yjn870/espcn-pytorch

PyTorch implementation of Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional Neural Network (CVPR 2016)

espcn-pytorch's Issues

Some questions...

Training with different scale factor results in error

python train.py --train-file "/BLAH_BLAH/SR/91-image_x3.h5" --eval-file "/BLAH_BLAHSR/Set5_x3.h5" --outputs-dir "/BLAH_BLAH/model_run/SR_outputs" --scale 2 --lr 1e-3 --batch-size 16 --num-epochs 200 --num-workers 8 --seed 123

epoch: 0/199: 0%| | 0/2688 [00:00<?, ?it/s]

***/code/ESPCN-pytorch/venv/lib/python3.7/site-packages/torch/nn/modules/loss.py:445: UserWarning: Using a target size (torch.Size([16, 1, 51, 51])) that is different to the input size (torch.Size([16, 1, 34, 34])). This will likely lead to incorrect results due to broadcasting. Please ensure they have the same size.
return F.mse_loss(input, target, reduction=self.reduction)

Traceback (most recent call last):
File "train.py", line 79, in
loss = criterion(preds, labels)
File "//code/ESPCN-pytorch/venv/lib/python3.7/site-packages/torch/nn/modules/module.py", line 722, in _call_impl
result = self.forward(input, kwargs)
File "//code/ESPCN-pytorch/venv/lib/python3.7/site-packages/torch/nn/modules/loss.py", line 445, in forward
return F.mse_loss(input, target, reduction=self.reduction)
File "//code/ESPCN-pytorch/venv/lib/python3.7/site-packages/torch/nn/functional.py", line 2647, in mse_loss
expanded_input, expanded_target = torch.broadcast_tensors(input, target)
File "/***/code/ESPCN-pytorch/venv/lib/python3.7/site-packages/torch/functional.py", line 65, in broadcast_tensors
return _VF.broadcast_tensors(tensors)
RuntimeError: The size of tensor a (34) must match the size of tensor b (51) at non-singleton dimension 3

I assume that somewhere in the .h5 file there is a hard-coded image that was created using a scale factor of 3. Is that correct? Because it seems that some images are not scaling properly with the dataset provided in the README.md

Where is the gaussian filter?

In the paper I read

"To synthesize the low-resolution samples LR, we blur HR using a Gaussian filter and sub-sample it by the upscaling factor."

But I cannot find the line in this repository where you apply a Gaussian filter. Or did I miss something?

No increase in resolution

Hi, I ran the inference on a test image

And used the pre-trained weights provided by you.
I got 2 results:

and

But I got no improvement. That is, the resolution of all the images remained almost same: 579x1032 pixels for all 3 images.

What went wrong ?

.h5 datasets could not download.

The error occured in train.py

I want to scale up x2 not x3.

So i need to use prepare.py.

And i made 2 files. One is Set5_x2.h5 for evaluating, and the other is Set14_x2.h5 for trainning.

But I have a problem when I train it.

And the error message is "ValueError: Field names only allowed for compound types".
(The error occured at line 95(for data in eval_dataloader:)

The command is python train.py --train-file "Set14_x2.h5" --eval-file "Set5_x2.h5"

Could you give me some advise?