Add mono models to huggingface model-zoo and incorporate into pipeline

[MXueguang]

No description provided.

[rodrigonogueira4]

Instead of try/except, should we pass a flag to distinguish PT from TF checkpoints?

[ronakice]

This entire segment of code should give an error since loader is not defined anymore?

[MXueguang]

@rodrigonogueira4 Yes, I feel a flag makes more sense. will add an flag --from_tf for evaluate_passage_ranker.py

[MXueguang]

@ronakice oh, I didn't notice there is a one change difference between my gcloud and local. Will remove that line.

[ronakice]

Huggingface's work-around is something like bool("ckpt" in model_name_or_path)

[MXueguang]

model_name_or_path is a directory may contains both "ckpt" and pytorch_model.bin ? https://huggingface.co/google/bert_uncased_L-10_H-128_A-2#list-files , so maybe still need to be stated by user?

[rodrigonogueira4]

To avoid problems with PT files containing ckpt in their name, I would just go with the flag.

[ronakice]

I feel like the model in the model-zoo is missing the vocab file?

[MXueguang]

I feel like the model in the model-zoo is missing the vocab file?

@ronakice Is this because we are using t5-base as tokenizer?

[ronakice]

https://huggingface.co/valhalla/t5-base-squad#list-files compare this with our list-files and it seems like our model is missing a few! Not sure if all are useful though!

[MXueguang]

Currently in construct_t5()

• we load model by T5ForConditionalGeneration.from_pretrained("castorini/monot5-base-msmarco").
• we load tokenizer by AutoTokenizer.from_pretrained("t5-base")
  which from different models on HuggingFace.

I think the missed files are from tokenizer.(e.g. config file and spiece.model or vocal file). The pretrained model itself doesn't require these files. (The original gs://neuralresearcher_data/doc2query/experiments/367 doesn't contains vocab files as well.)

2020-05-24 02:04:42 [INFO] configuration_utils: loading configuration file https://s3.amazona
ws.com/models.huggingface.co/bert/t5-base-config.json from cache 

2020-05-24 02:04:42 [INFO] tokenization_utils: loading file https://s3.amazonaws.com/models.huggingface.co/bert/t5-spiece.model from cache

Do we need to find a way to integrate the pretrained tokenizer files to our model files?

[MXueguang]

I don't think we need default='false' if action='store_true'. Could you double-check?

Yes, it can be removed.

[MXueguang]

revalidated experiment result on latest update

monot5

monoBERT

[rodrigonogueira4]

LGTM, thanks for implementing this!

[ronakice]

Great! We can merge this. Thanks for doing this @MXueguang