openmatch / openmatch Goto Github PK

In paper ANCE , the ANCE model uses cosine similarity; where in this repo, the T5-ANCE model seems to use dot similarity.

Have you tried using cosine similarity in T5-ANCE? Does it perform worse than dot similarity? Are there any criteria for determining which similarity function should be used in different models?

from Sentence-transformers format: https://huggingface.co/sentence-transformers/gtr-t5-base

I ran OpenMatch dense retrieval according to docs/dr-msmarco-passage.md, but when installing the dependencies according to README, the evaluation on the dev set during training has problems when the huggingface is up to date(4.27.1)

If 'self.head' is None, this will result in a saving error, such as using monot5.

G. Izacard, M. Caron, L. Hosseini, S. Riedel, P. Bojanowski, A. Joulin, E. Grave Unsupervised Dense Information Retrieval with Contrastive Learning

Torch >= 1.0 in README

Support and additional type of parameter efficient method from OpenDelta:

One key consideration for getting dense representations from such a model is to only mean pool the hidden state representations of real tokens.
For example:

elif self.pooling == "mean":
      soft_prompt_attention_mask = items.attention_mask
      soft_prompt_attention_mask[
          :, : model.soft_prompt_token_number
      ] = torch.zeros(
          (items.attention_mask.shape[0], model.soft_prompt_token_number)
      )
      reps = mean_pooling(
          hidden, soft_prompt_attention_mask
      )  # only pool hidden reps of real tokens

merge beir eval pipeline
fix split search for retrieve and successive search

transformers>=4.21.3
datasets>=2.10.1

There is an error in current has_answers function when meeting special characters, such as "café".

We should refer to 'pyserini':

from pyserini.eval.evaluate_dpr_retrieval import has_answers

1. manually set gpu
2. remove unuseful del stataments

ANCE uses a layer norm over the embeddings: https://github.com/microsoft/ANCE/blob/master/model/models.py#L146

Hello, can OpenMatch support Chinese corpus, and can you please show how to achieve it?

Zhuang et al. Asyncval: A Toolkit for Asynchronously Validating Dense Retriever Checkpoints during Training. SIGIR'22.