Why is the output image size of the test different from the input?

[Lkinyuu]

Thank you for your contribution. I am confused why the size of the output result and the input image are different, which is not conducive to the calculation of evaluation metrics such as PI, for example. What modifications should I make if I want to output the same size, thanks!

[Zj-BinXia]

Yes, that's because we need to 8x downsample input images to get images In different scales to achieve scale robustness. So, It seems that codes crop the image size to multiples of 8. This trick is commonly used in the Unet shape networks. If you want to keep the input image size, you can choose to pad the input image to multiples of 8 and crop the padding of the output. You can refer following codes:

`mod_pad_h, mod_pad_w = 0, 0

window_size=8

    _, _, h, w = lq.size()

    if h % window_size != 0:

        mod_pad_h = window_size - h % window_size

    if w % window_size != 0:

        mod_pad_w = window_size - w % window_size

    img = F.pad(lq, (0, mod_pad_w, 0, mod_pad_h), 'reflect')

` Then you can take img as input of the network, and obtain output. And crop the padding as follows:

`, , h, w = self.output.size()

scale=4

   output = self.output[:, :, 0:h - mod_pad_h * scale, 0:w - mod_pad_w * scale]

`

[Lkinyuu]

Yes, that's because we need to 8x downsample input images to get images In different scales to achieve scale robustness. So, It seems that codes crop the image size to multiples of 8. This trick is commonly used in the Unet shape networks. If you want to keep the input image size, you can choose to pad the input image to multiples of 8 and crop the padding of the output. You can refer following codes:

`mod_pad_h, mod_pad_w = 0, 0

window_size=8

    _, _, h, w = lq.size()

    if h % window_size != 0:

        mod_pad_h = window_size - h % window_size

    if w % window_size != 0:

        mod_pad_w = window_size - w % window_size

    img = F.pad(lq, (0, mod_pad_w, 0, mod_pad_h), 'reflect')

` Then you can take img as input of the network, and obtain output. And crop the padding as follows:

`, , h, w = self.output.size()

scale=4

   output = self.output[:, :, 0:h - mod_pad_h * scale, 0:w - mod_pad_w * scale]

`

I am afraid that such a step will not work. In the previous work, the original input image has been made to perform a (scale*8) calculation, and the pad operation with a multiplier of 8 on the input image will naturally be included in the above multiplier.

[Lkinyuu]

I think I solved it, thanks for the answer. I changed the rounding down in mod_crop to rounding up

[Lkinyuu]

By the way, when I test with gan.pth everything works fine, but when I test with mse.pth it suggests that the weights including dcn are missing.

[Lkinyuu]

@Zj-BinXia

[Zj-BinXia]

Can you download the checkpoint from google drive again? It may be the problem of downloading

[Zj-BinXia]

If you use our original code and checkpoint, you can run AMSA-mse and AMSA-GAN successfully

[Lkinyuu]

I am very sure that my download was correct, because after this problem I downloaded again accordingly

[Zj-BinXia]

I do not what's the problem because I can load the checkpoint. Can you use my original code and test it?

using command "python mmsr/test.py -opt "options/test/test_AMSA_mse.yml""?

[Zj-BinXia]

And do you use pytorch=1.4 and install DNCV2?

`python setup.py develop

cd mmsr/models/archs/DCNv2

python setup.py build develop`