我是 Pytorch 的新手，正在为我的 Django 应用程序使用预先训练的模型来恢复损坏的图像。我从这个存储库中给出的代码中获得了帮助。
我的 ML 模型类的代码如下：
class Hourglass(nn.Module):
def __init__(self):
super(Hourglass, self).__init__()
self.leaky_relu = nn.LeakyReLU()
self.d_conv_1 = nn.Conv2d(2, 8, 5, stride=2, padding=2)
self.d_bn_1 = nn.BatchNorm2d(8)
self.d_conv_2 = nn.Conv2d(8, 16, 5, stride=2, padding=2)
self.d_bn_2 = nn.BatchNorm2d(16)
self.d_conv_3 = nn.Conv2d(16, 32, 5, stride=2, padding=2)
self.d_bn_3 = nn.BatchNorm2d(32)
self.s_conv_3 = nn.Conv2d(32, 4, 5, stride=1, padding=2)
self.d_conv_4 = nn.Conv2d(32, 64, 5, stride=2, padding=2)
self.d_bn_4 = nn.BatchNorm2d(64)
self.s_conv_4 = nn.Conv2d(64, 4, 5, stride=1, padding=2)
self.d_conv_5 = nn.Conv2d(64, 128, 5, stride=2, padding=2)
self.d_bn_5 = nn.BatchNorm2d(128)
self.s_conv_5 = nn.Conv2d(128, 4, 5, stride=1, padding=2)
self.d_conv_6 = nn.Conv2d(128, 256, 5, stride=2, padding=2)
self.d_bn_6 = nn.BatchNorm2d(256)
self.u_deconv_5 = nn.ConvTranspose2d(256, 64, 4, stride=2, padding=1)
self.u_bn_5 = nn.BatchNorm2d(64)
self.u_deconv_4 = nn.ConvTranspose2d(128, 32, 4, stride=2, padding=1)
self.u_bn_4 = nn.BatchNorm2d(32)
self.u_deconv_3 = nn.ConvTranspose2d(64, 16, 4, stride=2, padding=1)
self.u_bn_3 = nn.BatchNorm2d(16)
self.u_deconv_2 = nn.ConvTranspose2d(32, 8, 4, stride=2, padding=1)
self.u_bn_2 = nn.BatchNorm2d(8)
self.u_deconv_1 = nn.ConvTranspose2d(16, 4, 4, stride=2, padding=1)
self.u_bn_1 = nn.BatchNorm2d(4)
self.out_deconv = nn.ConvTranspose2d(4, 4, 4, stride=2, padding=1)
self.out_bn = nn.BatchNorm2d(4)
def forward(self, noise):
down_1 = self.d_conv_1(noise)
down_1 = self.d_bn_1(down_1)
down_1 = self.leaky_relu(down_1)
down_2 = self.d_conv_2(down_1)
down_2 = self.d_bn_2(down_2)
down_2 = self.leaky_relu(down_2)
down_3 = self.d_conv_3(down_2)
down_3 = self.d_bn_3(down_3)
down_3 = self.leaky_relu(down_3)
skip_3 = self.s_conv_3(down_3)
down_4 = self.d_conv_4(down_3)
down_4 = self.d_bn_4(down_4)
down_4 = self.leaky_relu(down_4)
skip_4 = self.s_conv_4(down_4)
down_5 = self.d_conv_5(down_4)
down_5 = self.d_bn_5(down_5)
down_5 = self.leaky_relu(down_5)
skip_5 = self.s_conv_5(down_5)
down_6 = self.d_conv_6(down_5)
down_6 = self.d_bn_6(down_6)
down_6 = self.leaky_relu(down_6)
up_5 = self.u_deconv_5(down_6)
up_5 = torch.cat([up_5, skip_5], 1)
up_5 = self.u_bn_5(up_5)
up_5 = self.leaky_relu(up_5)
up_4 = self.u_deconv_4(up_5)
up_4 = torch.cat([up_4, skip_4], 1)
up_4 = self.u_bn_4(up_4)
up_4 = self.leaky_relu(up_4)
up_3 = self.u_deconv_3(up_4)
up_3 = torch.cat([up_3, skip_3], 1)
up_3 = self.u_bn_3(up_3)
up_3 = self.leaky_relu(up_3)
up_2 = self.u_deconv_2(up_3)
up_2 = self.u_bn_2(up_2)
up_2 = self.leaky_relu(up_2)
up_1 = self.u_deconv_1(up_2)
up_1 = self.u_bn_1(up_1)
up_1 = self.leaky_relu(up_1)
out = self.out_deconv(up_1)
out = self.out_bn(out)
out = nn.Sigmoid()(out)
return out
我正在使用它来恢复图像，如下所示：
def restore_image_deep_image_prior(original_image_path):
lr = 1e-2
device = 'cpu'
print('Using {} for computation'.format(device.upper()))
hg_net = Hourglass()
hg_net.to(device)
mse = nn.MSELoss()
optimizer = optim.Adam(hg_net.parameters(), lr=lr)
n_iter = 500
images = []
losses = []
to_tensor = tv.transforms.ToTensor()
z = torch.Tensor(np.mgrid[:542, :347]).unsqueeze(0).to(device) / 512 # Adjust the size according to your requirement
x = PILImage.open(original_image_path)
x = to_tensor(x).unsqueeze(0)
x, mask = pixel_thanos(x, 0.5)
mask = mask[:, :3, :, :].to(device).float() # Keep only the first 3 channels if mask has 4 channels
x = x.to(device)
for i in range(n_iter):
optimizer.zero_grad()
y = hg_net(z)
loss = mse(x, y*mask)
losses.append(loss.item())
loss.backward()
optimizer.step()
if i < 1000 and (i+1)%4==0 or i==0:
with torch.no_grad():
out = x + y * (1 - mask)
out = out[0].cpu().detach().permute(1, 2, 0)*255
out = np.array(out, np.uint8)
images.append(out)
if (i+1) % 20 == 0:
print('Iteration: {} Loss: {:.07f}'.format(i+1, losses[-1]))
restored_image_bytes = image_to_bytes(images)
return restored_image_bytes
def pixel_thanos(img, p=0.5):
assert p > 0 and p < 1, 'The probability value should lie in (0, 1)'
mask = torch.rand(1, 3, 542, 347)
img[mask < p,] = 0
mask = mask > p
mask = mask.repeat(1, 3, 1, 1)
return img, mask
def image_to_bytes(image):
buffer = BytesIO()
image.save(buffer, format="JPEG")
return buffer.getvalue()
现在，问题是我在标题中也遇到了错误监听，详细来说，是：
Traceback (most recent call last):
File "C:\Users\khubi\AppData\Local\Programs\Python\Python39\lib\site-packages\django\core\handlers\exception.py", line 55, in inner
response = get_response(request)
File "C:\Users\khubi\AppData\Local\Programs\Python\Python39\lib\site-packages\django\core\handlers\base.py", line 197, in _get_response
response = wrapped_callback(request, *callback_args, **callback_kwargs)
File "C:\Users\khubi\Desktop\Projects\VU\CS619\image restorer\image_restorer\image_app\views.py", line 35, in restore_image
restored_image = restore_image_deep_image_prior(original_image.image_file.path)
File "C:\Users\khubi\Desktop\Projects\VU\CS619\image restorer\image_restorer\image_app\views.py", line 76, in restore_image_deep_image_prior
y = hg_net(z)
File "C:\Users\khubi\AppData\Local\Programs\Python\Python39\lib\site-packages\torch\nn\modules\module.py", line 1518, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
File "C:\Users\khubi\AppData\Local\Programs\Python\Python39\lib\site-packages\torch\nn\modules\module.py", line 1527, in _call_impl
return forward_call(*args, **kwargs)
File "C:\Users\khubi\Desktop\Projects\VU\CS619\image restorer\image_restorer\image_app\views.py", line 182, in forward
up_5 = torch.cat([up_5, skip_5], 1)
RuntimeError: Sizes of tensors must match except in dimension 1. Expected size 18 but got size 17 for tensor number 1 in the list.
我知道我必须设置频道。我将其设置为 4，因为我用于恢复的图像有 4 个通道。但仍然收到错误。怎么解决？
主要问题来自模型的架构和输入图像的大小。您可以打印每个
nn.Conv2dnn.ConvTranspose2d根据您的代码，预期输出和错误消息如下：
预期输出：
input: torch.Size([1, 2, 542, 347])
down_1: torch.Size([1, 8, 271, 174])
down_2: torch.Size([1, 16, 136, 87])
skip_3: torch.Size([1, 4, 68, 44])
skip_4: torch.Size([1, 4, 34, 22])
skip_5: torch.Size([1, 4, 17, 11])
down_6: torch.Size([1, 256, 9, 6])
错误信息：
RuntimeError: Sizes of tensors must match except in dimension 1. Expected size 18 but got size 17 for tensor number 1 in the list.
问题源于这行代码：
up_5 = torch.cat([up_5, skip_5], 1)
在此操作中，
up_5[1, 64, 18, 12]skip_5[1, 4, 17, 11](18, 12)(17, 11)解决此问题的一种方法是调整输入数据形状。由于您的
nn.Conv2dnn.Conv2d[1, 2, 512, 320]但是，将输入大小更改为
[1, 2, 512, 320]RuntimeError: running_mean should contain 68 elements not 64
在具有更新大小
torch.cat([up_5, skip_5], 1)[1, 64, 16, 10][1, 4, 16, 10][1, 68, 16, 10]self.u_bn_5 = nn.BatchNorm2d(64)
其中预期特征大小为 68，但模块输入大小设置为 64。因此，将
self.u_deconv_560 + 4 = 64尽管如此，以下代码中出现了另一个问题：
self.u_deconv_4 = nn.ConvTranspose2d(128, 32, 4, stride=2, padding=1)
这里，输入通道大小 128 仍然与之前的特征大小 64 不匹配。
针对这些问题，建议修改后的模型架构如下，供大家参考：
class Hourglass(nn.Module):
def __init__(self):
super(Hourglass, self).__init__()
self.leaky_relu = nn.LeakyReLU()
self.d_conv_1 = nn.Conv2d(2, 8, 5, stride=2, padding=2)
self.d_bn_1 = nn.BatchNorm2d(8)
self.d_conv_2 = nn.Conv2d(8, 16, 5, stride=2, padding=2)
self.d_bn_2 = nn.BatchNorm2d(16)
self.d_conv_3 = nn.Conv2d(16, 32, 5, stride=2, padding=2)
self.d_bn_3 = nn.BatchNorm2d(32)
self.s_conv_3 = nn.Conv2d(32, 4, 5, stride=1, padding=2)
self.d_conv_4 = nn.Conv2d(32, 64, 5, stride=2, padding=2)
self.d_bn_4 = nn.BatchNorm2d(64)
self.s_conv_4 = nn.Conv2d(64, 4, 5, stride=1, padding=2)
self.d_conv_5 = nn.Conv2d(64, 128, 5, stride=2, padding=2)
self.d_bn_5 = nn.BatchNorm2d(128)
self.s_conv_5 = nn.Conv2d(128, 4, 5, stride=1, padding=2)
self.d_conv_6 = nn.Conv2d(128, 256, 5, stride=2, padding=2)
self.d_bn_6 = nn.BatchNorm2d(256)
self.u_deconv_5 = nn.ConvTranspose2d(256, 124, 4, stride=2, padding=1)
self.u_bn_5 = nn.BatchNorm2d(128)
self.u_deconv_4 = nn.ConvTranspose2d(128, 60, 4, stride=2, padding=1)
self.u_bn_4 = nn.BatchNorm2d(64)
self.u_deconv_3 = nn.ConvTranspose2d(64, 28, 4, stride=2, padding=1)
self.u_bn_3 = nn.BatchNorm2d(32)
self.u_deconv_2 = nn.ConvTranspose2d(32, 16, 4, stride=2, padding=1)
self.u_bn_2 = nn.BatchNorm2d(16)
self.u_deconv_1 = nn.ConvTranspose2d(16, 8, 4, stride=2, padding=1)
self.u_bn_1 = nn.BatchNorm2d(8)
self.out_deconv = nn.ConvTranspose2d(8, 4, 4, stride=2, padding=1)
self.out_bn = nn.BatchNorm2d(4)
def forward(self, noise):
print("input:",noise.size())
down_1 = self.d_conv_1(noise)
down_1 = self.d_bn_1(down_1)
down_1 = self.leaky_relu(down_1)
print("down_1:",down_1.size())
down_2 = self.d_conv_2(down_1)
down_2 = self.d_bn_2(down_2)
down_2 = self.leaky_relu(down_2)
print("down_2:",down_2.size())
down_3 = self.d_conv_3(down_2)
down_3 = self.d_bn_3(down_3)
down_3 = self.leaky_relu(down_3)
skip_3 = self.s_conv_3(down_3)
print("skip_3:",skip_3.size())
down_4 = self.d_conv_4(down_3)
down_4 = self.d_bn_4(down_4)
down_4 = self.leaky_relu(down_4)
skip_4 = self.s_conv_4(down_4)
print("skip_4:",skip_4.size())
down_5 = self.d_conv_5(down_4)
down_5 = self.d_bn_5(down_5)
down_5 = self.leaky_relu(down_5)
skip_5 = self.s_conv_5(down_5)
print("skip_5:",skip_5.size())
down_6 = self.d_conv_6(down_5)
down_6 = self.d_bn_6(down_6)
down_6 = self.leaky_relu(down_6)
print("down_6:",down_6.size())
up_5 = self.u_deconv_5(down_6)
up_5 = torch.cat([up_5, skip_5], 1)
up_5 = self.u_bn_5(up_5)
up_5 = self.leaky_relu(up_5)
print("up_5:",up_5.size())
up_4 = self.u_deconv_4(up_5)
up_4 = torch.cat([up_4, skip_4], 1)
up_4 = self.u_bn_4(up_4)
up_4 = self.leaky_relu(up_4)
print("up_4",up_4.size())
up_3 = self.u_deconv_3(up_4)
up_3 = torch.cat([up_3, skip_3], 1)
up_3 = self.u_bn_3(up_3)
up_3 = self.leaky_relu(up_3)
print("up_3",up_3.size())
up_2 = self.u_deconv_2(up_3)
up_2 = self.u_bn_2(up_2)
up_2 = self.leaky_relu(up_2)
print("up_2",up_2.size())
up_1 = self.u_deconv_1(up_2)
up_1 = self.u_bn_1(up_1)
up_1 = self.leaky_relu(up_1)
print("up_1",up_1.size())
out = self.out_deconv(up_1)
out = self.out_bn(out)
out = nn.Sigmoid()(out)
print("out",out.size())
return out
model = Hourglass()
x = torch.Tensor(np.mgrid[:512, :320]).unsqueeze(0)
y = model(x)
您可以将修改后的架构合并到您的代码中，以解决大小不一致的问题。