卷积层转全连接

卷积层转全连接测试代码

def conv2fn(conv, input_shape, output_shape):
    # padding=0, dialation=1 
    sh, sw = conv.stride                # row stride and col stride
    oc, ic, kh, kw = conv.weight.shape  # output_channels, input_channels, row of kernels, col of kernels
    ic, ih, iw = input_shape[-3:]       # batch channel height width or channel height width
    oc, oh, ow = output_shape[-3:]

    W = torch.zeros((oc*oh*ow, ic*ih*iw), dtype=torch.float)
    # we use tensor.view(-1) to flatten input and output, the order is col->row->channel
    for cha_out in range(oc):
         for row_out in range(oh):
            for col_out in range(ow):
                W_row = cha_out*oh*ow + row_out*ow + col_out

                row_in_start = sh*row_out
                col_in_start = sw*col_out
                for cha_in in range(ic):
                    for row_ker in range(kh):
                        W_col1 = cha_in*ih*iw + (row_in_start+row_ker)*iw + col_in_start
                        W_col2 = W_col1 + kw
                        W[W_row, W_col1:W_col2] = conv.weight[cha_out, cha_in, row_ker, :]
    return W

卷积前向传播相关：
链接1

反向传播相关：
推导1
推导2

卷积反向传播代码实现：

import torch
import torch.nn as nn 
import torch.nn.functional as F
from scipy import signal
from utils.Logger import reproduc
if __name__=="__main__":
    reproduc({'seed': 42, 'benchmark': False, 'deterministic': True})
    x = [
        [[1,2,3],[4,5,6],[7,8,9]],
        [[10,11,12],[13,14,15],[16,17,18]],
        [[19,20,21],[22,23,24],[25,26,27]]
             ]
    x = torch.tensor(x, dtype=torch.float)
    x.requires_grad = True
    conv = nn.Conv2d(3, 10, kernel_size=(2,2), stride=(1,1), bias=True)
    y = conv(x)
    loss = y.sum()

    x_grad = torch.autograd.grad(loss, x, grad_outputs=torch.ones_like(loss), create_graph=True)
    y_grad = torch.autograd.grad(loss, y, grad_outputs=torch.ones_like(loss), create_graph=True)
    print(x_grad[0].shape, y_grad[0].shape)
    print(x.shape, conv.weight.shape, y.shape)
    # y = F.conv2d(input=input, weight=conv.weight, bias=conv.bias, stride=conv.stride, padding=conv.padding, dilation=conv.dilation)
    # print(loss)
    W = conv.weight.transpose(0, 1) # in_channels, out_channels
    # W = W.transpose(2, 3)           # fully convolution
    W = torch.flip(W, dims=[-2,-1])
    x1_grad = F.conv2d(input=y_grad[0], weight=W, padding=1)
    res = x_grad[0] - x1_grad
    print(abs(res).max(), abs(x_grad[0]).max(), abs(x1_grad).max())