about upscale==3

[tarxs]

Hello,

The code worked well for me when the upscale factor was a multiple of 2. Thank you for providing these solutions. However, when I changed it to 3, I encountered some issues.

Firstly, in the model.py file, within the_UpsampleBlock class, I noticed that the parameters of nn.PixelShuffle need to be changed from 2 to the upscale_factor.

Furthermore, I encountered a problem related to pixel_loss = pixel_criterion(sr, gt). The error message "The size of tensor a (96) must match the size of tensor b (95) at non-singleton dimension 2" might be due to the image cropping process when the upscale factor is 3. It's possible that the cropping process introduces some floating-point discrepancies. However, the code structure is quite complex, and I'm unsure if I have the ability to fix it on my own.

Could you possibly help me address this issue so that the code works seamlessly with an upscale factor of 3, thank you?

[Lornatang]

Your origin image size is 96x96?

[tarxs]

Your origin image size is 96x96?

train->original's image size is 256x256, after run scripts/run.py, the image in train->GT is 128x128. I think is imgproc.py->random_crop_torch to crop the image to 96

[Lornatang]

Generally speaking, If GT image size is a multiple of 3, there is no problem.

[tarxs]

Generally speaking, If GT image size is a multiple of 3, there is no problem.

I change

lr_top = random.randint(0, lr_image_height - lr_patch_size)
lr_left = random.randint(0, lr_image_width - lr_patch_size)

to

lr_top = random.randint(0, lr_image_height - lr_patch_size - 1)
lr_left = random.randint(0, lr_image_width - lr_patch_size - 1) 

and it can work with scale==3,although i dont know why

[Lornatang]

There is a problem with the dismantling script, which has now been fixed