While I modified the unet, actually speaking, I just add a gaussian layer in the net, the code is as blow.
But when I run this code, it tells me that
RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation
So what is wrong with code?
# -*-coding:utf-8-*-
import torch
import torch.nn as nn
from torch.nn import init
import functools
import cv2
import numpy as np
from torch.optim import lr_scheduler
class unet256_2(nn.Module):
def __init__(self, input_nc, output_nc, num_downs, ngf,
norm_layer=nn.BatchNorm2d, use_dropout=False):
super(unet256_2, self).__init__()
# construct unet structure
unet_block = UnetSkipConnectionBlock(ngf * 8, ngf * 8, input_nc=None, submodule=None,
norm_layer=norm_layer, innermost=True) # 2*2
unet_block = UnetSkipConnectionBlock(ngf * 8, ngf * 8, input_nc=None, submodule=unet_block,
norm_layer=norm_layer, use_dropout=use_dropout, sigma=1, size=4) # 4*4
unet_block = UnetSkipConnectionBlock(ngf * 8, ngf * 8, input_nc=None, submodule=unet_block,
norm_layer=norm_layer, use_dropout=use_dropout, sigma=1, size=8) # 8*8
unet_block = UnetSkipConnectionBlock(ngf * 8, ngf * 8, input_nc=None, submodule=unet_block,
norm_layer=norm_layer, use_dropout=use_dropout, sigma=1, size=16) # 16*16
unet_block = UnetSkipConnectionBlock(ngf * 4, ngf * 8, input_nc=None, submodule=unet_block, norm_layer=norm_layer) # 32*32
unet_block = UnetSkipConnectionBlock(ngf * 2, ngf * 4, input_nc=None, submodule=unet_block, norm_layer=norm_layer)
unet_block = UnetSkipConnectionBlock(ngf, ngf * 2, input_nc=None, submodule=unet_block, norm_layer=norm_layer)
unet_block = UnetSkipConnectionBlock(output_nc, ngf, input_nc=input_nc, submodule=unet_block, outermost=True, norm_layer=norm_layer)
self.model = unet_block
def forward(self, input):
return self.model(input)
# Defines the submodule with skip connection.
# X -------------------identity---------------------- X
# |-- downsampling -- |submodule| -- upsampling --|
class UnetSkipConnectionBlock(nn.Module):
def __init__(self, outer_nc, inner_nc, input_nc=None,
submodule=None, outermost=False, innermost=False, norm_layer=nn.BatchNorm2d, use_dropout=False, sigma=None, size=None):
super(UnetSkipConnectionBlock, self).__init__()
self.outermost = outermost
# 对高斯layer进行声明
self.sigma = sigma
self.size = size
self.gaussian_layer=None
if sigma is not None:
self.gaussian_layer = Gaussian_filter(kernel_size=self.size, sigma=self.sigma)
# self.gaussian_layer = nn.Sequential(*[self.gaussian_layer])
if type(norm_layer) == functools.partial:
use_bias = norm_layer.func == nn.InstanceNorm2d
else:
use_bias = norm_layer == nn.InstanceNorm2d
if input_nc is None:
input_nc = outer_nc
downconv = nn.Conv2d(input_nc, inner_nc, kernel_size=4,
stride=2, padding=1, bias=use_bias)
downrelu = nn.LeakyReLU(0.2, True)
downnorm = norm_layer(inner_nc)
uprelu = nn.ReLU(True)
upnorm = norm_layer(outer_nc)
if outermost: # 如果是U-NET网络结构中的最外层部分
upconv = nn.ConvTranspose2d(inner_nc * 2, outer_nc,
kernel_size=4, stride=2,
padding=1)
down = [downconv]
up = [uprelu, upconv, nn.Tanh()]
model = down + [submodule] + up
elif innermost: # 如果是U-NET网络结构中的最低部
upconv = nn.ConvTranspose2d(inner_nc, outer_nc,
kernel_size=4, stride=2,
padding=1, bias=use_bias)
down = [downrelu, downconv]
up = [uprelu, upconv, upnorm]
model = down + up
else: # U-NET网络结构的中间部分
upconv = nn.ConvTranspose2d(inner_nc * 2, outer_nc,
kernel_size=4, stride=2,
padding=1, bias=use_bias)
down = [downrelu, downconv, downnorm]
up = [uprelu, upconv, upnorm]
if use_dropout:
model = down + [submodule] + up + [nn.Dropout(0.5)]
else:
model = down + [submodule] + up
self.model = nn.Sequential(*model)
def forward(self, x):
if self.outermost:
return self.model(x)
elif self.sigma is not None:
y = self.gaussian_layer(x)
z = self.model(x)
return torch.cat([y, z], 1) # 在第一个维度上将x与model(x)进行cat一起,向上up sampling
else:
return torch.cat([x, self.model(x)], 1)
class Gaussian_filter(nn.Module):
def __init__(self, kernel_size=8, sigma=1):
super(Gaussian_filter, self).__init__()
# 进行高斯初始化
k1 = cv2.getGaussianKernel(kernel_size, sigma)
k2 = cv2.getGaussianKernel(kernel_size, sigma)
n = int(kernel_size / 2)
K = np.dot(k1, np.transpose(k2))
N = np.ones((kernel_size, kernel_size))
K2 = K[n, n] * N - K
torch_data = torch.FloatTensor(K2)
self.gaussian_filter = nn.Parameter(torch_data, requires_grad=True)
def forward(self, x):
# print(x)
# x = self.test(x)
# print(x)
return torch.mul(x, self.gaussian_filter)
if __name__ =='__main__':
data = torch.rand(1,3,256,256)
label = torch.rand(1,3,256,256)
net = unet256_2(input_nc=3, output_nc=3, num_downs=8, ngf=64, norm_layer=nn.BatchNorm2d, use_dropout=True)
# net = Gaussian_filter()
output = net(data)
loss = torch.nn.L1Loss()
tmp_loss = loss(output, data)
optimizer = torch.optim.Adam(net.parameters(), lr=0.1)
optimizer.zero_grad()
tmp_loss.backward()
optimizer.step()