Significance of line 174 in train_query.py code

[Nishant3815]

Hi,

I was going through the code for query finetuning and I am not able to understand one condition in the code:

Is the above highlighted line redundant and if not what is the significance (I feel we can directly update the encoder). Just wanted to make sure that I am not missing anything.

[jhyuklee]

Hi @Nishant3815,

this line (v1.0.0) was written to sync the pre-trained encoder (for the retrieval) with the target encoder that is used for the query-side fine-tuning. As you mentioned, we updated the code (v1.1.0) recently to directly update the target encoder and use it for the retrieval as well: https://github.com/princeton-nlp/DensePhrases/blob/v1.1.0/train_query.py

It gives a slightly better accuracy, overall.

[Nishant3815]

Thanks for the confirmation on updating the target encoder. This makes sense, I had been following a previous version.

For your comment, "this line (v1.0.0) was written to sync the pre-trained encoder (for the retrieval) with the target encoder that is used for the query-side fine-tuning", in a scenario where we would like to keep both pretrained_encoder and target_encoder version of the code, can we keep line 175 and 176 and remove the "if" statement of 174.

[jhyuklee]

Yes you can do that. It has no effect in v1.0.0 for now unless you use a higher divisor (currently 1) for the updating period.

Note that v1.0.0 with deepcopy syncs the pre-trained encoder w/ the target encoder after each epoch, so freezes the pre-trained encoder during each epoch. But using only the target encoder in v1.1.0 retrieves phrases from the same target encoder that is being updated for every "step."

[Nishant3815]

Thanks, this helps.