batchsize=2，error

[lrn1314]

Hello,

Thank you for such a great job and for releasing your code. I want to train your network using a custom dataset. When I looked at the options.py file, the batch_size parameter is set to 1 and has not changed anywhere in other training files.

batchSize=2，Traceback (most recent call last): File "E:\LBP_idea\gabor+dense\gabor-test-dense\main.py", line 234, in train() File "E:\LBP_idea\gabor+dense\gabor-test-dense\main.py", line 151, in train I_g, I_o, loss_G = model.optimize_parameters() File "E:\LBP_idea\gabor+dense\gabor-test-dense\model.py", line 291, in optimize_parameters self.forward() File "E:\LBPidea\gabor+dense\gabor-test-dense\model.py", line 186, in forward , self.I_FEA = self.netG(self.I_g, self.L_g, self.mask) File "D:\Anaconda\envs\pytorch_3.8\lib\site-packages\torch\nn\modules\module.py", line 1501, in _call_impl return forward_call(*args, kwargs) File "E:\LBP_idea\gabor+dense\gabor-test-dense\networks.py", line 265, in forward up31 = self.myattention(up31[:, :256, :, :], up31[:, 256:, :, :], mask) File "D:\Anaconda\envs\pytorch_3.8\lib\site-packages\torch\nn\modules\module.py", line 1501, in _call_impl return forward_call(*args, *kwargs) File "E:\LBP_idea\gabor+dense\gabor-test-dense\attention.py", line 13, in forward return MyAttentionFunction.apply(generated, known, mask, self.device) File "D:\Anaconda\envs\pytorch_3.8\lib\site-packages\torch\autograd\function.py", line 506, in apply return super().apply(args, kwargs) # type: ignore[misc] File "E:\LBP_idea\gabor+dense\gabor-test-dense\attention.py", line 27, in forward patches = patches_all[ctx.flag == 1] IndexError: The shape of the mask [1, 2048] at index 0 does not match the shape of the indexed tensor [2, 1024, 256] at index 0

Process finished with exit code 1

[HighwayWu]

Thanks for your question. The attention module is designed for batch size = 1, and may not applied to > 2. This is because the number of known and unknown patches is different for each image, and they can't be processed in group. An additional 'for' loop may handle this problem.

[lrn1314]

Hello, thank you for your answer. Because I am a student who has just completed my graduate exam, your article has helped me a lot. How should I write an extra ‘for’ loop? For some basic training details, such as the celebA-HQ dataset of 27,000 images, what should I do if there are not so many irregular masks? Thank you very much for your prompt answer.

[HighwayWu]

The current attention is an inner loop, and the outer loop should obtain the feature for each input, and after processing, the features should be merged together.

BTW, I suggest you to ignore the attention module at beginning, and only train the standard U-Net.

[lrn1314]

Hello, very sorry to ask you again, how should this extra "for" loop be written? Where to add?

[Rahaf37]

Hello, how I should ignore the attention module?