Questions about how to make your code work about fewie

[leonhardhennig]

Hi,
there is no train.py, because we didn't do any training. Per few-shot batch, there is 'only', depending on the readout you configure, some fitting done. e.g. in src/fewie/evaluation/classifiers/logistic_regression.py we fit a LR function.

[kimkyeonghun]

Thank you for your quick response. I understand the answer well enough.

Looking at the paper, it appears that they used contrastive learning. Do you have any hints on how to use it?
Or is it something that is already included in the pre-trained BERT?

[leonhardhennig]

@chen-yuxuan did we include the code for the contrastive learning in the repo?

[chen-yuxuan]

Thank you for your quick response. I understand the answer well enough.

Looking at the paper, it appears that they used contrastive learning. Do you have any hints on how to use it? Or is it something that is already included in the pre-trained BERT?

Hi, @kimkyeonghun and @leonhardhennig
Yes, it is included in the contrastive-pre-training branch. I think after checking-out this branch, you can run it in quite the same way as for other encoders, since the configuration files for contrastive-variants are available here: https://github.com/DFKI-NLP/fewie/tree/contrastive-pretraining/config/encoder

Also we don't train for vanilla encoders, and don't train by ourselves for other-task/domain fine-tuned encoders (by just using the existing checkpoints from HF). However, we did train for contrastive learning, and in this case, the training logic is contained in the evaluate.py (calling

fewie/src/fewie/evaluation/scenarios/few_shot_contrastive_training.py

Line 24 in 68d7f12

def eval_few_show_contrastive_training(

) in this branch.

[kimkyeonghun]

Thank you for your quick and accurate response!