search

INK-USC / CrossFit

Code for paper "CrossFit :weight_lifting:: A Few-shot Learning Challenge for Cross-task Generalization in NLP" (https://arxiv.org/abs/2104.08835)
105 stars 6 forks source link

Describe a bug #3

Closed telxt closed 1 year ago

telxt commented 2 years ago

describe a bug

Whe I try to run example_scripts/finetune_boolq.sh , an error is raised

Traceback (most recent call last):
  File "/mnt/sfs_turbo/lxt/CrossFit/cli_singletask.py", line 147, in <module>
    main()
  File "/mnt/sfs_turbo/lxt/CrossFit/cli_singletask.py", line 144, in main
    run(args, logger)
  File "/mnt/sfs_turbo/lxt/CrossFit/run_singletask.py", line 84, in run
    best_dev_performance, best_model_state_dict = train(args, logger, model, train_data, dev_data, optimizer, scheduler)
  File "/mnt/sfs_turbo/lxt/CrossFit/run_singletask.py", line 175, in train
    curr_performance = inference(model if args.n_gpu==1 else model.module, dev_data)
  File "/mnt/sfs_turbo/lxt/CrossFit/run_singletask.py", line 219, in inference
    outputs = model.generate(input_ids=batch[0],
  File "/mnt/sfs_turbo/zhangshudan/anaconda3/envs/opendelta_dev/lib/python3.8/site-packages/torch/autograd/grad_mode.py", line 28, in decorate_context
    return func(*args, **kwargs)
  File "/mnt/sfs_turbo/zhangshudan/anaconda3/envs/opendelta_dev/lib/python3.8/site-packages/transformers/generation_utils.py", line 1053, in generate
    return self.beam_search(
  File "/mnt/sfs_turbo/zhangshudan/anaconda3/envs/opendelta_dev/lib/python3.8/site-packages/transformers/generation_utils.py", line 1787, in beam_search
    outputs = self(
  File "/mnt/sfs_turbo/zhangshudan/anaconda3/envs/opendelta_dev/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1051, in _call_impl
    return forward_call(*input, **kwargs)
TypeError: forward() got an unexpected keyword argument 'past_key_values'

System information

cherry979988 commented 2 years ago

Thank you for raising this issue. I am looking into this problem!

cherry979988 commented 2 years ago

This is probably due to the transformers library upgrade. Looks like new input variables are defined in the new version.

Solution 1: Use the transformer version stated in the README file. (But it's a pretty outdated version 😢) Solution 2: Replace the bart.py file with this new one (https://github.com/INK-USC/CrossFit/blob/676f801d7cc2c431ddd0e21b9593183d8e95f580/bart.py). **model_kwargs will automatically handle new input variables due to transformer versioning. I have tested this with Python 3.6.9, transformers 4.10.0, torch 1.7.1. I hope it solves your issue.