Huggingface AutoTokenizer can't load from local path

Question

I'm trying to run language model finetuning script (run_language_modeling.py) from huggingface examples with my own tokenizer(just added in several tokens, see the comments). I have problem loading the tokenizer. I think the problem is with AutoTokenizer.from_pretrained('local/path/to/directory').

Code:

from transformers import *

tokenizer = AutoTokenizer.from_pretrained('bert-base-uncased')
# special_tokens = ['<HASHTAG>', '<URL>', '<AT_USER>', '<EMOTICON-HAPPY>', '<EMOTICON-SAD>']
# tokenizer.add_tokens(special_tokens)
tokenizer.save_pretrained('../twitter/twittertokenizer/')
tmp = AutoTokenizer.from_pretrained('../twitter/twittertokenizer/')

Error Message:

OSError                                   Traceback (most recent call last)
/z/huggingface_venv/lib/python3.7/site-packages/transformers/configuration_utils.py in get_config_dict(cls, pretrained_model_name_or_path, pretrained_config_archive_map, **kwargs)
    248                 resume_download=resume_download,
--> 249                 local_files_only=local_files_only,
    250             )

/z/huggingface_venv/lib/python3.7/site-packages/transformers/file_utils.py in cached_path(url_or_filename, cache_dir, force_download, proxies, resume_download, user_agent, extract_compressed_file, force_extract, local_files_only)
    265         # File, but it doesn't exist.
--> 266         raise EnvironmentError("file {} not found".format(url_or_filename))
    267     else:

OSError: file ../twitter/twittertokenizer/config.json not found

During handling of the above exception, another exception occurred:

OSError                                   Traceback (most recent call last)
<ipython-input-32-662067cb1297> in <module>
----> 1 tmp = AutoTokenizer.from_pretrained('../twitter/twittertokenizer/')

/z/huggingface_venv/lib/python3.7/site-packages/transformers/tokenization_auto.py in from_pretrained(cls, pretrained_model_name_or_path, *inputs, **kwargs)
    190         config = kwargs.pop("config", None)
    191         if not isinstance(config, PretrainedConfig):
--> 192             config = AutoConfig.from_pretrained(pretrained_model_name_or_path, **kwargs)
    193 
    194         if "bert-base-japanese" in pretrained_model_name_or_path:

/z/huggingface_venv/lib/python3.7/site-packages/transformers/configuration_auto.py in from_pretrained(cls, pretrained_model_name_or_path, **kwargs)
    192         """
    193         config_dict, _ = PretrainedConfig.get_config_dict(
--> 194             pretrained_model_name_or_path, pretrained_config_archive_map=ALL_PRETRAINED_CONFIG_ARCHIVE_MAP, **kwargs
    195         )
    196 

/z/huggingface_venv/lib/python3.7/site-packages/transformers/configuration_utils.py in get_config_dict(cls, pretrained_model_name_or_path, pretrained_config_archive_map, **kwargs)
    270                     )
    271                 )
--> 272             raise EnvironmentError(msg)
    273 
    274         except json.JSONDecodeError:

OSError: Can't load '../twitter/twittertokenizer/'. Make sure that:

- '../twitter/twittertokenizer/' is a correct model identifier listed on 'https://huggingface.co/models'

- or '../twitter/twittertokenizer/' is the correct path to a directory containing a 'config.json' file

If I change AutoTokenizer to BertTokenizer, the code above can work. Also I can run the script without any problem is I load by shortcut name instead of path. But in the script run_language_modeling.py it uses AutoTokenizer. I'm looking for a way to get it running.

Any idea? Thanks!

Answer 1

The problem is that you are using nothing that would indicate the correct tokenizer to instantiate.

For reference, see the rules defined in the Huggingface docs. Specifically, since you are using BERT:

contains bert: BertTokenizer (Bert model)

Otherwise, you have to specify the exact type yourself, as you mentioned.

Answer 2

AutoTokenizer.from_pretrained fails if the specified path does not contain the model configuration files, which are required solely for the tokenizer class instantiation.

In the context of run_language_modeling.py the usage of AutoTokenizer is buggy (or at least leaky).
There is no point to specify the (optional) tokenizer_name parameter if it's identical to the model name or path. Therefore, to my understanding, it supposes to support exactly the case of a modified tokenizer. I also found this issue very confusing.

The best workaround that I have found is to add config.json to the tokenizer directory with only the "missing" configuration:

{
  "model_type": "bert"
}

Answer 3

when loading modified tokenizer or pretrained tokenizer you should load it as follows:

tokenizer = AutoTokenizer.from_pretrained(path_to_json_file_of_tokenizer, config=AutoConfig.from_pretrained('path to thefolderthat contains the config file of the model'))