search

allanj / pytorch_neural_crf

Pytorch implementation of LSTM/BERT-CRF for named entity recognition
359 stars 62 forks source link

get error when running in torch 1.81 #24

Closed Hansyvea closed 3 years ago

Hansyvea commented 3 years ago

return torch.gather(word_rep[:, 1:, :], 1, orig_to_token_index.unsqueeze(-1).expand(batch_size, max_sent_len, rep_size)) RuntimeError: gather_out_cuda(): Expected dtype int64 for index

allanj commented 3 years ago

Let me check and reproduce the error

Hansyvea commented 3 years ago

Let me check and reproduce the error

have ya solved it yet?

allanj commented 3 years ago

Sorry. I didn't run into any issues using PyTorch 1.8.1 and Cuda 10.2.

What CUDA version are you using? also your detailed environments?

Or you can make a .long() to orig_to_token_index.unsqueeze(-1).expand(batch_size, max_sent_len, rep_size).long()

see if that solves your issue

Hansyvea commented 3 years ago

Sorry. I didn't run into any issues using PyTorch 1.8.1 and Cuda 10.2.

What CUDA version are you using? also your detailed environments?

Or you can make a .long() to orig_to_token_index.unsqueeze(-1).expand(batch_size, max_sent_len, rep_size).long()

see if that solves your issue

thanks for your reply! I am using Cuda 11.2 and the other packages are all up-to-date, so I assume the problem is caused by operating system since I have been running it on Windows