拥抱面孔 - PyTorch RuntimeError：nll_loss_forward_reduce_cuda_kernel_2d_index 未为 Int 实现

Question

I attempted to fine-tune the Wav2vec2 model for an audio classification task by following the tutorial at https://huggingface.co/docs/transformers/tasks/audio_classification and copying and pasting the code provided.

However, when I trained the model, I encountered the following error:

nll_loss_forward_reduce_cuda_kernel_2d_index not implemented for Int.

I believe the problem may lie with PyTorch . It seems to be working on Ubuntu WSL 22.04 , but not on Powershell .

My system specifications include:

- Powershell
- datasets 2.8.0
- transformers 4.25.1
- torch                   1.13.1+cu117
- torchaudio              0.13.1+cu117
- torchvision             0.14.1+cu117

How can I resolve this issue?

Answer 1

Convert your target dtype to torch.int64 .

I received the same error when my target dtype was torch.int32 . I resolved the issue by adding the following line to my code:

target = target.to(torch.int64)