RuntimeError: only batches of spatial targets supported (3D tensors) but got targets of size: : [1, 215, 245, 3]

[40618233]

Hi,
I have a problem with this place, can you help me?

[xhh12381]

I have the same question!

[MotsPassant]

I fixed this issue by making sure my masks files only had 1 channel and not 3.

With open cv it was :
img = cv2.imread('path/to/your/mask')
img_1_channel = img[:,:,0]
cv2.imwrite('path/to/new/mask', img_1_channel)

[xhh12381]

Thanks, I know exactly what the problem is too.
Pytorch nn.CrossEntropyLoss() requires input dimension is (N,C,H,W), target dimension is (N,H,W).
So the mask must be pre-processed as a single channel.
I learned this from the following blog, hope it can give you some help.
https://blog.csdn.net/Fcc_bd_stars/article/details/105158215

[milesial]

Hi, can you try the latest master? I improved the support for custom datasets.