yjn870 / rednet-pytorch Goto Github PK

Very nice job!
One little extension concerns the number of channels, ie. RGB=3 but it can nicely be genralized

def init(self, num_layers=5, num_features=64, num_channels=1):

then the 1st conv_layer

nn.Conv2d(num_channels, ...
and the last deconv_layer
deconv_layers.append(nn.ConvTranspose2d(num_features, num_channels,...
Thanks

Hi @yjn870, thank you for the implementation.

It works reasonably well for the most cases.

However, for the images that has some even dimensions such as [370, 545], [663, 962], [359, 478],

the forward pass does not preserve the dimensions at their original form. As a result, I receive the following error :

...
in forward
    x += residual
RuntimeError: The expanded size of the tensor (360) 
              must match the existing size (359) at non-singleton dimension 2
torch.Size([1, 3, 360, 478]) torch.Size([1, 3, 359, 478])

Can you suggest a solution to this problem?

I was confused that you are using MSE loss for the input and output but there is a direct residual adding of original input in the model. Would this model just make the encoder decoder to output 0 to reduce the MSE loss. Hoping you could point me out.
Best regards.

Hello @yjn870, thank you for the implementation.
After training with your code, I found that the psnr and ssim is lower than that of the original paper. Is it caused by different data processing or training methods？

Getting following error during testing. Original image size is 481x321.

File "example.py", line 55, in
pred = model(input)
File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py", line 541, in call
result = self.forward(*input, **kwargs)
File "/content/gdrive/My Drive/rednet/model.py", line 123, in forward
x += residual
RuntimeError: The size of tensor a (482) must match the size of tensor b (481) at non-singleton dimension 3

Hello.
I used the BSD dataset as stated in the paper, and your code for training. I trained for 20 epochs using all of train data, validation data, and test data, but the final loss was around 0.002 ~ 0.003. This is not enough for PSNR to be 30, which is not the quality of the sample output you uploaded.
Can you tell me your learning environment? Thanks in advance.