Transformers older version failling, hack included

[Maddy12]

Describe the bug When attempting to run precompute_text.py, there is an issue.

To Reproduce

 python precompute_text.py youcook2  --cuda --metadata_name "all_${PERTURBATION}"  --data_path ${DATA_PATH}

Screenshots

******************** Loading model bert-base-uncased from transformers
Running model on device cuda:0
Maximum input length 512
Loading meta file of 1 MB
Took 0.1 seconds for 1790.
Compute total_words and max_words: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1790/1790 [00:00<00:00, 247806.85it/s]
Total 115316 average 64.42 max 242
******************** Loading and testing dataset.
Keyword arguments {'add_special_tokens': False} not recognized.
Keyword arguments {'add_special_tokens': False} not recognized.
Keyword arguments {'add_special_tokens': False} not recognized.
Keyword arguments {'add_special_tokens': False} not recognized.
Keyword arguments {'add_special_tokens': False} not recognized.
Keyword arguments {'add_special_tokens': False} not recognized.
Keyword arguments {'add_special_tokens': False} not recognized.
Keyword arguments {'add_special_tokens': False} not recognized.
Keyword arguments {'add_special_tokens': False} not recognized.
Keyword arguments {'add_special_tokens': False} not recognized.
key: --bv0V6ZjWI

text: ['[CLS] crush and chop the garlic [SEP]', 'add oil garlic and salt to a bowl [SEP]', 'mix the tomoatoes with the oil mixture [SEP]', 'chop the basil [SEP]', 'spread the oil mixture onto the dough [SEP]', 'place provolone cheese and mozzarella cheese onto the dough [SEP]', 'add the basil to the pizza [SEP]', 'place the tomatoes on the pizza [SEP]', 'sprinkle cheese onto the pizza [SEP]', 'bake the pizza in an oven [SEP]']

text_tokenized: [['[CLS]', 'crush', 'and', 'chop', 'the', 'garlic', '[SEP]'], ['add', 'oil', 'garlic', 'and', 'salt', 'to', 'a', 'bowl', '[SEP]'], ['mix', 'the', 'tom', '##oat', '##oes', 'with', 'the', 'oil', 'mixture', '[SEP]'], ['chop', 'the', 'basil', '[SEP]'], ['spread', 'the', 'oil', 'mixture', 'onto', 'the', 'dough', '[SEP]'], ['place', 'pro', '##vo', '##lone', 'cheese', 'and', 'mo', '##zza', '##rella', 'cheese', 'onto', 'the', 'dough', '[SEP]'], ['add', 'the', 'basil', 'to', 'the', 'pizza', '[SEP]'], ['place', 'the', 'tomatoes', 'on', 'the', 'pizza', '[SEP]'], ['sp', '##rin', '##kle', 'cheese', 'onto', 'the', 'pizza', '[SEP]'], ['ba', '##ke', 'the', 'pizza', 'in', 'an', 'oven', '[SEP]']]

tokens: tensor([  101, 10188,  1998, 24494,  1996, 20548,   102,  5587,  3514, 20548,
         1998,  5474,  2000,  1037,  4605,   102,  4666,  1996,  3419, 16503,
        22504,  2007,  1996,  3514,  8150,   102, 24494,  1996, 14732,   102,
         3659,  1996,  3514,  8150,  3031,  1996, 23126,   102,  2173,  4013,
         6767, 27165,  8808,  1998,  9587, 20715, 21835,  8808,  3031,  1996,
        23126,   102,  5587,  1996, 14732,  2000,  1996, 10733,   102,  2173,
         1996, 12851,  2006,  1996, 10733,   102, 11867,  6657, 19099,  8808,
         3031,  1996, 10733,   102,  8670,  3489,  1996, 10733,  1999,  2019,
        17428,   102])

sentence_lengths: [7, 9, 10, 4, 8, 14, 7, 7, 8, 8]

******************** Running the encoding.
Encoding text with model: bert-base-uncased, layers: [-2, -1], batch size: 1, workers: 0
compute text features:   0%|                                                                                                                                                                                              | 0/1790 [00:00<?, ?it/s]Keyword arguments {'add_special_tokens': False} not recognized.
Keyword arguments {'add_special_tokens': False} not recognized.
Keyword arguments {'add_special_tokens': False} not recognized.
Keyword arguments {'add_special_tokens': False} not recognized.
Keyword arguments {'add_special_tokens': False} not recognized.
Keyword arguments {'add_special_tokens': False} not recognized.
Keyword arguments {'add_special_tokens': False} not recognized.
Keyword arguments {'add_special_tokens': False} not recognized.
Keyword arguments {'add_special_tokens': False} not recognized.
Keyword arguments {'add_special_tokens': False} not recognized.
Traceback (most recent call last):
  File "models/coot-videotext/precompute_text.py", line 460, in <module>
    main()
  File "/home/.conda/envs/multimodal_py38/lib/python3.8/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context
    return func(*args, **kwargs)
  File "models/coot-videotext/precompute_text.py", line 210, in main
    hidden_states = model_outputs["hidden_states"]
TypeError: tuple indices must be integers or slices, not str
compute text features:   0%|               

The current output of model_outputs is a tuple:

(pdb) model_outputs[1].shape
torch.Size([1, 768])
(Pdb) model_outputs[0].shape
torch.Size([1, 82, 768])

System Info: transformers==3.4.0

Additional context I am sure it is an issue with the versioning of transformers, so package information on what version is required would be helpful as it is not indicated in the requirements.txt

[Maddy12]

My hack based on the current version of transformers is the following:

tokenizer = AutoTokenizer.from_pretrained(model_name, cache_dir=args.model_path,
                                                  add_special_tokens=args.add_special_tokens)
config = BertConfig.from_pretrained(model_name, output_hidden_states=True)
model = BertModel.from_pretrained(model_name, config=config)

and to no longer pass add_special_tokens in function TextConverterDataset when calling:

 sentence_tokens_str = self.tokenizer.tokenize(sentence) # , add_special_tokens=self.add_special_tokens)

[Maddy12]

I see the tranformers version now in requrements_frozen.txt. The hack may be useful if want to update to more recent versions. Thanks!