Checkpoints for CALM model

[metodj]

Hi!

Thanks for the great repo! I was just wondering if it would be possible to release checkpoints for CALM model as well? To enable experimentation with this baseline model without having to fine-tune it on my own.

Best,

Metod

[raymin0223]

Hi @metodj,

Thanks for the interest in our paper! You can download checkpoints here.
Note that "weighted_ce" checkpoints are for CALM model, and please let me know if there are any problems.

[metodj]

Hi @raymin0223 ,

thanks for your answer!

I downloaded the checkpoints from the link you provided, however, it seems like a tokeniser is missing. Concretely, downloading the CNN/DM weighted_ce pretrained model to a folder ./save/cnndm_t5_large_weighted_ce and then running run_sum_cnndm.sh

CUDA_VISIBLE_DEVICES=0 python -m torch.distributed.run --nproc_per_node=1 \
    run_summarization.py \
    --model_name_or_path ./save/cnndm_t5_large_weighted_ce/ \
    --do_eval \
    --dataset_name cnn_dailymail \
    --dataset_config_name "3.0.0" \
    --output_dir ./save/cnndm_t5_large_weighted_ce/ \
    --per_device_eval_batch_size 1 \
    --deploy_scenario True \
    --use_synchronize True \
    --overwrite_output_dir \
    --predict_with_generate \
    --source_prefix "summarize: " \
    --use_early_exit True \
    --exit_conf_type softmax \
    --exit_conf_threshold 0.9 \
    --exit_min_layer 4 \

results in the following error:

Traceback (most recent call last):                                                                                                                                               
  File "run_summarization.py", line 585, in <module>                                                                                                                             
    main(model_args, data_args, training_args, additional_args, model_cls, trainer_cls)                                                                                          
  File "run_summarization.py", line 222, in main                                                                                                                                 
    tokenizer = AutoTokenizer.from_pretrained(                                                                                                                                   
  File "/home/mjazbec/miniconda3/envs/calm/lib/python3.8/site-packages/transformers/models/auto/tokenization_auto.py", line 720, in from_pretrained                              
    return tokenizer_class_fast.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)                                                                                
  File "/home/mjazbec/miniconda3/envs/calm/lib/python3.8/site-packages/transformers/tokenization_utils_base.py", line 1795, in from_pretrained                                   
    raise EnvironmentError(                                                                                                                                                      
OSError: Can't load tokenizer for './save/cnndm_t5_large_weighted_ce/'. If you were trying to load it from 'https://huggingface.co/models', make sure you don't have a local dire
ctory with the same name. Otherwise, make sure './save/cnndm_t5_large_weighted_ce/' is the correct path to a directory containing all relevant files for a T5TokenizerFast tokenizer.

Could you upload the tokenisers as well?

Thanks!

[metodj]

Aha, I resolved the issue by downloading tokenizer.json from here. Might be worth adding this step to the README.md in case someone else will want to use your checkpoints in the future and will run into a similar issue

[raymin0223]

Hi @metodj,

Thanks so much for finding and fixing the errors! I've updated the README.md as you advised.
If you have any more questions, please feel free to reopen the issues. Thanks again!

[metodj]

In run_qa_squad.sh, run_summarization.py is called which I suppose should be run_question_answering.py instead, right? And same for run_tr_iwlst.sh?

[raymin0223]

Oh no, so silly mistakes 😂 You are right and thanks so so much for reporting bugs!!