single gpu train

[wmrenr]

when I use single GPU to train VOC datasets ,I got an error:

can you help me?

[wmrenr]

I print torch.from_numpy(labels_cmap).shape :[512,512,3] and the 3D transpose should be performed, so I cant understand '.permute([0, 3, 1, 2])' in your codes.

[HenryQUQ]

I print torch.from_numpy(labels_cmap).shape :[512,512,3] and the 3D transpose should be performed, so I cant understand '.permute([0, 3, 1, 2])' in your codes.

torch.from_numpy(labels_cmap).shape should be [1,512,512,3] because it is followed the RGB rule, and after permuting, it will turn the default tensor form [1,3,512,512].

The result why your issue happens is you read only one image each time, so the shape will be (512, 512, 3). You can simply add '.unsqueeze(0)' after ' torch.from_numpy(labels_cmap)'. This line should be 'torch.from_numpy(labels_cmap).unsqueeze(0).permute([0, 3, 1, 2])

[llttisme]

Excuse me, can it run successfully on a single GPU?

[yxsec]

Excuse me, can it run successfully on a single GPU?

Hi, have you solved it?