No need to resize input image

[MasKogNavier]

I tried issue #10 because I don't need to resize the input image, but it doesn't work.
And i tried commented out the PIL.resize part
ex) l.129 in ref_cufed_dataset.py
before:
img_in_lq = img_in_pil.resize((lq_w, lq_h), Image.BICUBIC)
after:
img_in_lq = img_in_pil#.resize((lq_w, lq_h), Image.BICUBIC)
I got below error.(input and ref image size are 128x128)

RuntimeError: invalid argument 0: Sizes of tensors must match except in dimension 1. Got 32 and 128 in dimension 2 at /opt/conda/conda-bld/pytorch_1579040055865/work/aten/src/THC/generic/THCTensorMath.cu:71
How do I use an input image of my own defined size(no need to resize image)?

[yumingj]

Hi, thanks for your interest in our work!

Can you have a look at the size of img_in_up? It should have the same size as img_ref_up.

[MasKogNavier]

Thank you for your reply.
With the current settings, I intend to super-resolution 128x128 input image (img_in_lr) to x4.
As a confirmation, if the output is 512x512, does it mean that 512x512 should be prepared for img_ref?

[yumingj]

Yes. In the current implementation, if the reference images are smaller than 512x512, we will do the padding operation to make the reference images have that resolution.