2.Conv2d实现

[C++ 基于Eigen库实现CRN前向推理]

第二部分：Conv2d实现

• 前言：(Eigen库使用记录)
• 第一部分：WavFile.class (实现读取wav/pcm,实现STFT)
• 第二部分：Conv2d实现
• 第三部分：TransposedConv2d实现 (mimo,padding,stride,dilation,kernel,outpadding)
• 第四部分：NonLinearity （Sigmoid，Tanh，ReLU，ELU，Softplus）
• 第五部分：LSTM
• GITHUB仓库

1. Conv2d介绍

1.1 参数介绍

这是pytorch官方源码

def __init__(
        self,
        in_channels: int,
        out_channels: int,
        kernel_size: _size_2_t,
        stride: _size_2_t = 1,
        padding: _size_2_t = 0,
        dilation: _size_2_t = 1,
        groups: int = 1,
        bias: bool = True,
        padding_mode: str = 'zeros'  # TODO: refine this type
    ):

• in_channels：网络输入的通道数，简单理解为每个输入样本包含多个个FeatureMap。

• out_channels：网络输出的通道数。即卷积核的个数

• kernel_size：卷积核的大小，如果该参数是一个整数q，那么卷积核的大小是qXq。
  至此，一个简单的卷积如图
  

• stride：步长。是卷积过程中移动的步长。默认情况下是1。一般卷积核在输入图像上的移动是自左至右，自上至下。如果参数是一个整数那么就默认在水平和垂直方向都是该整数。如果参数是stride=(2, 1),2代表着高（h）进行步长为2，1代表着宽（w）进行步长为1。
  加入步长后，当步长为2时，卷积如图：
  

• padding：填充，默认是0值填充。改参数指定的是在边缘填充多少行或列的0值
  如padding为1时，卷积如图
  

• dilation：扩张。一般情况下，卷积核与输入图像对应的位置之间的计算是相同尺寸的，也就是说卷积核的大小是3X3，那么它在输入图像上每次作用的区域是3X3，这种情况下dilation=0。当dilation=1时，表示的是下图这种情况。
  

1.2 python实现

2. 基于Eigen的C++实现

2.1 Layer_Conv2d.h

//
// Created by Koer on 2022/10/31.
//

#ifndef CRN_LAYER_CONV2D_H
#define CRN_LAYER_CONV2D_H


#include "vector"
#include "mat.h"
#include "Eigen"
#include "tuple"

#include "Eigen/CXX11/Tensor"

class Layer_Conv2d {
public:
    Layer_Conv2d();

    Layer_Conv2d(int64_t in_ch, int64_t out_ch, std::pair<int64_t, int64_t> kernel = std::make_pair(1, 1),
                 std::pair<int64_t, int64_t> stride = std::make_pair(1, 1),
                 std::pair<int64_t, int64_t> dilation = std::make_pair(1, 1),
                 std::pair<int64_t, int64_t> padding = std::make_pair(0, 0));

    void LoadState(MATFile *pmFile, const std::string &state_preffix);

    void LoadTestState();

    Eigen::Tensor<float_t, 4> forward(Eigen::Tensor<float_t, 4> &input);

private:
    int64_t in_channels;
    int64_t out_channels;
    std::pair<int64_t, int64_t> kernel_size;
    std::pair<int64_t, int64_t> stride;
    std::pair<int64_t, int64_t> dilation;
    std::pair<int64_t, int64_t> padding;
    Eigen::Tensor<float_t, 4> weights;
    Eigen::Tensor<float_t, 2> bias;

};


#endif //CRN_LAYER_CONV2D_H

2.2 Layer_Conv2d.cpp

//
// Created by Koer on 2022/10/31.
//

#include "iostream"
#include "../include/Layer_Conv2d.h"


Layer_Conv2d::Layer_Conv2d() {
    this->in_channels = 1;
    this->out_channels = 1;
    this->kernel_size = std::make_pair(1, 1);
    this->stride = std::make_pair(1, 1);
    this->padding = std::make_pair(0, 0);
}

Layer_Conv2d::Layer_Conv2d(int64_t in_ch, int64_t out_ch,
                           std::pair<int64_t, int64_t> kernel,
                           std::pair<int64_t, int64_t> stride,
                           std::pair<int64_t, int64_t> dilation,
                           std::pair<int64_t, int64_t> padding) {
    /* code */
    this->in_channels = in_ch;
    this->out_channels = out_ch;
    this->kernel_size = kernel;
    this->stride = stride;
    this->dilation = dilation;
    this->padding = padding;
}

void Layer_Conv2d::LoadState(MATFile *pmFile, const std::string &state_preffix) {
    std::string weight_name = state_preffix + "_weight";
    std::string bias_name = state_preffix + "_bias";

    // Read weight
    mxArray *pa = matGetVariable(pmFile, weight_name.c_str());
    auto *values = (float_t *) mxGetData(pa);
    // First Dimension  eg.(16,1,2,3)  ===> M=16
    long long dim1 = mxGetM(pa);
    // Rest Total Dimension eg.(16,1,2,3) ===>N = 1 * 2 * 3 = 6
    long long dim2 = mxGetN(pa);
    dim2 = dim2 / this->kernel_size.first / this->kernel_size.second;
    this->weights.resize(dim1, dim2, this->kernel_size.first, this->kernel_size.second);
    int idx = 0;
    for (int i = 0; i < this->kernel_size.second; i++) {
        for (int j = 0; j < this->kernel_size.first; j++) {
            for (int k = 0; k < dim2; k++) {
                for (int l = 0; l < dim1; l++) {
                    this->weights(l, k, j, i) = values[idx++];
                }
            }
        }
    }
    // std::cout << this->weights << std::endl;

    // Read bias
    pa = matGetVariable(pmFile, bias_name.c_str());
    values = (float_t *) mxGetData(pa);
    dim1 = mxGetM(pa);
    dim2 = mxGetN(pa);
    this->bias.resize(dim1, dim2);
    idx = 0;
    for (int i = 0; i < dim2; i++) {
        for (int j = 0; j < dim1; j++) {
            this->bias(j, i) = values[idx++];
        }
    }
    // std::cout << this->bias << std::endl;
    // std::cout << " Finish Loading State of " + state_preffix << std::endl;
}

void Layer_Conv2d::LoadTestState() {
    Eigen::Tensor<float_t, 4> w(this->out_channels, this->in_channels, this->kernel_size.first,
                                this->kernel_size.second);
    w.setConstant(1.0);
    this->weights = w;
    Eigen::Tensor<float_t, 2> b(1, this->out_channels);
    b.setConstant(0.0);
    this->bias = b;
}

Eigen::Tensor<float_t, 4> Layer_Conv2d::forward(Eigen::Tensor<float_t, 4> &input) {
    const Eigen::Tensor<size_t, 4>::Dimensions &dim_inp = input.dimensions();

    /* Sequence channel × T × F */
    size_t pad_size_time = this->padding.first;
    size_t pad_size_freq = this->padding.second;
    int64_t batch = dim_inp[0], C_in = dim_inp[1], H_in = dim_inp[2], W_in = dim_inp[3];
    int64_t H_pad = H_in + pad_size_time * 2;
    int64_t W_pad = W_in + pad_size_freq * 2;

    /* padding tensor */
    Eigen::Tensor<float_t, 4> padded_input = Eigen::Tensor<float_t, 4>(batch, C_in, H_pad, W_pad);
    padded_input.setZero();
    padded_input.slice(Eigen::array<size_t, 4>{0, 0, pad_size_time, pad_size_freq}, dim_inp) = input;

    /* output shape */
    int64_t H_out = (H_pad - this->dilation.first * (this->kernel_size.first - 1) - 1) / this->stride.first + 1;
    int64_t W_out = (W_pad - this->dilation.second * (this->kernel_size.second - 1) - 1) / this->stride.second + 1;
    Eigen::Tensor<float_t, 4> output = Eigen::Tensor<float_t, 4>(batch, this->out_channels, H_out, W_out);
    output.setZero();

    /* params
     * region: tmp storage of map to be convolved
     * kernel: tmp storage of kernel of the out_channels idx_outc
     * tmp_res: tmp storage of res (convolve all in_channels and sum up)
     * dim_sum: the origin tmp_res is at view of (1,ic,k1,k2), sum along the 1,2,3 axis
     * h_region: the h of convolve region - 1
     * w_region: the w of convolve region - 1
    */
    Eigen::Tensor<float_t, 4> region;
    Eigen::Tensor<float_t, 4> kernel;
    Eigen::Tensor<float_t, 1> tmp_res;
    Eigen::array<int, 3> dim_sum{1, 2, 3};
    int64_t h_region = (this->kernel_size.first - 1) * this->dilation.first;
    int64_t w_region = (this->kernel_size.second - 1) * this->dilation.second;
    for (int64_t idx_batch = 0; idx_batch < batch; idx_batch++) {
        for (int64_t idx_outc = 0; idx_outc < this->out_channels; idx_outc++) {
            kernel = this->weights.slice(Eigen::array<int64_t, 4>{idx_outc, 0, 0, 0},
                                         Eigen::array<int64_t, 4>{1, this->in_channels, this->kernel_size.first,
                                                                  this->kernel_size.second}
            );
            for (int64_t idx_h = 0; idx_h < H_pad - h_region; idx_h += stride.first) {
                for (int64_t idx_w = 0; idx_w < W_pad - w_region; idx_w += stride.second) {
                    region = padded_input.stridedSlice(
                            Eigen::array<int64_t, 4>{idx_batch, 0, idx_h, idx_w},
                            Eigen::array<int64_t, 4>{idx_batch + 1, this->in_channels, idx_h + h_region + 1,
                                                     idx_w + w_region + 1},
                            Eigen::array<int64_t, 4>{1, 1, this->dilation.first, this->dilation.second});
                    tmp_res = (region * kernel).sum(dim_sum);
                    output(idx_batch, idx_outc, idx_h / this->stride.first, idx_w / this->stride.second) =
                            tmp_res(0) + this->bias(0, idx_outc);
                }
            }
        }
    }
    return output;
}

这是基于循环写的，效率十分十分低。后面要写成unfold形式进行并行运算。

参考链接

[1] Conv2d介绍