers[j]->type != "BatchNorm")
                continue;

            if (layers[j]->bottoms.size() != 1)
                continue;

            if (layers[j]->bottoms[0] == top_blob_index)
                break;
        }

        if (j == layer_count)
            continue;

        // fuse Deconvolution - BatchNorm to Deconvolution
        ncnn::Deconvolution* deconvolution = (ncnn::Deconvolution*)layers[i];
        ncnn::BatchNorm* batchnorm = (ncnn::BatchNorm*)layers[j];

        fprintf(stderr, "fuse_deconvolution_batchnorm %s %s\n", deconvolution->name.c_str(), batchnorm->name.c_str());

        {
            int channels = batchnorm->channels;
            float eps = batchnorm->eps;

            // a = bias - slope * mean / sqrt(var + eps)
            // b = slope / sqrt(var + eps)
            // value = value * b + a

            std::vector<float> a(channels);
            std::vector<float> b(channels);
            for (int i=0; i<channels; i++)
            {
                float sqrt_var = static_cast<float>(sqrt(batchnorm->var_data[i] + eps));
                a[i] = batchnorm->bias_data[i] - batchnorm->slope_data[i] * batchnorm->mean_data[i] / sqrt_var;
                b[i] = batchnorm->slope_data[i] / sqrt_var;
            }

            if (deconvolution->bias_term == 0)
            {
                // init bias as zero
                deconvolution->bias_term = 1;
                deconvolution->bias_data = ncnn::Mat(channels);
                deconvolution->bias_data.fill(0.f);
            }

            const int weight_per_outch = deconvolution->weight_data_size / channels;

            float* weight = deconvolution->weight_data;
            float* bias = deconvolution->bias_data;
            for (int i=0; i<channels; i++)
            {
                float* conv_weight_outch = weight + weight_per_outch * i;
                for (int j=0; j<weight_per_outch; j++)
                {
                    conv_weight_outch[j] *= b[i];
                }

                bias[i] = bias[i] * b[i] + a[i];
            }
        }

        int top_blob_index_final = batchnorm->tops[0];
        deconvolution->tops[0] = top_blob_index_final;
        blobs[top_blob_index_final].producer = i;
        batchnorm->type = "ncnnfused";
    }

    return 0;
}

int NetOptimize::fuse_deconvolutiondepthwise_batchnorm()
{
    const size_t layer_count = layers.size();
    for (int i=0; i<layer_count; i++)
    {
        if (layers[i]->type != "DeconvolutionDepthWise")
            continue;

        // DeconvolutionDepthWise - BatchNorm
        int top_blob_index = layers[i]->tops[0];

        int j = i + 1;
        for (; j<layer_count; j++)
        {
            if (layers[j]->type != "BatchNorm")
                continue;

            if (layers[j]->bottoms.size() != 1)
                continue;

            if (layers[j]->bottoms[0] == top_blob_index)
                break;
        }

        if (j == layer_count)
            continue;

        // fuse DeconvolutionDepthWise - BatchNorm to DeconvolutionDepthWise
        ncnn::DeconvolutionDepthWise* deconvolutiondepthwise = (ncnn::DeconvolutionDepthWise*)layers[i];
        ncnn::BatchNorm* batchnorm = (ncnn::BatchNorm*)layers[j];

        fprintf(stderr, "fuse_deconvolutiondepthwise_batchnorm %s %s\n", deconvolutiondepthwise->name.c_str(), batchnorm->name.c_str());

        {
            int channels = batchnorm->channels;
            float eps = batchnorm->eps;

            // a = bias - slope * mean / sqrt(var + eps)
            // b = slope / sqrt(var + eps)
            // value = value * b + a

            std::vector<float> a(channels);
            std::vector<float> b(channels);
            for (int i=0; i<channels; i++)
            {
                float sqrt_var = static_cast<float>(sqrt(batchnorm->var_data[i] + eps));
                a[i] = batchnorm->bias_data[i] - batchnorm->slope_data[i] * batchnorm->mean_data[i] / sqrt_var;
                b[i] = batchnorm->slope_data[i] / sqrt_var;
            }

            if (deconvolutiondepthwise->bias_term == 0)
            {
                // init bias as zero
                deconvolutiondepthwise->bias_term = 1;
                deconvolutiondepthwise->bias_data = ncnn::Mat(channels);
                deconvolutiondepthwise->bias_data.fill(0.f);
            }

            const int weight_per_outch = deconvolutiondepthwise->weight_data_size / channels;

            float* weight = deconvolutiondepthwise->weight_data;
            float* bias = deconvolutiondepthwise->bias_data;
            for (int i=0; i<channels; i++)
            {
                float* conv_weight_outch = weight + weight_per_outch * i;
                for (int j=0; j<weight_per_outch; j++)
                {
                    conv_weight_outch[j] *= b[i];
                }

                bias[i] = bias[i] * b[i] + a[i];
            }
        }

        int top_blob_index_final = batchnorm->tops[0];
        deconvolutiondepthwise->tops[0] = top_blob_index_final;
        blobs[top_blob_index_final].producer = i;
        batchnorm->type = "ncnnfused";
    }

    return 0;
}

int NetOptimize::fuse_innerproduct_batchnorm()
{
    const size_t layer_count = layers.size();
    for (int i=0; i<layer_count; i++)
    {
        if (layers[i]->type != "InnerProduct")
            continue;

        // InnerProduct - BatchNorm
        int top_blob_index = layers[i]->tops[0];

        int j = i + 1;
        for (; j<layer_count; j++)
        {
            if (layers[j]->type != "BatchNorm")
                continue;

            if (layers[j]->bottoms.size() != 1)
                continue;

            if (layers[j]->bottoms[0] == top_blob_index)
                break;
        }

        if (j == layer_count)
            continue;

        // fuse InnerProduct - BatchNorm to InnerProduct
        ncnn::InnerProduct* innerproduct = (ncnn::InnerProduct*)layers[i];
        ncnn::BatchNorm* batchnorm = (ncnn::BatchNorm*)layers[j];

        fprintf(stderr, "fuse_innerproduct_batchnorm %s %s\n", innerproduct->name.c_str(), batchnorm->name.c_str());

        {
            int channels = batchnorm->channels;
            float eps = batchnorm->eps;

            // a = bias - slope * mean / sqrt(var + eps)
            // b = slope / sqrt(var + eps)
            // value = value * b + a

            std::vector<float> a(channels);
            std::vector<float> b(channels);
            for (int i=0; i<channels; i++)
            {
                float sqrt_var = static_cast<float>(sqrt(batchnorm->var_data[i] + eps));
                a[i] = batchnorm->bias_data[i] - batchnorm->slope_data[i] * batchnorm->mean_data[i] / sqrt_var;
                b[i] = batchnorm->slope_data[i] / sqrt_var;
            }

            if (innerproduct->bias_term == 0)
            {
                // init bias as zero
                innerproduct->bias_term = 1;
                innerproduct->bias_data = ncnn::Mat(channels);
                innerproduct->bias_data.fill(0.f);
            }

            const int weight_per_outch = innerproduct->weight_data_size / channels;

            float* weight = innerproduct->weight_data;
            float* bias = innerproduct->bias_data;
            for (int i=0; i<channels; i++)
            {
                float* conv_weight_outch = weight + weight_per_outch * i;
                for (int j=0; j<weight_per_outch; j++)
                {
                    conv_weight_outch[j] *= b[i];
                }

                bias[i] = bias[i] * b[i] + a[i];
            }
        }

        int top_blob_index_final = batchnorm->tops[0];
        innerproduct->tops[0] = top_blob_index_final;
        blobs[top_blob_index_final].producer = i;
        batchnorm->type = "ncnnfused";
    }

    return 0;
}

int NetOptimize::fuse_innerproduct_dropout()
{
    const size_t layer_count = layers.size();
    for (int i=0; i<layer_count; i++)
    {
        if (layers[i]->type != "InnerProduct")
            continue;

        // InnerProduct - Dropout
        int top_blob_index = layers[i]->tops[0];

        int j = i + 1;
        for (; j<layer_count; j++)
        {
            if (layers[j]->type != "Dropout")
                continue;

            if (layers[j]->bottoms.size() != 1)
                continue;

            if (layers[j]->bottoms[0] == top_blob_index)
                break;
        }

        if (j == layer_count)
            continue;

        // fuse InnerProduct - Dropout to InnerProduct
        ncnn::InnerProduct* innerproduct = (ncnn::InnerProduct*)layers[i];
        ncnn::Dropout* dropout = (ncnn::Dropout*)layers[j];

        fprintf(stderr, "fuse_innerproduct_dropout %s %s\n", innerproduct->name.c_str(), dropout->name.c_str());

        float scale = dropout->scale;
        if (scale != 1.f)
        {
            const int num_output = innerproduct->num_output;
            const int weight_per_outch = innerproduct->weight_data_size / num_output;

            float* weight = innerproduct->weight_data;
            for (int i=0; i<num_output; i++)
            {
                float* conv_weight_outch = weight + weight_per_outch * i;
                for (int j=0; j<weight_per_outch; j++)
                {
                    conv_weight_outch[j] *= scale;
                }
            }

            if (innerproduct->bias_term)
            {
                float* bias = innerproduct->bias_data;
                for (int i=0; i<num_output; i++)
                {
                    bias[i] *= scale;
                }
            }
        }

        int top_blob_index_final = dropout->tops[0];
        innerproduct->tops[0] = top_blob_index_final;
        blobs[top_blob_index_final].producer = i;
        dropout->type = "ncnnfused";
    }

    return 0;
}

int NetOptimize::fuse_convolution_activation()
{
    const size_t layer_count = layers.size();
    for (int i=0; i<layer_count; i++)
    {
        if (layers[i]->type != "Convolution")
            continue;

        // Convolution - Activation
        int top_blob_index = layers[i]->tops[0];

        int j = i + 1;
        for (; j<layer_count; j++)
        {
            if (layers[j]->type != "ReLU" && layers[j]->type != "Clip" && layers[j]->type != "Sigmoid")
                continue;

            if (layers[j]->bottoms.size() != 1)
                continue;

            if (layers[j]->bottoms[0] == top_blob_index)
                break;
        }

        if (j == layer_count)
            continue;

        // fuse Convolution - Activation to Convolution
        ncnn::Convolution* convolution = (ncnn::Convolution*)layers[i];
        ncnn::Layer* activation = layers[j];

        fprintf(stderr, "fuse_convolution_activation %s %s\n", convolution->name.c_str(), activation->name.c_str());

        if (activation->type == "ReLU")
        {
            ncnn::ReLU* relu = (ncnn::ReLU*)activation;

            if (relu->slope == 0.f)
            {
                convolution->activation_type = 1;
            }
            else
            {
                convolution->activation_type = 2;
                convolution->activation_params = ncnn::Mat(1);
                convolution->activation_params[0] = relu->slope;
            }
        }
        else if (activation->type == "Clip")
        {
            ncnn::Clip* clip = (ncnn::Clip*)activation;

            convolution->activation_type = 3;
            convolution->activation_params = ncnn::Mat(2);
            convolution->activation_params[0] = clip->min;
            convolution->activation_params[1] = clip->max;
        }
        else if (activation->type == "Sigmoid")
        {
            convolution->activation_type = 4;
        }

        int top_blob_index_final = activation->tops[0];
        convolution->tops[0] = top_blob_index_final;
        blobs[top_blob_index_final].producer = i;
        activation->type = "ncnnfused";
    }

    return 0;
}

int NetOptimize::fuse_convolutiondepthwise_activation()
{
    const size_t layer_count = layers.size();
    for (int i=0; i<layer_count; i++)
    {
        if (layers[i]->type != "ConvolutionDepthWise")
            continue;

        // ConvolutionDepthWise - Activation
        int top_blob_index = layers[i]->tops[0];

        int j = i + 1;
        for (; j<layer_count; j++)
        {
            if (layers[j]->type != "ReLU" && layers[j]->type != "Clip" && layers[j]->type != "Sigmoid")
                continue;

            if (layers[j]->bottoms.size() != 1)
                continue;

            if (layers[j]->bottoms[0] == top_blob_index)
                break;
        }

        if (j == layer_count)
            continue;

        // fuse ConvolutionDepthWise - Activation to ConvolutionDepthWise
        ncnn::ConvolutionDepthWise* convolutiondepthwise = (ncnn::ConvolutionDepthWise*)layers[i];
        ncnn::Layer* activation = layers[j];

        fprintf(stderr, "fuse_convolutiondepthwise_activation %s %s\n", convolutiondepthwise->name.c_str(), activation->name.c_str());

        if (activation->type == "ReLU")
        {
            ncnn::ReLU* relu = (ncnn::ReLU*)activation;

            if (relu->slope == 0.f)
            {
                convolutiondepthwise->activation_type = 1;
            }
            else
            {
                convolutiondepthwise->activation_type = 2;
                convolutiondepthwise->activation_params = ncnn::Mat(1);
                convolutiondepthwise->activation_params[0] = relu->slope;
            }
        }
        else if (activation->type == "Clip")
        {
            ncnn::Clip* clip = (ncnn::Clip*)activation;

            convolutiondepthwise->activation_type = 3;
            convolutiondepthwise->activation_params = ncnn::Mat(2);
            convolutiondepthwise->activation_params[0] = clip->min;
            convolutiondepthwise->activation_params[1] = clip->max;
        }
        else if (activation->type == "Sigmoid")
        {
            convolutiondepthwise->activation_type = 4;
        }

        int top_blob_index_final = activation->tops[0];
        convolutiondepthwise->tops[0] = top_blob_index_final;
        blobs[top_blob_index_final].producer = i;
        activation->type = "ncnnfused";
    }

    return 0;
}

int NetOptimize::fuse_deconvolution_activation()
{
    const size_t layer_count = layers.size();
    for (int i=0; i<layer_count; i++)
    {
        if (layers[i]->type != "Deconvolution")
            continue;

        // Deconvolution - Activation
        int top_blob_index = layers[i]->tops[0];

        int j = i + 1;
        for (; j<layer_count; j++)
        {
            if (layers[j]->type != "ReLU" && layers[j]->type != "Clip" && layers[j]->type != "Sigmoid")
                continue;

            if (layers[j]->bottoms.size() != 1)
                continue;

            if (layers[j]->bottoms[0] == top_blob_index)
                break;
        }

        if (j == layer_count)
            continue;

        // fuse Deconvolution - Activation to Deconvolution
        ncnn::Deconvolution* deconvolution = (ncnn::Deconvolution*)layers[i];
        ncnn::Layer* activation = layers[j];

        fprintf(stderr, "fuse_deconvolution_activation %s %s\n", deconvolution->name.c_str(), activation->name.c_str());

        if (activation->type == "ReLU")
        {
            ncnn::ReLU* relu = (ncnn::ReLU*)activation;

            if (relu->slope == 0.f)
            {
                deconvolution->activation_type = 1;
            }
            else
            {
                deconvolution->activation_type = 2;
                deconvolution->activation_params = ncnn::Mat(1);
                deconvolution->activation_params[0] = relu->slope;
            }
        }
        else if (activation->type == "Clip")
        {
            ncnn::Clip* clip = (ncnn::Clip*)activation;

            deconvolution->activation_type = 3;
            deconvolution->activation_params = ncnn::Mat(2);
            deconvolution->activation_params[0] = clip->min;
            deconvolution->activation_params[1] = clip->max;
        }
        else if (activation->type == "Sigmoid")
        {
            deconvolution->activation_type = 4;
        }

        int top_blob_index_final = activation->tops[0];
        deconvolution->tops[0] = top_blob_index_final;
        blobs[top_blob_index_final].producer = i;
        activation->type = "ncnnfused";
    }

    return 0;
}

int NetOptimize::fuse_deconvolutiondepthwise_activation()
{
    const size_t layer_count = layers.size();
    for (int i=0; i<layer_count; i++)
    {
        if (layers[i]->type != "DeconvolutionDepthWise")
            continue;

        // DeconvolutionDepthWise - Activation
        int top_blob_index = layers[i]->tops[0];

        int j = i + 1;
        for (; j<layer_count; j++)
        {
            if (layers[j]->type != "ReLU" && layers[j]->type != "Clip" && layers[j]->type != "Sigmoid")
                continue;

            if (layers[j]->bottoms.size() != 1)
                continue;

            if (layers[j]->bottoms[0] == top_blob_index)
                break;
        }

        if (j == layer_count)
            continue;

        // fuse DeconvolutionDepthWise - Activation to DeconvolutionDepthWise
        ncnn::DeconvolutionDepthWise* deconvolutiondepthwise = (ncnn::DeconvolutionDepthWise*)layers[i];
        ncnn::Layer* activation = layers[j];

        fprintf(stderr, "fuse_deconvolutiondepthwise_activation %s %s\n", deconvolutiondepthwise->name.c_str(), activation->name.c_str());

        if (activation->type == "ReLU")
        {
            ncnn::ReLU* relu = (ncnn::ReLU*)activation;

            if (relu->slope == 0.f)
            {
                deconvolutiondepthwise->activation_type = 1;
            }
            else
            {
                deconvolutiondepthwise->activation_type = 2;
                deconvolutiondepthwise->activation_params = ncnn::Mat(1);
                deconvolutiondepthwise->activation_params[0] = relu->slope;
            }
        }
        else if (activation->type == "Clip")
        {
            ncnn::Clip* clip = (ncnn::Clip*)activation;

            deconvolutiondepthwise->activation_type = 3;
            deconvolutiondepthwise->activation_params = ncnn::Mat(2);
            deconvolutiondepthwise->activation_params[0] = clip->min;
            deconvolutiondepthwise->activation_params[1] = clip->max;
        }
        else if (activation->type == "Sigmoid")
        {
            deconvolutiondepthwise->activation_type = 4;
        }

        int top_blob_index_final = activation->tops[0];
        deconvolutiondepthwise->tops[0] = top_blob_index_final;
        blobs[top_blob_index_final].producer = i;
        activation->type = "ncnnfused";
    }

    return 0;
}

int NetOptimize::fuse_innerproduct_activation()
{
    const size_t layer_count = layers.size();
    for (int i=0; i<layer_count; i++)
    {
        if (layers[i]->type != "InnerProduct")
            continue;

        // InnerProduct - Activation
        int top_blob_index = layers[i]->tops[0];

        int j = i + 1;
        for (; j<layer_count; j++)
        {
            if (layers[j]->type != "ReLU" && layers[j]->type != "Clip" && layers[j]->type != "Sigmoid")
                continue;

            if (layers[j]->bottoms.size() != 1)
                continue;

            if (layers[j]->bottoms[0] == top_blob_index)
                break;
        }

        if (j == layer_count)
            continue;

        // fuse InnerProduct - Activation to InnerProduct
        ncnn::InnerProduct* innerproduct = (ncnn::InnerProduct*)layers[i];
        ncnn::Layer* activation = layers[j];

        fprintf(stderr, "fuse_innerproduct_activation %s %s\n", innerproduct->name.c_str(), activation->name.c_str());

        if (activation->type == "ReLU")
        {
            ncnn::ReLU* relu = (ncnn::ReLU*)activation;

            if (relu->slope == 0.f)
            {
                innerproduct->activation_type = 1;
            }
            else
            {
                innerproduct->activation_type = 2;
                innerproduct->activation_params = ncnn::Mat(1);
                innerproduct->activation_params[0] = relu->slope;
            }
        }
        else if (activation->type == "Clip")
        {
            ncnn::Clip* clip = (ncnn::Clip*)activation;

            innerproduct->activation_type = 3;
            innerproduct->activation_params = ncnn::Mat(2);
            innerproduct->activation_params[0] = clip->min;
            innerproduct->activation_params[1] = clip->max;
        }
        else if (activation->type == "Sigmoid")
        {
            innerproduct->activation_type = 4;
        }

        int top_blob_index_final = activation->tops[0];
        innerproduct->tops[0] = top_blob_index_final;
        blobs[top_blob_index_final].producer = i;
        activation->type = "ncnnfused";
    }

    return 0;
}

int NetOptimize::eliminate_dropout()
{
    const size_t layer_count = layers.size();
    for (int i=0; i<layer_count; i++)
    {
        if (layers[i]->type != "Dropout")
            continue;

        ncnn::Dropout* dropout = (ncnn::Dropout*)layers[i];
        if (dropout->scale != 1.f)
            continue;

        // Any - Dropout
        int bottom_blob_index = layers[i]->bottoms[0];

        int j = i - 1;
        for (; j>=0; j--)
        {
            if (layers[j]->type == "ncnnfused")
                continue;

            if (layers[j]->tops.size() != 1)
                continue;

            if (layers[j]->tops[0] == bottom_blob_index)
                break;
        }

        if (j == -1)
            continue;

        ncnn::Layer* any = layers[j];

        fprintf(stderr, "eliminate_dropout %s %s\n", any->name.c_str(), dropout->name.c_str());

        int top_blob_index_final = dropout->tops[0];
        any->tops[0] = top_blob_index_final;
        blobs[top_blob_index_final].producer = j;
        dropout->type = "ncnnfused";
    }

    return 0;
}

int NetOptimize::eliminate_pooling1x1()
{
    const size_t layer_count = layers.size();
    for (int i=0; i<layer_count; i++)
    {
        if (layers[i]->type != "Pooling")
            continue;

        ncnn::Pooling* pooling = (ncnn::Pooling*)layers[i];
        if (pooling->pad_left != 0 || pooling->pad_right != 0 || pooling->pad_top != 0 || pooling->pad_bottom != 0)
            continue;

        if (pooling->kernel_w != 1 || pooling->kernel_h != 1 || pooling->stride_w != 1 || pooling->stride_h != 1)
            continue;

        if (pooling->global_pooling != 0)
            continue;

        // Any - Pooling
        int bottom_blob_index = layers[i]->bottoms[0];

        int top_i = -1;
        int j = i - 1;
        for (; j>=0; j--)
        {
            if (layers[j]->type == "ncnnfused")
                continue;

            for (int k=0; k<layers[j]->tops.size(); k++)
            {
                if (layers[j]->tops[k] == bottom_blob_index)
                {
                    top_i = k;
                    break;
                }
            }

            if (top_i != -1)
                break;
        }

        if (j == -1)
            continue;

        ncnn::Layer* any = layers[j];

        fprintf(stderr, "eliminate_pooling1x1 %s %s\n", any->name.c_str(), pooling->name.c_str());

        int top_blob_index_final = pooling->tops[0];
        any->tops[top_i] = top_blob_index_final;
        blobs[top_blob_index_final].producer = j;
        pooling->type = "ncnnfused";
    }

    return 0;
}

int NetOptimize::eliminate_noop()
{
    const size_t layer_count = layers.size();
    for (int i=0; i<layer_count; i++)
    {
        if (layers[i]->type != "Noop")
            continue;

        ncnn::Layer* noop = layers[i];

        if (noop->bottoms.empty())
        {
            // Noop
            fprintf(stderr, "eliminate_noop %s\n", noop->name.c_str());

            size_t top_blob_count = noop->tops.size();
            for (int k=0; k<top_blob_count; k++)
            {
                int top_blob_index_final = noop->tops[k];
                blobs[top_blob_index_final].producer = -1;
            }
            noop->type = "ncnnfused";

            continue;
        }

        // Any - Noop
        int bottom_blob_index = layers[i]->bottoms[0];

        int j = i - 1;
        for (; j>=0; j--)
        {
            if (layers[j]->type == "ncnnfused")
                continue;

            if (layers[j]->tops.size() != 1)
                continue;

            if (layers[j]->tops[0] == bottom_blob_index)
                break;
        }

        if (j == -1)
            continue;

        ncnn::Layer* any = layers[j];

        fprintf(stderr, "eliminate_noop %s %s\n", any->name.c_str(), noop->name.c_str());

        size_t top_blob_count = std::min(noop->tops.size(), any->tops.size());
        for (int k=0; k<top_blob_count; k++)
        {
            int top_blob_index_final = noop->tops[k];
            any->tops[k] = top_blob_index_final;
            blobs[top_blob_index_final].producer = j;
        }
        noop->type = "ncnnfused";
    }

    return 0;
}

int NetOptimize::eliminate_orphaned_memorydata()
{
    const size_t layer_count = layers.size();
    for (int i=0; i<layer_count; i++)
    {
        if (layers[i]->type != "MemoryData")
            continue;

        // MemoryData - X
        int top_blob_index = layers[i]->tops[0];

        int j = i + 1;
        for (; j<layer_count; j++)
        {
            if (layers[j]->type == "ncnnfused")
                continue;

            bool orphaned = true;
            for (int k=0; k<layers[j]->bottoms.size(); k++)
            {
                if (layers[j]->bottoms[k] == top_blob_index)
                {
                    orphaned = false;
                    break;
                }
            }

            if (!orphaned)
                break;
        }

        if (j < layer_count)
            continue;

        // assert orphaned == true
        fprintf(stderr, "eliminate_orphaned_memorydata %s\n", layers[i]->name.c_str());

        layers[i]->type = "ncnnfused";
    }

    return 0;
}

int NetOptimize::eliminate_reshape_after_global_pooling()
{
    const size_t layer_count = layers.size();
    for (int i=0; i<layer_count; i++)
    {
        if (layers[i]->type != "Pooling")
            continue;

        ncnn::Pooling* pooling = (ncnn::Pooling*)layers[i];
        if (pooling->global_pooling == 0)
            continue;

        // Pooling - Reshape
        int top_blob_index = layers[i]->tops[0];

        int j = i + 1;
        for (; j<layer_count; j++)
        {
            if (layers[j]->type != "Reshape")
                continue;

            if (layers[j]->bottoms.size() != 1)
                continue;

            if (layers[j]->bottoms[0] == top_blob_index)
                break;
        }

        if (j == layer_count)
            continue;

        ncnn::Reshape* reshape = (ncnn::Reshape*)layers[j];
        if (reshape->h != -233 || reshape->c != -233 || reshape->permute != 0)
            continue;

        fprintf(stderr, "eliminate_reshape_after_global_pooling %s %s\n", pooling->name.c_str(), reshape->name.c_str());

        int top_blob_index_final = reshape->tops[0];
        pooling->tops[0] = top_blob_index_final;
        blobs[top_blob_index_final].producer = i;
        reshape->type = "ncnnfused";
    }

    return 0;
}

int NetOptimize::eliminate_flatten_after_global_pooling()
{
    const size_t layer_count = layers.size();
    for (int i=0; i<layer_count; i++)
    {
        if (layers[i]->type != "Pooling")
            continue;

        ncnn::Pooling* pooling = (ncnn::Pooling*)layers[i];
        if (pooling->global_pooling == 0)
            continue;

        // Pooling - Flatten
        int top_blob_index = layers[i]->tops[0];

        int j = i + 1;
        for (; j<layer_count; j++)
        {
            if (layers[j]->type != "Flatten")
                continue;

            if (layers[j]->bottoms.size() != 1)
                continue;

            if (layers[j]->bottoms[0] == top_blob_index)
                break;
        }

        if (j == layer_count)
            continue;

        ncnn::Flatten* flatten = (ncnn::Flatten*)layers[j];

        fprintf(stderr, "eliminate_flatten_after_global_pooling %s %s\n", pooling->name.c_str(), flatten->name.c_str());

        int top_blob_index_final = flatten->tops[0];
        pooling->tops[0] = top_blob_index_final;
        blobs[top_blob_index_final].producer = i;
        flatten->type = "ncnnfused";
    }

    return 0;
}

int NetOptimize::eliminate_flatten_after_innerproduct()
{
    const size_t layer_count = layers.size();
    for (int i=0; i<layer_count; i++)
    {
        if (layers[i]->type != "InnerProduct")
            continue;

        // InnerProduct - Flatten
        int top_blob_index = layers[i]->tops[0];

        int j = i + 1;
        for (; j<layer_count; j++)
        {
            if (layers[j]->type != "Flatten")
                continue;

            if (layers[j]->bottoms.size() != 1)
                continue;

            if (layers[j]->bottoms[0] == top_blob_index)
                break;
        }

        if (j == layer_count)
            continue;

        ncnn::InnerProduct* innerproduct = (ncnn::InnerProduct*)layers[i];
        ncnn::Flatten* flatten = (ncnn::Flatten*)layers[j];

        fprintf(stderr, "eliminate_flatten_after_innerproduct %s %s\n", innerproduct->name.c_str(), flatten->name.c_str());

        int top_blob_index_final = flatten->tops[0];
        innerproduct->tops[0] = top_blob_index_final;
        blobs[top_blob_index_final].producer = i;
        flatten->type = "ncnnfused";
    }

    return 0;
}

int NetOptimize::eliminate_reshape_before_binaryop()
{
    const size_t layer_count = layers.size();
    for (int i=0; i<layer_count; i++)
    {
        if (layers[i]->type != "Reshape")
            continue;

        ncnn::Reshape* reshape = (ncnn::Reshape*)layers[i];
        if (reshape->w != 1 || reshape->h != 1 || reshape->permute != 0)
            continue;

        // Reshape - BinaryOp
        int top_blob_index = layers[i]->tops[0];

        int j = i + 1;
        for (; j<layer_count; j++)
        {
            if (layers[j]->type != "BinaryOp")
                continue;

            if (layers[j]->bottoms.size() != 2)
                continue;

            if (layers[j]->bottoms[0] == top_blob_index || layers[j]->bottoms[1] == top_blob_index)
                break;
        }

        if (j == layer_count)
            continue;

        ncnn::BinaryOp* binaryop = (ncnn::BinaryOp*)layers[j];

        fprintf(stderr, "eliminate_reshape_before_binaryop %s %s\n", reshape->name.c_str(), binaryop->name.c_str());

        int bottom_blob_index_final = reshape->bottoms[0];
        if (layers[j]->bottoms[0] == top_blob_index)
            binaryop->bottoms[0] = bottom_blob_index_final;
        if (layers[j]->bottoms[1] == top_blob_index)
            binaryop->bottoms[1] = bottom_blob_index_final;
        blobs[bottom_blob_index_final].consumers.erase(std::find(blobs[bottom_blob_index_final].consumers.begin(), blobs[bottom_blob_index_final].consumers.end(), i));
        blobs[bottom_blob_index_final].consumers.push_back(j);
        reshape->type = "ncnnfused";
    }

    return 0;
}

int NetOptimize::replace_convolution_with_innerproduct_after_global_pooling()
{
    const size_t layer_count = layers.size();
    for (int i=0; i<layer_count; i++)
    {
        if (layers[i]->type != "Pooling")
            continue;

        ncnn::Pooling* pooling = (ncnn::Pooling*)layers[i];
        if (pooling->global_pooling == 0)
            continue;

        // Pooling - Convolution
        int top_blob_index = layers[i]->tops[0];

        int j = i + 1;
        for (; j<layer_count; j++)
        {
            if (layers[j]->type != "Convolution")
                continue;

            if (layers[j]->bottoms.size() != 1)
                continue;

            if (layers[j]->bottoms[0] == top_blob_index)
                break;
        }

        if (j == layer_count)
            continue;

        ncnn::Convolution* convolution = (ncnn::Convolution*)layers[j];

        fprintf(stderr, "replace_convolution_with_innerproduct_after_global_pooling %s %s\n", pooling->name.c_str(), convolution->name.c_str());

        ncnn::InnerProduct* innerproduct = (ncnn::InnerProduct*)ncnn::create_layer("InnerProduct");

        innerproduct->type = "InnerProduct";
        innerproduct->name = convolution->name;
        innerproduct->bottoms = convolution->bottoms;
        innerproduct->tops = convolution->tops;

        ncnn::ParamDict pd;
        innerproduct->load_param(pd);

        innerproduct->num_output = convolution->num_output;
        innerproduct->bias_term = convolution->bias_term;
        innerproduct->weight_data_size = convolution->weight_data_size;

        innerproduct->weight_data = convolution->weight_data;
        innerproduct->bias_data = convolution->bias_data;

        innerproduct->activation_type = convolution->activation_type;
        innerproduct->activation_params = convolution->activation_params;

        layers[j] = innerproduct;
        delete convolution;
    }

    return 0;
}

int NetOptimize::replace_convolution_with_innerproduct_after_innerproduct()
{
    const size_t layer_count = layers.size();
    for (;;)
    {
    bool replaced = false;

    for (int i=0; i<layer_count; i++)
    {
        if (layers[i]->type != "InnerProduct")
            continue;

        // InnerProduct - Convolution
        int top_blob_index = layers[i]->tops[0];

        int j = i + 1;
        for (; j<layer_count; j++)
        {
            if (layers[j]->type != "Convolution")
                continue;

            if (layers[j]->bottoms.size() != 1)
                continue;

            if (layers[j]->bottoms[0] == top_blob_index)
                break;
        }

        if (j == layer_count)
            continue;

        ncnn::InnerProduct* innerproduct = (ncnn::InnerProduct*)layers[i];
        ncnn::Convolution* convolution = (ncnn::Convolution*)layers[j];

        fprintf(stderr, "replace_convolution_with_innerproduct_after_innerproduct %s %s\n", innerproduct->name.c_str(), convolution->name.c_str());

        ncnn::InnerProduct* innerproduct2 = (ncnn::InnerProduct*)ncnn::create_layer("InnerProduct");

        innerproduct2->type = "InnerProduct";
        innerproduct2->name = convolution->name;
        innerproduct2->bottoms = convolution->bottoms;
        innerproduct2->tops = convolution->tops;

        ncnn::ParamDict pd;
        innerproduct2->load_param(pd);

        innerproduct2->num_output = convolution->num_output;
        innerproduct2->bias_term = convolution->bias_term;
        innerproduct2->weight_data_size = convolution->weight_data_size;

        innerproduct2->weight_data = convolution->weight_data;
        innerproduct2->bias_data = convolution->bias_data;

        innerproduct2->activation_type = convolution->activation_type;
        innerproduct2->activation_params = convolution->activation_params;

        layers[j] = innerproduct2;
        delete convolution;

        replaced = true;
    }

    if (!replaced)
        break;
    }

    return 0;
}

int NetOptimize::fprintf_param_int_array(int id, const ncnn::Mat& m, FILE* pp)
{
    const int count = m.w;
    const int* ptr = m;

    fprintf(pp, " -%d=%d", 23300 + id, count);
    for (int i=0; i<count; i++)
    {
        fprintf(pp, ",%d", ptr[i]);
    }

    return 0;
}

int NetOptimize::fprintf_param_float_array(int id, const ncnn::Mat& m, FILE* pp)
{
    const int count = m.w;
    const float* ptr = m;

    fprintf(pp, " -%d=%d", 23300 + id, count);
    for (int i=0; i<count; i++)
    {
        fprintf(pp, ",%e", ptr[i]);
    }

    return 0;
}

static inline size_t alignSize(size_t sz, int n)
{
    return (sz + n-1) & -n;
}

int NetOptimize::fwrite_weight_tag_data(int tag, const ncnn::Mat& data, FILE* bp)
{
    int p0 = ftell(bp);

    ncnn::Mat data_flattened = data.reshape(data.w * data.h * data.c);
    if (storage_type == 1 && tag == 0)
    {
        tag = 0x01306B47; // fp16 magic
        fwrite(&tag, sizeof(int), 1, bp);
        ncnn::Mat data_flattened_fp16;
        ncnn::cast_float32_to_float16(data_flattened, data_flattened_fp16);
        fwrite(data_flattened_fp16.data, data_flattened_fp16.elemsize, data_flattened_fp16.w, bp);
    }
    else
    {
        fwrite(&tag, sizeof(int), 1, bp);
        fwrite(data_flattened.data, data_flattened.elemsize, data_flattened.w, bp);
    }

    // padding to 32bit align
    int nwrite = ftell(bp) - p0;
    size_t nalign = alignSize(nwrite, 4);
    unsigned char padding[4] = {0x00, 0x00, 0x00, 0x00};
    fwrite(padding, sizeof(unsigned char), nalign - nwrite, bp);

    return 0;
}

int NetOptimize::fwrite_weight_data(const ncnn::Mat& data, FILE* bp)
{
    int p0 = ftell(bp);

    ncnn::Mat data_flattened = data.reshape(data.w * data.h * data.c);
    fwrite(data_flattened.data, data_flattened.elemsize, data_flattened.w, bp);

    // padding to 32bit align
    int nwrite = ftell(bp) - p0;
    size_t nalign = alignSize(nwrite, 4);
    unsigned char padding[4] = {0x00, 0x00, 0x00, 0x00};
    fwrite(padding, sizeof(unsigned char), nalign - nwrite, bp);

    return 0;
}

int NetOptimize::save(const char* parampath, const char* binpath)
{
    FILE* pp = fopen(parampath, "wb");
    FILE* bp = fopen(binpath, "wb");

    fprintf(pp, "7767517\n");

    const size_t layer_count = layers.size();

    int layer_count_fused = 0;
    std::set<std::string> blob_names;
    for (int i=0; i<layer_count; i++)
    {
        const ncnn::Layer* layer = layers[i];
        if (layer->type == "ncnnfused")
            continue;

        layer_count_fused++;

        size_t bottom_count = layer->bottoms.size();
        for (int j=0; j<bottom_count; j++)
        {
            int bottom_blob_index = layer->bottoms[j];
            blob_names.insert(blobs[bottom_blob_index].name);
        }

        size_t top_count = layer->tops.size();
        for (int j=0; j<top_count; j++)
        {
            int top_blob_index = layer->tops[j];
            blob_names.insert(blobs[top_blob_index].name);
        }
    }

    size_t blob_count_fused = blob_names.size();

    fprintf(pp, "%d %zd\n", layer_count_fused, blob_count_fused);

    for (int i=0; i<layer_count; i++)
    {
        const ncnn::Layer* layer = layers[i];
        if (layer->type == "ncnnfused")
            continue;

        size_t bottom_count = layer->bottoms.size();
        size_t top_count = layer->tops.size();

        fprintf(pp, "%-24s %-24s %zd %zd", layer->type.c_str(), layer->name.c_str(), bottom_count, top_count);

        for (int j=0; j<bottom_count; j++)
        {
            int bottom_blob_index = layer->bottoms[j];
            fprintf(pp, " %s", blobs[bottom_blob_index].name.c_str());
        }
        for (int j=0; j<top_count; j++)
        {
            int top_blob_index = layer->tops[j];
            fprintf(pp, " %s", blobs[top_blob_index].name.c_str());
        }

        ncnn::Layer* layer_default = ncnn::create_layer(layer->typeindex);

        ncnn::ParamDict pd;
        layer_default->load_param(pd);

#define fprintf_param_value(format, phase) \
        { if (op->phase != op_default->phase) fprintf(pp, format, op->phase); }

        if (layer->type == "BatchNorm")
        {
            ncnn::BatchNorm* op = (ncnn::BatchNorm*)layer;
            ncnn::BatchNorm* op_default = (ncnn::BatchNorm*)layer_default;

            fprintf_param_value(" 0=%d", channels)
            fprintf_param_value(" 1=%e", eps)

            fwrite_weight_data(op->slope_data, bp);
            fwrite_weight_data(op->mean_data, bp);
            fwrite_weight_data(op->var_data, bp);
            fwrite_weight_data(op->bias_data, bp);
        }
        else if (layer->type == "Bias")
        {
            ncnn::Bias* op = (ncnn::Bias*)layer;
            ncnn::Bias* op_default = (ncnn::Bias*)layer_default;

            fprintf_param_value(" 0=%d", bias_data_size)

            fwrite_weight_data(op->bias_data, bp);
        }
        else if (layer->type == "BinaryOp")
        {
            ncnn::BinaryOp* op = (ncnn::BinaryOp*)layer;
            ncnn::BinaryOp* op_default = (ncnn::BinaryOp*)layer_default;

            fprintf_param_value(" 0=%d", op_type)
            fprintf_param_value(" 1=%d", with_scalar)
            fprintf_param_value(" 2=%e", b)
        }
        else if (layer->type == "Clip")
        {
            ncnn::Clip* op = (ncnn::Clip*)layer;
            ncnn::Clip* op_default = (ncnn::Clip*)layer_default;

            fprintf_param_value(" 0=%e", min)
            fprintf_param_value(" 1=%e", max)
        }
        else if (layer->type == "Concat")
        {
            ncnn::Concat* op = (ncnn::Concat*)layer;
            ncnn::Concat* op_default = (ncnn::Concat*)layer_default;

            fprintf_param_value(" 0=%d", axis)
        }
        else if (layer->type == "Convolution")
        {
            ncnn::Convolution* op = (ncnn::Convolution*)layer;
            ncnn::Convolution* op_default = (ncnn::Convolution*)layer_default;

            fprintf_param_value(" 0=%d", num_output)
            fprintf_param_value(" 1=%d", kernel_w)
            { if (op->kernel_h != op->kernel_w) fprintf(pp, " 11=%d", op->kernel_h); }
            fprintf_param_value(" 2=%d", dilation_w)
            { if (op->dilation_h != op->dilation_w) fprintf(pp, " 12=%d", op->dilation_h); }
            fprintf_param_value(" 3=%d", stride_w)
            { if (op->stride_h != op->stride_w) fprintf(pp, " 13=%d", op->stride_h); }
            fprintf_param_value(" 4=%d", pad_left)
            { if (op->pad_top != op->pad_left) fprintf(pp, " 14=%d", op->pad_top); }
            { if (op->pad_right != op->pad_left) fprintf(pp, " 15=%d", op->pad_right); }
            { if (op->pad_bottom != op->pad_top) fprintf(pp, " 16=%d", op->pad_bottom); }
            fprintf_param_value(" 18=%e", pad_value)
            fprintf_param_value(" 5=%d", bias_term)
            fprintf_param_value(" 6=%d", weight_data_size)
            fprintf_param_value(" 8=%d", int8_scale_term)
            fprintf_param_value(" 9=%d", activation_type)
            { if (!op->activation_params.empty()) fprintf_param_float_array(10, op->activation_params, pp); }
            fprintf_param_value(" 17=%d", impl_type)

            fwrite_weight_tag_data(0, op->weight_data, bp);
            fwrite_weight_data(op->bias_data, bp);
        }
        else if (layer->type == "ConvolutionDepthWise")
        {
            ncnn::ConvolutionDepthWise* op = (ncnn::ConvolutionDepthWise*)layer;
            ncnn::ConvolutionDepthWise* op_default = (ncnn::ConvolutionDepthWise*)layer_default;

            fprintf_param_value(" 0=%d", num_output)
            fprintf_param_value(" 1=%d", kernel_w)
            { if (op->kernel_h != op->kernel_w) fprintf(pp, " 11=%d", op->kernel_h); }
            fprintf_param_value(" 2=%d", dilation_w)
            { if (op->dilation_h != op->dilation_w) fprintf(pp, " 12=%d", op->dilation_h); }
            fprintf_param_value(" 3=%d", stride_w)
            { if (op->stride_h != op->stride_w) fprintf(pp, " 13=%d", op->stride_h); }
            fprintf_param_value(" 4=%d", pad_left)
            { if (op->pad_top != op->pad_left) fprintf(pp, " 14=%d", op->pad_top); }
            { if (op->pad_right != op->pad_left) fprintf(pp, " 15=%d", op->pad_right); }
            { if (op->pad_bottom != op->pad_top) fprintf(pp, " 16=%d", op->pad_bottom); }
            fprintf_param_value(" 18=%e", pad_value)
            fprintf_param_value(" 5=%d", bias_term)
            fprintf_param_value(" 6=%d", weight_data_size)
            fprintf_param_value(" 7=%d", group)
            fprintf_param_value(" 8=%d", int8_scale_term)
            fprintf_param_value(" 9=%d", activation_type)
            { if (!op->activation_params.empty()) fprintf_param_float_array(10, op->activation_params, pp); }

            fwrite_weight_tag_data(0, op->weight_data, bp);
            fwrite_weight_data(op->bias_data, bp);
        }
        else if (layer->type == "Crop")
        {
            ncnn::Crop* op = (ncnn::Crop*)layer;
            ncnn::Crop* op_default = (ncnn::Crop*)layer_default;

            fprintf_param_value(" 0=%d", woffset)
            fprintf_param_value(" 1=%d", hoffset)
            fprintf_param_value(" 2=%d", coffset)
            fprintf_param_value(" 3=%d", outw)
            fprintf_param_value(" 4=%d", outh)
            fprintf_param_value(" 5=%d", outc)
            fprintf_param_value(" 6=%d", woffset2)
            fprintf_param_value(" 7=%d", hoffset2)
            fprintf_param_value(" 8=%d", coffset2)
            { if (!op->starts.empty()) fprintf_param_int_array(9, op->starts, pp); }
            { if (!op->ends.empty()) fprintf_param_int_array(10, op->ends, pp); }
            { if (!op->axes.empty()) fprintf_param_int_array(11, op->axes, pp); }
        }
        else if (layer->type == "Deconvolution")
        {
            ncnn::Deconvolution* op = (ncnn::Deconvolution*)layer;
            ncnn::Deconvolution* op_default = (ncnn::Deconvolution*)layer_default;

            fprintf_param_value(" 0=%d", num_output)
            fprintf_param_value(" 1=%d", kernel_w)
            { if (op->kernel_h != op->kernel_w) fprintf(pp, " 11=%d", op->kernel_h); }
            fprintf_param_value(" 2=%d", dilation_w)
            { if (op->dilation_h != op->dilation_w) fprintf(pp, " 12=%d", op->dilation_h); }
            fprintf_param_value(" 3=%d", stride_w)
            { if (op->stride_h != op->stride_w) fprintf(pp, " 13=%d", op->stride_h); }
            fprintf_param_value(" 4=%d", pad_left)
            { if (op->pad_top != op->pad_left) fprintf(pp, " 14=%d", op->pad_top); }
            { if (op->pad_right != op->pad_left) fprintf(pp, " 15=%d", op->pad_right); }
            { if (op->pad_bottom != op->pad_top) fprintf(pp, " 16=%d", op->pad_bottom); }
            fprintf_param_value(" 18=%d", output_pad_right)
            { if (op->output_pad_bottom != op->output_pad_right) fprintf(pp, " 19=%d", op->output_pad_bottom); }
            fprintf_param_value(" 20=%d", output_w)
            { if (op->output_h != op->output_w) fprintf(pp, " 21=%d", op->output_h); }
            fprintf_param_value(" 5=%d", bias_term)
            fprintf_param_value(" 6=%d", weight_data_size)
            fprintf_param_value(" 9=%d", activation_type)
            { if (!op->activation_params.empty()) fprintf_param_float_array(10, op->activation_params, pp); }

            fwrite_weight_tag_data(0, op->weight_data, bp);
            fwrite_weight_data(op->bias_data, bp);
        }
        else if (layer->type == "DeconvolutionDepthWise")
        {
            ncnn::DeconvolutionDepthWise* op = (ncnn::DeconvolutionDepthWise*)layer;
            ncnn::DeconvolutionDepthWise* op_default = (ncnn::DeconvolutionDepthWise*)layer_default;

            fprintf_param_value(" 0=%d", num_output)
            fprintf_param_value(" 1=%d", kernel_w)
            { if (op->kernel_h != op->kernel_w) fprintf(pp, " 11=%d", op->kernel_h); }
            fprintf_param_value(" 2=%d", dilation_w)
            { if (op->dilation_h != op->dilation_w) fprintf(pp, " 12=%d", op->dilation_h); }
            fprintf_param_value(" 3=%d", stride_w)
            { if (op->stride_h != op->stride_w) fprintf(pp, " 13=%d", op->stride_h); }
            fprintf_param_value(" 4=%d", pad_left)
            { if (op->pad_top != op->pad_left) fprintf(pp, " 14=%d", op->pad_top); }
            { if (op->pad_right != op->pad_left) fprintf(pp, " 15=%d", op->pad_right); }
            { if (op->pad_bottom != op->pad_top) fprintf(pp, " 16=%d", op->pad_bottom); }
            fprintf_param_value(" 18=%d", output_pad_right)
            { if (op->output_pad_bottom != op->output_pad_right) fprintf(pp, " 19=%d", op->output_pad_bottom); }
            fprintf_param_value(" 20=%d", output_w)
            { if (op->output_h != op->output_w) fprintf(pp, " 21=%d", op->output_h); }
            fprintf_param_value(" 5=%d", bias_term)
            fprintf_param_value(" 6=%d", weight_data_size)
            fprintf_param_value(" 7=%d", group)
            fprintf_param_value(" 9=%d", activation_type)
            { if (!op->activation_params.empty()) fprintf_param_float_array(10, op->activation_params, pp); }

            fwrite_weight_tag_data(0, op->weight_data, bp);
            fwrite_weight_data(op->bias_data, bp);
        }
        else if (layer->type == "DetectionOutput")
        {
            ncnn::DetectionOutput* op = (ncnn::DetectionOutput*)layer;
            ncnn::DetectionOutput* op_default = (ncnn::DetectionOutput*)layer_default;

            fprintf_param_value(" 0=%d", num_class)
            fprintf_param_value(" 1=%e", nms_threshold)
            fprintf_param_value(" 2=%d", nms_top_k)
            fprintf_param_value(" 3=%d", keep_top_k)
            fprintf_param_value(" 4=%e", confidence_threshold)
            fprintf_param_value(" 5=%e", variances[0])
            fprintf_param_value(" 6=%e", variances[1])
            fprintf_param_value(" 7=%e", variances[2])
            fprintf_param_value(" 8=%e", variances[3])
        }
        else if (layer->type == "Dropout")
        {
            ncnn::Dropout* op = (ncnn::Dropout*)layer;
            ncnn::Dropout* op_default = (ncnn::Dropout*)layer_default;

            fprintf_param_value(" 0=%e", scale)
        }
        else if (layer->type == "Eltwise")
        {
            ncnn::Eltwise* op = (ncnn::Eltwise*)layer;
            ncnn::Eltwise* op_default = (ncnn::Eltwise*)layer_default;

            fprintf_param_value(" 0=%d", op_type)
            { if (!op->coeffs.empty()) fprintf_param_float_array(1, op->coeffs, pp); }
        }
        else if (layer->type == "ELU")
        {
            ncnn::ELU* op = (ncnn::ELU*)layer;
            ncnn::ELU* op_default = (ncnn::ELU*)layer_default;

            fprintf_param_value(" 0=%e", alpha)
        }
        else if (layer->type == "Exp")
        {
            ncnn::Exp* op = (ncnn::Exp*)layer;
            ncnn::Exp* op_default = (ncnn::Exp*)layer_default;

            fprintf_param_value(" 0=%e", base)
            fprintf_param_value(" 1=%e", scale)
            fprintf_param_value(" 2=%e", shift)
        }
        else if (layer->type == "ExpandDims")
        {
            ncnn::ExpandDims* op = (ncnn::ExpandDims*)layer;
            ncnn::ExpandDims* op_default = (ncnn::ExpandDims*)layer_default;

            fprintf_param_value(" 0=%d", expand_w)
            fprintf_param_value(" 1=%d", expand_h)
            fprintf_param_value(" 2=%d", expand_c)
            { if (!op->axes.empty()) fprintf_param_int_array(0, op->axes, pp); }
        }
        else if (layer->type == "HardSigmoid")
        {
            ncnn::HardSigmoid* op = (ncnn::HardSigmoid*)layer;
            ncnn::HardSigmoid* op_default = (ncnn::HardSigmoid*)layer_default;

            fprintf_param_value(" 0=%e", alpha)
            fprintf_param_value(" 1=%e", beta)
        }
        else if (layer->type == "HardSwish")
        {
            ncnn::HardSwish* op = (ncnn::HardSwish*)layer;
            ncnn::HardSwish* op_default = (ncnn::HardSwish*)layer_default;

            fprintf_param_value(" 0=%e", alpha)
            fprintf_param_value(" 1=%e", beta)
        }
        else if (layer->type == "InnerProduct")
        {
            ncnn::InnerProduct* op = (ncnn::InnerProduct*)layer;
            ncnn::InnerProduct* op_default = (ncnn::InnerProduct*)layer_default;

            fprintf_param_value(" 0=%d", num_output)
            fprintf_param_value(" 1=%d", bias_term)
            fprintf_param_value(" 2=%d", weight_data_size)
            fprintf_param_value(" 8=%d", int8_scale_term)
            fprintf_param_value(" 9=%d", activation_type)
            { if (!op->activation_params.empty()) fprintf_param_float_array(10, op->activation_params, pp); }

            fwrite_weight_tag_data(0, op->weight_data, bp);
            fwrite_weight_data(op->bias_data, bp);
        }
        else if (layer->type == "Input")
        {
            ncnn::Input* op = (ncnn::Input*)layer;
            ncnn::Input* op_default = (ncnn::Input*)layer_default;

            fprintf_param_value(" 0=%d", w)
            fprintf_param_value(" 1=%d", h)
            fprintf_param_value(" 2=%d", c)
        }
        else if (layer->type == "InstanceNorm")
        {
            ncnn::InstanceNorm* op = (ncnn::InstanceNorm*)layer;
            ncnn::InstanceNorm* op_default = (ncnn::InstanceNorm*)layer_default;

            fprintf_param_value(" 0=%d", channels)
            fprintf_param_value(" 1=%e", eps)

            fwrite_weight_data(op->gamma_data, bp);
            fwrite_weight_data(op->beta_data, bp);
        }
        else if (layer->type == "Interp")
        {
            ncnn::Interp* op = (ncnn::Interp*)layer;
            ncnn::Interp* op_default = (ncnn::Interp*)layer_default;

            fprintf_param_value(" 0=%d", resize_type)
            fprintf_param_value(" 1=%e", height_scale)
            fprintf_param_value(" 2=%e", width_scale)
            fprintf_param_value(" 3=%d", output_height)
            fprintf_param_value(" 4=%d", output_width)
        }
        else if (layer->type == "Log")
        {
            ncnn::Log* op = (ncnn::Log*)layer;
            ncnn::Log* op_default = (ncnn::Log*)layer_default;

            fprintf_param_value(" 0=%e", base)
            fprintf_param_value(" 1=%e", scale)
            fprintf_param_value(" 2=%e", shift)
        }
        else if (layer->type == "LRN")
        {
            ncnn::LRN* op = (ncnn::LRN*)layer;
            ncnn::LRN* op_default = (ncnn::LRN*)layer_default;

            fprintf_param_value(" 0=%d", region_type)
            fprintf_param_value(" 1=%d", local_size)
            fprintf_param_value(" 2=%e", alpha)
            fprintf_param_value(" 3=%e", beta)
            fprintf_param_value(" 4=%e", bias)
        }
        else if (layer->type == "MemoryData")
        {
            ncnn::MemoryData* op = (ncnn::MemoryData*)layer;
            ncnn::MemoryData* op_default = (ncnn::MemoryData*)layer_default;

            fprintf_param_value(" 0=%d", w)
            fprintf_param_value(" 1=%d", h)
            fprintf_param_value(" 2=%d", c)
            fwrite_weight_data(op->data, bp);
        }
        else if (layer->type == "MVN")
        {
            ncnn::MVN* op = (ncnn::MVN*)layer;
            ncnn::MVN* op_default = (ncnn::MVN*)layer_default;

            fprintf_param_value(" 0=%d", normalize_variance)
            fprintf_param_value(" 1=%d", across_channels)
            fprintf_param_value(" 2=%e", eps)
        }
        else if (layer->type == "Normalize")
        {
            ncnn::Normalize* op = (ncnn::Normalize*)layer;
            ncnn::Normalize* op_default = (ncnn::Normalize*)layer_default;

            fprintf_param_value(" 0=%d", across_spatial)
            fprintf_param_value(" 1=%d", channel_shared)
            fprintf_param_value(" 2=%e", eps)
            fprintf_param_value(" 3=%d", scale_data_size)
            fprintf_param_value(" 4=%d", across_channel)
            fprintf_param_value(" 9=%d", eps_mode)

            fwrite_weight_data(op->scale_data, bp);
        }
        else if (layer->type == "Padding")
        {
            ncnn::Padding* op = (ncnn::Padding*)layer;
            ncnn::Padding* op_default = (ncnn::Padding*)layer_default;

            fprintf_param_value(" 0=%d", top)
            fprintf_param_value(" 1=%d", bottom)
            fprintf_param_value(" 2=%d", left)
            fprintf_param_value(" 3=%d", right)
            fprintf_param_value(" 4=%d", type)
            fprintf_param_value(" 5=%e", value)
        }
        else if (layer->type == "Permute")
        {
            ncnn::Permute* op = (ncnn::Permute*)layer;
            ncnn::Permute* op_default = (ncnn::Permute*)layer_default;

            fprintf_param_value(" 0=%d", order_type)
        }
        else if (layer->type == "Pooling")
        {
            ncnn::Pooling* op = (ncnn::Pooling*)layer;
            ncnn::Pooling* op_default = (ncnn::Pooling*)layer_default;

            fprintf_param_value(" 0=%d", pooling_type)
            fprintf_param_value(" 1=%d", kernel_w)
            { if (op->kernel_h != op->kernel_w) fprintf(pp, " 11=%d", op->kernel_h); }
            fprintf_param_value(" 2=%d", stride_w)
            { if (op->stride_h != op->stride_w) fprintf(pp, " 12=%d", op->stride_h); }
            fprintf_param_value(" 3=%d", pad_left)
            { if (op->pad_top != op->pad_left) fprintf(pp, " 13=%d", op->pad_top); }
            { if (op->pad_right != op->pad_left) fprintf(pp, " 14=%d", op->pad_right); }
            { if (op->pad_bottom != op->pad_top) fprintf(pp, " 15=%d", op->pad_bottom); }
            fprintf_param_value(" 4=%d", global_pooling)
            fprintf_param_value(" 5=%d", pad_mode)
        }
        else if (layer->type == "Power")
        {
            ncnn::Power* op = (ncnn::Power*)layer;
            ncnn::Power* op_default = (ncnn::Power*)layer_default;

            fprintf_param_value(" 0=%e", power)
            fprintf_param_value(" 1=%e", scale)
            fprintf_param_value(" 2=%e", shift)
        }
        else if (layer->type == "PReLU")
        {
            ncnn::PReLU* op = (ncnn::PReLU*)layer;
            ncnn::PReLU* op_default = (ncnn::PReLU*)layer_default;

            fprintf_param_value(" 0=%d", num_slope)

            fwrite_weight_data(op->slope_data, bp);
        }
        else if (layer->type == "PriorBox")
        {
            ncnn::PriorBox* op = (ncnn::PriorBox*)layer;
            ncnn::PriorBox* op_default = (ncnn::PriorBox*)layer_default;

            { if (!op->min_sizes.empty()) fprintf_param_float_array(0, op->min_sizes, pp); }
            { if (!op->max_sizes.empty()) fprintf_param_float_array(1, op->max_sizes, pp); }
            { if (!op->aspect_ratios.empty()) fprintf_param_float_array(2, op->aspect_ratios, pp); }
            fprintf_param_value(" 3=%e", variances[0])
            fprintf_param_value(" 4=%e", variances[1])
            fprintf_param_value(" 5=%e", variances[2])
            fprintf_param_value(" 6=%e", variances[3])
            fprintf_param_value(" 7=%d", flip)
            fprintf_param_value(" 8=%d", clip)
            fprintf_param_value(" 9=%d", image_width)
            fprintf_param_value(" 10=%d", image_height)
            fprintf_param_value(" 11=%e", step_width)
            fprintf_param_value(" 12=%e", step_height)
            fprintf_param_value(" 13=%e", offset)
        }
        else if (layer->type == "Proposal")
        {
            ncnn::Proposal* op = (ncnn::Proposal*)layer;
            ncnn::Proposal* op_default = (ncnn::Proposal*)layer_default;

            fprintf_param_value(" 0=%d", feat_stride)
            fprintf_param_value(" 1=%d", base_size)
            fprintf_param_value(" 2=%d", pre_nms_topN)
            fprintf_param_value(" 3=%d", after_nms_topN)
            fprintf_param_value(" 4=%e", nms_thresh)
            fprintf_param_value(" 5=%d", min_size)
        }
        else if (layer->type == "PSROIPooling")
        {
            ncnn::PSROIPooling* op = (ncnn::PSROIPooling*)layer;
            ncnn::PSROIPooling* op_default = (ncnn::PSROIPooling*)layer_default;

            fprintf_param_value(" 0=%d", pooled_width)
            fprintf_param_value(" 1=%d", pooled_height)
            fprintf_param_value(" 2=%e", spatial_scale)
            fprintf_param_value(" 3=%d", output_dim)
        }
        else if (layer->type == "Quantize")
        {
            ncnn::Quantize* op = (ncnn::Quantize*)layer;
            ncnn::Quantize* op_default = (ncnn::Quantize*)layer_default;

            fprintf_param_value(" 0=%e", scale)
        }
        else if (layer->type == "Reduction")
        {
            ncnn::Reduction* op = (ncnn::Reduction*)layer;
            ncnn::Reduction* op_default = (ncnn::Reduction*)layer_default;

            fprintf_param_value(" 0=%d", operation)
            fprintf_param_value(" 1=%d", reduce_all)
            fprintf_param_value(" 2=%e", coeff)
            { if (!op->axes.empty()) fprintf_param_int_array(3, op->axes, pp); }
            fprintf_param_value(" 4=%d", keepdims)
        }
        else if (layer->type == "ReLU")
        {
            ncnn::ReLU* op = (ncnn::ReLU*)layer;
            ncnn::ReLU* op_default = (ncnn::ReLU*)layer_default;

            fprintf_param_value(" 0=%e", slope)
        }
        else if (layer->type == "Reorg")
        {
            ncnn::Reorg* op = (ncnn::Reorg*)layer;
            ncnn::Reorg* op_default = (ncnn::Reorg*)layer_default;

            fprintf_param_value(" 0=%d", stride)
        }
        else if (layer->type == "Requantize")
        {
            ncnn::Requantize* op = (ncnn::Requantize*)layer;
            ncnn::Requantize* op_default = (ncnn::Requantize*)layer_default;

            fprintf_param_value(" 0=%e", scale_in)
            fprintf_param_value(" 1=%e", scale_out)
            fprintf_param_value(" 2=%d", bias_term)
            fprintf_param_value(" 3=%d", bias_data_size)
            fprintf_param_value(" 4=%d", fusion_relu)
        }
        else if (layer->type == "Reshape")
        {
            ncnn::Reshape* op = (ncnn::Reshape*)layer;
            ncnn::Reshape* op_default = (ncnn::Reshape*)layer_default;

            fprintf_param_value(" 0=%d", w)
            fprintf_param_value(" 1=%d", h)
            fprintf_param_value(" 2=%d", c)
            fprintf_param_value(" 3=%d", permute)
        }
        else if (layer->type == "ROIAlign")
        {
            ncnn::ROIAlign* op = (ncnn::ROIAlign*)layer;
            ncnn::ROIAlign* op_default = (ncnn::ROIAlign*)layer_default;

            fprintf_param_value(" 0=%d", pooled_width)
            fprintf_param_value(" 1=%d", pooled_height)
            fprintf_param_value(" 2=%e", spatial_scale)
        }
        else if (layer->type == "ROIPooling")
        {
            ncnn::ROIPooling* op = (ncnn::ROIPooling*)layer;
            ncnn::ROIPooling* op_default = (ncnn::ROIPooling*)layer_default;

            fprintf_param_value(" 0=%d", pooled_width)
            fprintf_param_value(" 1=%d", pooled_height)
            fprintf_param_value(" 2=%e", spatial_scale)
        }
        else if (layer->type == "Scale")
        {
            ncnn::Scale* op = (ncnn::Scale*)layer;
            ncnn::Scale* op_default = (ncnn::Scale*)layer_default;

            fprintf_param_value(" 0=%d", scale_data_size)
            fprintf_param_value(" 1=%d", bias_term)

            fwrite_weight_data(op->scale_data, bp);
            fwrite_weight_data(op->bias_data, bp);
        }
        else if (layer->type == "ShuffleChannel")
        {
            ncnn::ShuffleChannel* op = (ncnn::ShuffleChannel*)layer;
            ncnn::ShuffleChannel* op_default = (ncnn::ShuffleChannel*)layer_default;

            fprintf_param_value(" 0=%d", group)
        }
        else if (layer->type == "Slice")
        {
            ncnn::Slice* op = (ncnn::Slice*)layer;
            ncnn::Slice* op_default = (ncnn::Slice*)layer_default;

            { if (!op->slices.empty()) fprintf_param_int_array(0, op->slices, pp); }
            fprintf_param_value(" 1=%d", axis)
        }
        else if (layer->type == "Softmax")
        {
            ncnn::Softmax* op = (ncnn::Softmax*)layer;
            ncnn::Softmax* op_default = (ncnn::Softmax*)layer_default;

            fprintf_param_value(" 0=%d", axis)

            // HACK
            if (op->axis != 0)
            {
                int fixbug0 = 1;
                fprintf(pp, " 1=%d", fixbug0);
            }
        }
        else if (layer->type == "Squeeze")
        {
            ncnn::Squeeze* op = (ncnn::Squeeze*)layer;
            ncnn::Squeeze* op_default = (ncnn::Squeeze*)layer_default;

            fprintf_param_value(" 0=%d", squeeze_w)
            fprintf_param_value(" 1=%d", squeeze_h)
            fprintf_param_value(" 2=%d", squeeze_c)
            { if (!op->axes.empty()) fprintf_param_int_array(0, op->axes, pp); }
        }
        else if (layer->type == "Threshold")
        {
            ncnn::Threshold* op = (ncnn::Threshold*)layer;
            ncnn::Threshold* op_default = (ncnn::Threshold*)layer_default;

            fprintf_param_value(" 0=%e", threshold)
        }
        else if (layer->type == "UnaryOp")
        {
            ncnn::UnaryOp* op = (ncnn::UnaryOp*)layer;
            ncnn::UnaryOp* op_default = (ncnn::UnaryOp*)layer_default;

            fprintf_param_value(" 0=%d", op_type)
        }
        else if (layer->type == "YoloDetectionOutput")
        {
            ncnn::YoloDetectionOutput* op = (ncnn::YoloDetectionOutput*)layer;
            ncnn::YoloDetectionOutput* op_default = (ncnn::YoloDetectionOutput*)layer_default;

            fprintf_param_value(" 0=%d", num_class)
            fprintf_param_value(" 1=%d", num_box)
            fprintf_param_value(" 2=%e", confidence_threshold)
            fprintf_param_value(" 3=%e", nms_threshold)
            { if (!op->biases.empty()) fprintf_param_float_array(4, op->biases, pp); }
        }
        else if (layer->type == "Yolov3DetectionOutput")
        {
            ncnn::Yolov3DetectionOutput* op = (ncnn::Yolov3DetectionOutput*)layer;
            ncnn::Yolov3DetectionOutput* op_default = (ncnn::Yolov3DetectionOutput*)layer_default;

            fprintf_param_value(" 0=%d", num_class)
            fprintf_param_value(" 1=%d", num_box)
            fprintf_param_value(" 2=%e", confidence_threshold)
            fprintf_param_value(" 3=%e", nms_threshold)
            { if (!op->biases.empty()) fprintf_param_float_array(4, op->biases, pp); }
            { if (!op->mask.empty()) fprintf_param_int_array(5, op->mask, pp); }
            { if (!op->anchors_scale.empty()) fprintf_param_float_array(6, op->anchors_scale, pp); }
        }

#undef fprintf_param_value

        fprintf(pp, "\n");

        delete layer_default;
    }

    fclose(pp);
    fclose(bp);

    return 0;
}

int main(int argc, char** argv)
{
#if defined(__aarch64__) && defined(LINUX)
    if (argc != 10)
    {
        fprintf(stderr, "usage: %s [inparam] [inbin] [outparam] [outbin] [flag] [dataname] [w] [h] [c]\n", argv[0]);
        return -1;
    }
    const char* dataname = argv[6];
    int inw = atoi(argv[7]);
    int inh = atoi(argv[8]);
    int inc = atoi(argv[9]);
#else
    if (argc != 6)
    {
        fprintf(stderr, "usage: %s [inparam] [inbin] [outparam] [outbin] [flag]\n", argv[0]);
        return -1;
    }
#endif // defined(__aarch64__) && defined(LINUX)

    const char* inparam = argv[1];
    const char* inbin = argv[2];
    const char* outparam = argv[3];
    const char* outbin = argv[4];
    int flag = atoi(argv[5]);

    NetOptimize optimizer;

    if (flag == 65536)
    {
        optimizer.storage_type = 1;
    }
    else
    {
        optimizer.storage_type = 0;
    }

    optimizer.load_param(inparam);
    if (strcmp(inbin, "null") == 0)
    {
        DataReaderFromEmpty dr;
        optimizer.load_model(dr);
    }
    else
        optimizer.load_model(inbin);

#if defined(__aarch64__) && defined(LINUX)
    optimizer.find_fastest_fp32_conv(dataname, inw, inh, inc);
#endif // defined(__aarch64__) && defined(LINUX)
    optimizer.fuse_batchnorm_scale();
    optimizer.fuse_convolution_batchnorm();
    optimizer.fuse_convolutiondepthwise_batchnorm();
    optimizer.fuse_deconvolution_batchnorm();
    optimizer.fuse_deconvolutiondepthwise_batchnorm();
    optimizer.fuse_innerproduct_batchnorm();
    optimizer.fuse_innerproduct_dropout();
    optimizer.fuse_convolution_activation();
    optimizer.fuse_convolutiondepthwise_activation();
    optimizer.fuse_deconvolution_activation();
    optimizer.fuse_deconvolutiondepthwise_activation();
    optimizer.fuse_innerproduct_activation();

    optimizer.eliminate_dropout();
    optimizer.eliminate_pooling1x1();
    optimizer.eliminate_noop();
    optimizer.eliminate_flatten_after_global_pooling();
    optimizer.eliminate_reshape_after_global_pooling();
    optimizer.eliminate_reshape_before_binaryop();

    optimizer.replace_convolution_with_innerproduct_after_global_pooling();
    optimizer.replace_convolution_with_innerproduct_after_innerproduct();

    optimizer.eliminate_flatten_after_innerproduct();
    optimizer.eliminate_orphaned_memorydata();

    optimizer.save(outparam, outbin);

    return 0;
}