Error during learning process

[hw1204]

I found this error while running finetune_dualstylegan.py in the last step of learning.
I changed it to jpg because I thought the format of the picture might be a problem, but there must be a problem.
I would appreciate it if you could tell me how to solve it.

load model: /content/drive/MyDrive/generator-pretrain.pt
Loading pSp from checkpoint: /content/drive/MyDrive/encoder.pt
Loading ResNet ArcFace
Encoder model successfully loaded!
Traceback (most recent call last):
File "finetune_dualstylegan.py", line 529, in
Simg = transform2(Simg).unsqueeze(dim=3)
File "/usr/local/lib/python3.7/dist-packages/torchvision/transforms/transforms.py", line 67, in call
img = t(img)
File "/usr/local/lib/python3.7/dist-packages/torch/nn/modules/module.py", line 727, in call_impl
result = self.forward(*input, **kwargs)
File "/usr/local/lib/python3.7/dist-packages/torchvision/transforms/transforms.py", line 226, in forward
return F.normalize(tensor, self.mean, self.std, self.inplace)
File "/usr/local/lib/python3.7/dist-packages/torchvision/transforms/functional.py", line 284, in normalize
tensor.sub(mean).div_(std)
RuntimeError: The size of tensor a (4) must match the size of tensor b (3) at non-singleton dimension 0

[williamyang1991]

I changed it to jpg because I thought the format of the picture might be a problem, but there must be a problem.

So the error occurs when you use jpg or png?
The code uses mean and std to adjust 3 channels of your data.

RuntimeError: The size of tensor a (4) must match the size of tensor b (3) at non-singleton dimension 0

It seems that your data has 4 channels.

    transform2 = transforms.Compose(
    [
        transforms.ToTensor(),
        transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5), inplace=True),
    ])     

File "finetune_dualstylegan.py", line 529, in
Simg = transform2(Simg).unsqueeze(dim=3)

My original code is Simg = transform2(Simg).unsqueeze(dim=0)
Did you make modifications to the code, which leads to the error?

[hw1204]

@williamyang1991 It was originally a png format, but I changed it to jpg, but there is still an error. I tried your original code, but it's the same! Do you think the png format is a problem?

[williamyang1991]

you can add
print(np.array(Simg).shape) before
Simg = transform2(Simg).unsqueeze(dim=0)
to see the shape of your Simg.
It should be (1024, 1024, 3)
otherwise, there is something wrong with your style image.

[hw1204]

Alpha Channel was the problem! Thanks for letting me know. One question is that input image is definitely channel 3, but Simg is still channel 4. Can I assume there was a malfunction in the middle?

[williamyang1991]

My code is just read the image and transform it.

Simg = Image.open(os.path.join(path, filename))
Simg = transform2(Simg).unsqueeze(dim=0)

It was originally a png format, but I changed it to jpg

It should be that you incorrectly changed png to jpg without discarding its alpah channel.