Perceptual Loss computation

[rogertrullo]

Hi, thanks for sharing your code. I was reading the implementation of the PerceptualLoss and I am wondering why only the first image of the batch is normalized: https://github.com/VITA-Group/DeblurGANv2/blob/c5bbf780b152d5b877d9911b8323898d5f2e0a50/models/losses.py#L50-L51

Is the code intended to only work with batchsize=1?

[diyar-m]

Hi @rogertrullo, have you found the answer?

[rogertrullo]

Hi @rogertrullo, have you found the answer?

Hi @diyar-m although I never got a reply, I assumed it was a bug. I made some mods to the code but I haven’t tested it

[diyar-m]

I'm gonna change it and test it too.

I'll let you know if I'd got a better result.

[rogertrullo]

Great! Let me know how it went once you have it :)

[diyar-m]

I got this error message. @rogertrullo

Traceback (most recent call last): | File "train.py", line 256, in | trainer.train(resume_train=config['resume']['resume_training']) | File "train.py", line 56, in train | self._run_epoch(epoch) | File "train.py", line 125, in _run_epoch | loss_content = self.criterionG(outputs, targets) | File "/home/avir/Desktop/DeblurGANv2/models/losses.py", line 59, in call | return self.get_loss(fakeIm, realIm) | File "/home/avir/Desktop/DeblurGANv2/models/losses.py", line 50, in get_loss | fakeIm[:, :, :, :] = self.transform(fakeIm[:, :, :, :]) | File "/home/avir/Desktop/DeblurGANv2/.env/lib/python3.6/site-packages/torchvision/transforms/trans| forms.py", line 212, in call | return F.normalize(tensor, self.mean, self.std, self.inplace) | File "/home/avir/Desktop/DeblurGANv2/.env/lib/python3.6/site-packages/torchvision/transforms/funct| ional.py", line 284, in normalize | '{}.'.format(tensor.size())) | ValueError: Expected tensor to be a tensor image of size (C, H, W). Got tensor.size() = torch.Size([| 6, 3, 256, 256]).

[rogertrullo]

I dont have the code at hand now, but yeah the thing is that the transform expects individial images and not a batch of images, I cant remember exactly how I handled it but it is definetly doable

[diyar-m]

Yeah. I'm using a for to iterate between images and transform each one individially.

[rogertrullo]

yeah I think I did the same with a map or sth but escentially the same

[ldlshizhu]

是的，我想我在地图或某物上做了相同的操作，但在外观上却是相同的 你好，请问您是如何修改的代码