Grayscale dataset

[Elwarfalli]

I'd like to ask about your great work.

Is it possible to run it on a grayscale dataset? If so what should I change? 
I changed a number of input channels but it is not working for me.

network structuresnetwork_g: 

type: HAT 
upscale: 3 
in_chans: 1

I am looking forward to hearing back from you.
Thank you,

[chxy95]

@Elwarfalli You may directly input the gray images without any change and you would get "gray images" with three channels. Then you may simply convert the three-channel images to one-channel ones.

[Elwarfalli]

When I applied that where the input is grayscale and I tested my pre-trained model the results were saved in 3 channels of grayscale. I am curious about PSNR; has PSNR computed for one grayscale channel? or how? and how can I change the code to save the results on a one-channel grayscale?

Thank you for your fast reply,

[chxy95]

@Elwarfalli PSNR is comupted on Y channel in the default settings. If you want to compute PSNR for comparing models on gray images, the best way is to retrain a model for gray images by setting the channel as 1 for complete fairness. If you just want to see the results of pretrained model on gray images, the most convenient way is to wirte a post-processing script to convert the three-channel images into gray channel like using Opencv BGR2Gray, and then compute PSNR on the one-channel images.

[Elwarfalli]

Thank you,

[Elwarfalli]

When I set the parameter as follows:

network structures

network_g:
type: HAT
upscale: 3
in_chans: 1
img_size: 64
window_size: 16
compress_ratio: 3
squeeze_factor: 30
conv_scale: 0.01
overlap_ratio: 0.5
img_range: 1.
depths: [6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6]
embed_dim: 180
num_heads: [6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6]
mlp_ratio: 2
upsampler: 'pixelshuffle'
resi_connection: '1conv'

I got an error:

output = module(*input, **kwargs)

File "C:\Users\Student.conda\envs\hat\lib\site-packages\torch\nn\modules\module.py", line 1194, in _call_impl
return forward_call(*input, **kwargs)
File "e:\hamed\pycode\benchmarks\hat-master\hat\archs\hat_arch.py", line 980, in forward
x = self.conv_first(x)
File "C:\Users\Student.conda\envs\hat\lib\site-packages\torch\nn\modules\module.py", line 1194, in _call_impl
return forward_call(*input, **kwargs)
File "C:\Users\Student.conda\envs\hat\lib\site-packages\torch\nn\modules\conv.py", line 463, in forward
return self._conv_forward(input, self.weight, self.bias)
File "C:\Users\Student.conda\envs\hat\lib\site-packages\torch\nn\modules\conv.py", line 459, in _conv_forward
return F.conv2d(input, weight, bias, self.stride,
RuntimeError: Given groups=1, weight of size [180, 1, 3, 3], expected input[1, 3, 16, 16] to have 1 channels, but got 3 channels instead

Any help, please?

Thank you,

[chxy95]

@Elwarfalli You need modify the data loader for 1-channel images I/O. Create a custom data loader referring to paired_image_dataset.py.

[Elwarfalli]

Based on my understating, BasicSR paired_image_dataset.py papers the dataset. I have my own grayscale dataset training/validation.

[chxy95]

@Elwarfalli I mean you need to create your own data loader to fit your dataset. Read this file just as the reference.