[RAG] Generate Dense Embeddings Fix

[klshuster]

Patch description
#3868 pointed out a couple of current shortcomings in the custom embeddings generation process for RAG/FiD DPR Models.

1. If specifying a ParlAI RAG or FiD model, one needs to set --dpr-model True; this has been clarified in the README.
2. If specifying specifically a FiD-RAG model, we need to make sure that the document encoder is initialized appropriately (this was not correctly handled prior)

Testing steps
I tried running the provided command in #3868 with --dpr-model True, after the change, and it appears the script works correctly:

$ python parlai/agents/rag/scripts/generate_dense_embeddings.py -mf zoo:hallucination/bart_fid_rag/model --dpr-model true --passages-file data/models/hallucination/wow_passages/wow_articles.paragraphs.tsv --outfile /tmp/out --num-shards 1 --shard-id 0 -bs 32
17:12:05 | Overriding opt["model"] to dpr_agent (previously: fid)
17:12:05 | Overriding opt["interactive_candidates"] to inline (previously: fixed)
17:12:05 | Overriding opt["share_encoders"] to False (previously: True)
17:12:06 | Using CUDA
17:12:06 | DPR: full interactive mode on.
17:12:42 | Total parameters: 241,228,800 (240,442,368 trainable)
17:12:42 | Loading existing model parameters from /private/home/kshuster/ParlAI/data/models/hallucination/bart_fid_rag/model
17:12:49 | Loading data/models/hallucination/wow_passages/wow_articles.paragraphs.tsv
17:12:49 | Reading data from: data/models/hallucination/wow_passages/wow_articles.paragraphs.tsv
2863it [00:00, 105418.20it/s]
17:12:49 | Shard 0 of 1 encoding psg index 0 to 2862, out of 2862
17:12:51 | Encoded 320 out of 2862 passages
17:12:51 | Encoded 640 out of 2862 passages
17:12:52 | Encoded 960 out of 2862 passages
17:12:52 | Encoded 1280 out of 2862 passages
17:12:53 | Encoded 1600 out of 2862 passages
17:12:53 | Encoded 1920 out of 2862 passages
17:12:54 | Encoded 2240 out of 2862 passages
17:12:54 | Encoded 2560 out of 2862 passages
17:12:55 | Writing results to /tmp/out_0.pt
17:12:55 | Writing ids to /tmp/ids_0

[mojtaba-komeili]

Does this slow things down if you have a fid model and you load document_path later again? In that case does it make sense to move this to the else condition of the following if?

[klshuster]

im not entirely sure what you're asking here --> we're only getting the document path from the .opt file. We should never actually build a FiD model in this scenario

[mojtaba-komeili]

LGTM.