[Discriminator] Forward() results in 'NoneType' object has no attribute 'dtype'

[LeWentz]

When training a new dataset using the train.py likes this on Windows 11 in a conda env:

python train.py --outdir="C:\\Models\\styleGANv3" --cfg=stylegan3-t --data="C:\\Datasets\\ABC\\" --batch=32 --gpus=1 --gamma=8.2 --aug=noaug

I get an Error on the Forward() of the Discriminator.

'NoneType' object has no attribute 'dtype'

The reason for that is x=None in the following code snippet.

stylegan3/training/networks_stylegan2.py

Lines 778 to 783 in c233a91

	def forward(self, img, c, update_emas=False, **block_kwargs):
	_ = update_emas # unused
	x = None
	for res in self.block_resolutions:
	block = getattr(self, f'b{res}')
	x, img = block(x, img, **block_kwargs)

The actual Error is then raised here, when x.dtype is accessed:

stylegan3/training/networks_stylegan2.py

Lines 176 to 177 in c233a91

	x = conv2d_resample.conv2d_resample(x=x, w=w.to(x.dtype), f=self.resample_filter, up=self.up, down=self.down, padding=self.padding, flip_weight=flip_weight)

Also the first argument just passes x to the function which directly asserts x is not None:

stylegan3/torch_utils/ops/conv2d_resample.py

Line 71 in c233a91

assert isinstance(x, torch.Tensor) and (x.ndim == 4)

My Question is: Why is x=None and is that supposed to work?
I feel like x should be a Tensor of an actual image and not None.
I see others here training successfully and think, that the error seems on my side.

[Neilstid]

I think this kind of error may appear if you have either an empty directory or images in a wrong format (and thus can't be red by the loader). Is your images in either .jpg or .png ? Also try to print the path when it loads an image this could help.