Generating .t7 file for inferencing

[saswatidana]

Hello,
I am trying to generate .t7 file for a trained model. For that I am running scripts/generate_candidates.py . This python file needs another input file saved_candidates_ids. How do I create this candidate_ids file?
Any pointer will help me to run inference code.

[ledw]

@saswatidana Thanks for the reporting. We'll update an instruction on generating candidates shortly.

[JinfengXiao]

@ledw I'm also curious about how to generate the .t7 file for a trained model. What is the format of the saved_candidates_ids file required in scripts/generate_candidates.py? Could you give me more instructions on this?

[abhinavkulkarni]

@ledw: Any updates on this?

I was also looking at scripts/generate_candidates.py script and it looks like it expects another pre-generated input file saved_candidates_ids. Digging more into the code reveals that this is a torch tensor of token_idxs of candidates.

Can you please let us know how to generate this file?

This is needed so that we can introduce new candidates from newer versions of Wikipedia.

Thanks!

[ledw-2]

Sorry for not updating on this.
The token_idx are generated from BERT tokenizers. The format is batch x vec where vec is the BERT token id vector of input.
It's the input "ids" from this function:
https://github.com/facebookresearch/BLINK/blob/main/blink/biencoder/data_process.py#L96

[abhinavkulkarni]

Thanks @ledw-2: I was able to follow your reply and generate embeddings for candidate entities: #106 (comment)