devsinghsachan / emdr2 Goto Github PK

the precomputed evidence embedding file is only 19GB if I download it by Google,and then I have a error message

Unpickling BlockData: /disk2/qby/Desktop/emdr2-main/embedding-path/emdr2-finetuning-embedding/psgs_w100-retriever-nq-emdr2-finetuning-base-topk50-epochs10-bsize64-async-indexer.pkl
Traceback (most recent call last):
File "tasks/run.py", line 67, in
main()
File "/disk2/qby/Desktop/emdr2-main/tasks/openqa/e2eqa/run.py", line 72, in main
open_retrieval_generative_qa(dataset_cls)
File "/disk2/qby/Desktop/emdr2-main/tasks/openqa/e2eqa/run.py", line 60, in open_retrieval_generative_qa
end_of_training_callback_provider=distributed_metrics_func_provider)
File "/disk2/qby/Desktop/emdr2-main/tasks/openqa/e2eqa/train_e2eqa.py", line 583, in train
model, optimizer, lr_scheduler = setup_model_and_optimizer(model_provider)
File "/disk2/qby/Desktop/emdr2-main/megatron/training.py", line 134, in setup_model_and_optimizer
model = get_model(model_provider_func)
File "/disk2/qby/Desktop/emdr2-main/megatron/training.py", line 43, in get_model
model = model_provider_func()
File "/disk2/qby/Desktop/emdr2-main/tasks/openqa/e2eqa/run.py", line 36, in model_provider
evidence_retriever = PreComputedEvidenceDocsRetriever()
File "/disk2/qby/Desktop/emdr2-main/megatron/model/emdr2_model.py", line 387, in init
self.precomputed_index_wrapper()
File "/disk2/qby/Desktop/emdr2-main/megatron/model/emdr2_model.py", line 417, in precomputed_index_wrapper
self.get_evidence_embedding(args.embedding_path)
File "/disk2/qby/Desktop/emdr2-main/megatron/model/emdr2_model.py", line 412, in get_evidence_embedding
load_from_path=True)
File "/disk2/qby/Desktop/emdr2-main/megatron/data/emdr2_index.py", line 28, in init
self.load_from_file()
File "/disk2/qby/Desktop/emdr2-main/megatron/data/emdr2_index.py", line 50, in load_from_file
state_dict = pickle.load(open(self.embedding_path, 'rb'))
_pickle.UnpicklingError: pickle data was truncated

Hi @DevSinghSachan ,

Thanks for sharing the code and resources. I was trying to reproduce the reported results with the released checkpoints, and I'm able to reproduce most of them except for TriviaQA. The reported reader scores are 71.13/71.43 (Dev/Test), but my reproduced scores are 68.6/68.8, which looks very close to one of the FiD variants (MSS + DPR retriever, T5 reader). Can you check if the released ckpts for TriviaQA are correct?

Besides, I'd also like to know:

1. I didn't see the code for MSS pretraining? Will you release it as well?
2. A bert_110m checkpoint is used to initialize the retriever. I wonder where I can download it?
3. Can you elaborate how to reproduce the ablations in Table 2 (Our Implementation part)? It's not clear to me the difference between (1) FiD (MSS retriever, MSS reader); (2) FiD (MSS retriever, MSS reader) and (3) EMDR2(MSS retriever, MSS reader).

Thank you in advance!
Rui

Hi Devendra, thanks for open sourcing this great project! I want to apply your code on my own Chinese dataset, but I am confused on how to process my dataset to get files like Pre-tokenized evidence passages and their titles and Wikipedia evidence passages from DPR paper you provided for open-domain QA tasks. Could you give me some advice to build them? Thanks in advance.