CrossEntropyLoss() Error

vision
1

Hello,
I’ve been working on a semantic segmentation task in point clouds, with RandLA-Net.
I’m trying to train the network on the DALES dataset, which is 8 labels semantic segmentation dataset, while doing so, I’ve got the following error:

image
image1186×177 8.5 KB

Now, when I searched online I found that it might be triggered due to the difference between the logarithm distributed scores shape, and the labels shape. when printing the shapes of them I’m getting the following:

code:

 scores = model(points)

 logp = torch.distributions.utils.probs_to_logits(scores, is_binary=False)
          
 print(logp.shape)

 print(labels.shape)

and the output is:

torch.Size([1, 8, 40960])
torch.Size([1, 40960])

I’ve been struggling to understand if the shapes are incorrect, and how to fix them, will appreciate any help.
Thanks in advance.

2

Anyone? Still struggling to solve this issue. Thanks

3

Could you rerun the script via:

CUDA_LAUNCH_BLOCKING=1 python setup.py args

and post the stack trace here?
Alternatively, run the code on the CPU and post the complete stack trace.

1 Like