Validation in training loop - dimensionality problems

[sptom]

Hi! I uncommented the "eval" part in the training loop, and want to use it for validation during training (with color images). However, I get one of two errors:

Because of the line: noise_map_val = np.zeros((1, 2*c, img_w, img_h))

In the code The marked line causes a dimension error, as it tries to input a 3 channel image into a 6 channel tensor (as defined above with 2*c).

Otherwise, if I change the above line to noise_map_val = np.zeros((1, 1*c, img_w, img_h))

The line out_val_nb = torch.clamp( imgn_val - model(imgn_val, NM_tensor_val), 0., 1.) Fails, because NM_tensor val has 3 channels and not 6.

What is the right way to input 6 channels into the generated noise map from 3 channels?

Thanks, Tom

[yzhouas]

Hi Tom, the codes are very old so I am not sure whether my answer will address your problem. Basically, in the paper we used two noise types: AWGN and RVIN, so the noise map contains 6 channels (RGB for AWGN and RGB for RVIN).

I guess the commented validation codes are deprecated and it only generates one single type AWGN noise levels. It's better to refer https://github.com/yzhouas/PD-Denoising-pytorch/blob/master/train.py#L111-L133 to generate the validation noise level maps.

The codes are not clean enough and sorry for causing so many troubles.

[sptom]

Hi Yuqian,

first of all thanks for the quick reply!

I replaced the deprecated commands with newer versions. In the meantime I just replicated the 3 channel generated noise map twice so it fits the 6 channels, and the code runs now. I'll review the lines you pointed to and try to adjust.

BTW, I see that "parepare_data" can create a validation data loader as well. I guess that if I want to use it for validation instead of just running on files in the val folder I'll have to write the validation loop myself, right?

Thanks again, Tom