Refactor DPR from FB to Transformers codebase

[kolk]

Breaking changes:

• DPR instantiation:
  signature:

DensePassageRetriever( document_store: BaseDocumentStore,
                query_embedding_model: str,
                passage_embedding_model: str,
                max_seq_len: int = 256,
                projection_dim: int = 0,
                use_gpu: bool = True,
                batch_size: int = 16,
                do_lower_case: bool = False,
                use_amp: str = None)

example call:

dpr_object = DensePassageRetriever(document_store=document_store, query_embedding_model="facebook/dpr-question_encoder-single-nq-base", passage_embedding_model="facebook/dpr-ctx_encoder-single-nq-base", use_gpu=True)

• tokenizers: Query and passage tokenizers instantiated from DPRQuestionEncoderTokenizer and DPRContextEncoderTokenizer

example call:

query_config = DPRConfig(projection_dim=0, config="facebook/dpr-question_encoder-single-nq-base", dropout=0.0)
query_tokenizer = DPRQuestionEncoderTokenizer.from_pretrained("facebook/dpr-question_encoder-single-nq-base"
, config=query_config)

• tensorizer:
  signature
  _tensorizer(tokenizer, title: Optional[List[str]], text: List[str], add_special_tokens: bool = True)

example call:

 token_ids_batch, token_seg_batch, token_attn_mask = dpr_instance._tensorizer(tokenizer, text=ctx_text, title=ctx_title)

• Encoders: Query and passage encoders instantiated from DPRQuestionEncoder and DPRContextEncoder
  example call:

query_config = DPRConfig(projection_dim=0, config="facebook/dpr-question_encoder-single-nq-base", dropout=0.0)
query_encoder = DPRQuestionEncoder.from_pretrained("facebook/dpr-question_encoder-single-nq-base", config=query_config)

• generate_batch_predictions:
  Modified to include titles_mask parameter for handling batches with some missing titles
  example call:

dpr_object._generate_batch_predictions(texts,  encoder_model, tokenizer, titles, titles_mask, batch_size) 

Title inclusion in passages:
embed_title == True:
- No titles in batch: no "name" key in any Document in batch or no meta attribute => titles: List["", "", ..] and titles_mask: List[0, 0, ..]
- Some titles in batch: some Document has "name" key and value => titles: List["", title1, ..] and titles_mask: List[0, 1, ..]
- all samples has titles: all Document has "name" key and value => titles: List[title1, title2, ..] and titles_mask: List[1, 1, ..]

embed_title == False: => titles: List["", "", ..] and titles_mask: List[0, 0, ..]

TO-DO:

• Change DPR query and passage encoders from official FB HFBertEncoder to transformers DPREncoders.
• Modify _generate_batch_predictions to reflect DPREncoderOutput
• Modify _prepare_model function to load the saved checkpoint into DPREncoder models
• Remove BertTensorizer
• Refactor DPR model download and saved state parameters
  Test performance on top_k=20 retrieved documents on nq_dev.json

1. Official FB encoder modules

Retriever Recall: 0.9345536176826109
Retriever Mean Avg Precision: 0.7782416555082636

2. transformers DPRQuestionEncoder and DPRContextEncoder and modelhub models

Retriever Recall: 0.9345536176826109
Retriever Mean Avg Precision: 0.7782502895956392

• Modify Tutorial-5 eval scripts
• Include transformers DPR modules in dpr_utils until it is supported by transformers==3.0.2
• Update Tutorial-5 and Tutoriacl-6 ipython notebooks
• Handle title use-cases for embed_passages

[Timoeller]

Added a suggestion to make the code simpler.

[Timoeller]

There is still some very custom code in there that should be changed

[Timoeller]

I just made some minor suggestions (might be a bit nit picky, but lets strive for best code possible)

Lets move the HF DPR classes to dpr_utils, remove the old fb code and then this looks very much like we can merge! : )

[tholor]

Looking good. Small changes required. Mostly naming and docstrings.

[Timoeller]

Hey, a minor thing I noticed that happened to others here before. Its not a big thing but lets try to avoid this in the future:
I have seen you also commited the output of tutorial notebooks 5+6. Please remove the cell output before commiting since it pollutes the changes a lot.

[Timoeller]

Added some thoughts on loading the models with potential for improvement. Happy to start a discussion here, I might not be aware of all necessary details.

[tholor]

Some smaller changes required. I am also not sure yet about the handling of titles when some docs have it and some don't.

[tholor]

Looking good now! I only updated one docstring slightly.