BERTConfigs in example usages in `modeling.py` are not OK (?)

[davidefiocco]

Hi!

In the config definition https://github.com/huggingface/pytorch-pretrained-BERT/blob/21f0196412115876da1c38652d22d1f7a14b36ff/pytorch_pretrained_bert/modeling.py#L848
in the Example usage of BertForSequenceClassification in modeling.py, there's things I don't understand:

• vocab_size in not an acceptable parameter name, by looking at the BertConfig class definition https://github.com/huggingface/pytorch-pretrained-BERT/blob/21f0196412115876da1c38652d22d1f7a14b36ff/pytorch_pretrained_bert/modeling.py#L70

• even by changing vocab_size into vocab_size_or_config_json_file, for the choice of the other params given in the example i.e.
  vocab_size=32000, hidden_size=512, num_hidden_layers=8, num_attention_heads=6, intermediate_size=1024
  I get:
  ValueError: The hidden size (512) is not a multiple of the number of attention heads (6)
  I think that something similar may be true for the other classes as well, BertForQuestionAnswering, BertForNextSentencePrediction, etc.

Am I missing something?

[thomwolf]

Hi @davidefiocco, you are right, I updated the docstrings in the new release 0.3.0.