手动实现部分卷积功能（stride,padding）以便更好理解卷积行为

多看一眼就爆炸靠近点就被融化

已于 2023-05-03 16:35:30 修改

阅读量1.4k

点赞数 1

文章标签： python 深度学习开发语言

于 2022-08-13 22:05:21 首次发布

本文链接：https://blog.csdn.net/qq_48355075/article/details/126324634

版权

1.卷积运算行为

2.卷积的超参数之一：stride（步幅）、padding（填充，通常以0填充）

3.基于1和2的原理手动实现最普通的卷积行为

def corr2d(X, K, stride=1, padding=0):
    '''
    :param X: (H, W)
    :param K: (k_h, k_w)
    :return: (Y_H, Y_W)
    '''

    H, W = X.shape
    k_h, k_w = K.shape
    Y_H = int(math.floor((H - k_h + 2 * padding) / stride + 1))
    Y_W = int(math.floor((W - k_w + 2 * padding) / stride + 1))
    if padding > 0:
        new_X = torch.zeros((H + 2 * padding, W + 2 * padding))
        new_X[padding: - padding, padding: -padding] = X
        X2multi = new_X
        if new_X.shape[0] % 2 == 0 and new_X.shape[1] % 2 != 0:
            X2multi = new_X[0: -1, :]
        if new_X.shape[1] % 2 == 0 and new_X.shape[0] % 2 != 0:
            X2multi = new_X[:, 0: -1]
        if new_X.shape[1] % 2 == 0 and new_X.shape[0] % 2 == 0:
            X2multi = new_X[0: -1, 0: -1]
    else:
        X2multi = X
    Y = torch.zeros(Y_H, Y_W)
    for i in range(Y.shape[0]):
        for j in range(Y.shape[1]):
            row_start = stride * i
            colum_start = stride * j
            Y[i, j] = (X2multi[row_start : row_start + k_h, colum_start : colum_start + k_w] * K).sum()
    return Y

需要注意，如下情况。

0.在被卷积的张量X的分辨率H×W中，H或W为奇数时，以H为例，卷积之后的分辨率遵循如下计算公式： $H2 = (H1 - K + 2P)/ S + 1$

1.H或W为偶数时，以H为例，卷积之后的分辨率遵循如下公式： $H2 = floor[(H1 - K + 2P)/ S] + 1$ ，其中floor表示向下取整。

2.padding为P时其实表示原来被卷积的张量X，所属的H和W，变为了H+2P和W+2P

3.举例：H=W=5,S=1,P=0 --> H'=W'=3;H=W=5,S=2,P=0 --> H'=W'=2;

H=W=5,S=1,P=1 --> H'=W'=5;H=W=5,S=2,P=1 --> H'=W'=3;

H=W=4,S=1,P=0 --> H'=W'=2;H=W=4,S=2,P=0 --> H'=W'=1;

H=W=4,S=1,P=1 --> H'=W'=4;H=W=4,S=2,P=1 --> H'=W'=2;

上述指的是K=3的情况

4.以上四项示例在上面代码中已经相应给出，可自行debug理解

4.多输入通道

应注意卷积核的通道数与输入通道数一致

代码手动实现如下。

def corr2d_muti_in(X, K, stride=1, padding=0):
    '''
    :param X: (C_in, H, W)
    :param K: (C_in, k_h, k_w)
    :return: (Y_H, Y_W)
    '''
    res = corr2d(X[0, :, :], K[0, :, :], stride, padding)
    for i in range(1, X.shape[0]):
        res += corr2d(X[i, :, :], K[i, :, :], stride, padding)
    return res

5.多输出通道

代码实现如下段所示。

def corr2d_muti_in_out(X, K, stride=1, padding=0):
    out = torch.stack([corr2d_muti_in(X, k, stride, padding) for k in K])
    return out

6.Convd2D的部分功能等价手动实现

直接是代码哦。

class MyConv2D(nn.Module):
    def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0):
        super(MyConv2D, self).__init__()
        if isinstance(kernel_size, int):
            size = (out_channels, in_channels, kernel_size, kernel_size)
        elif isinstance(kernel_size, tuple):
            size = (out_channels, in_channels, kernel_size[0], kernel_size[1])
        self.weight = nn.Parameter(torch.randn(size))
        self.bias = nn.Parameter(torch.randn(out_channels, 1, 1))
        self.stride = stride
        self.padding = padding

    def forward(self, x):
        '''
        :param x: (C_in, H, W)
        :return: (C_out, Y_H, Y_W)
        '''
        return corr2d_muti_in_out(x, self.weight, self.stride, self.padding) + self.bias

可以自行debug尝试，用此可以看出，一个卷积层的参数数量为 $out_c* in_c * H_k*W_k+out_c$ 。即一个卷积层为MyConv2D(16, 3, 3, 1, 0)或者为MyConv2D(16, 3, (3, 3), 1, 0)，这个卷积层的参数为：16*3*3*3+16=448。

7.总体代码

# -*- coding: utf-8 -*-
"""
@Time ： 2022/8/13 9:00
@Auth ： Fanteng Meng
@File ：conv_manual.py
@IDE ：PyCharm
@Github : https://github.com/FT115

"""
import torch
from torch import nn
import math

def corr2d(X, K, stride=1, padding=0):
    '''
    :param X: (H, W)
    :param K: (k_h, k_w)
    :return: (Y_H, Y_W)
    '''

    H, W = X.shape
    k_h, k_w = K.shape
    Y_H = int(math.floor((H - k_h + 2 * padding) / stride + 1))
    Y_W = int(math.floor((W - k_w + 2 * padding) / stride + 1))
    if padding > 0:
        new_X = torch.zeros((H + 2 * padding, W + 2 * padding))
        new_X[padding: - padding, padding: -padding] = X
        X2multi = new_X
        if new_X.shape[0] % 2 == 0 and new_X.shape[1] % 2 != 0:
            X2multi = new_X[0: -1, :]
        if new_X.shape[1] % 2 == 0 and new_X.shape[0] % 2 != 0:
            X2multi = new_X[:, 0: -1]
        if new_X.shape[1] % 2 == 0 and new_X.shape[0] % 2 == 0:
            X2multi = new_X[0: -1, 0: -1]
    else:
        X2multi = X
    Y = torch.zeros(Y_H, Y_W)
    for i in range(Y.shape[0]):
        for j in range(Y.shape[1]):
            row_start = stride * i
            colum_start = stride * j
            Y[i, j] = (X2multi[row_start : row_start + k_h, colum_start : colum_start + k_w] * K).sum()
    return Y

def corr2d_muti_in(X, K, stride=1, padding=0):
    '''
    :param X: (C_in, H, W)
    :param K: (C_in, k_h, k_w)
    :return: (Y_H, Y_W)
    '''
    res = corr2d(X[0, :, :], K[0, :, :], stride, padding)
    for i in range(1, X.shape[0]):
        res += corr2d(X[i, :, :], K[i, :, :], stride, padding)
    return res

def corr2d_muti_in_out(X, K, stride=1, padding=0):
    out = torch.stack([corr2d_muti_in(X, k, stride, padding) for k in K])
    return out

class MyConv2D(nn.Module):
    def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0):
        super(MyConv2D, self).__init__()
        if isinstance(kernel_size, int):
            size = (out_channels, in_channels, kernel_size, kernel_size)
        elif isinstance(kernel_size, tuple):
            size = (out_channels, in_channels, kernel_size[0], kernel_size[1])
        self.weight = nn.Parameter(torch.randn(size))
        self.bias = nn.Parameter(torch.randn(out_channels, 1, 1))
        self.stride = stride
        self.padding = padding

    def forward(self, x):
        '''
        :param x: (C_in, H, W)
        :return: (C_out, Y_H, Y_W)
        '''
        return corr2d_muti_in_out(x, self.weight, self.stride, self.padding) + self.bias

记录自己学习到的知识的同时，希望对你有所帮助！