RuntimeError: stack expects a non-empty TensorList

[hxqstree]

Hi，thank you for your nice work!

I get the following error:

File "/CMD-SE-release-main/models/model.py", line 621, in text_to_embedding x = torch.stack(all_token_embeddings, dim=0) # [batch_size, n_ctx, d_model] RuntimeError: stack expects a non-empty TensorList

How should this be solved?

[ltttpku]

Hi, what command do you use to run the Python script?

[hxqstree]

I set the hyperparameters as follows: --batch_size 2 \ --output_dir [path to save checkpoint] \ --epochs 80 \ --lr 1e-4 --min-lr 1e-7 \ --hoi_token_length 25 \ --enable_dec \ --dataset_file hico --multi_scale true --f_idxs 5 8 11 --set_cost_hoi_type 5 --use_aux_text true \ --enable_focal_loss --description_file_path hico_hoi_descriptions.json After setting the parameters, I ran "main" directly. And sometimes there is no problem in the first round of training, but this problem will occur in the second round Is it possible that it is due to batch_size?

[ltttpku]

Yes, it is possible that the issue is related to the batch size. Would it be possible to try a larger batch size and see if that helps?

[hxqstree]

Thanks for your help！