Can custom loss function be used for model.fit?

[shangh1]

Hi, I'm using this git to run a cross encoder model, hope to use a customized function. The code is like below. The thing is that using model.fit, it does not specify which is the output.
In future, I'll also add another argument in this customized loss to compute lambdarank loss. Thanks!

def my_loss(output, label):
    loss = ... 
    return loss

model.fit(train_dataloader=train_dataloader,
              loss_fct = myloss)

[tomaarsen]

Hello!

Yes, this is totally possible. If you provide your loss function this way, then it'll be called right here:

sentence-transformers/sentence_transformers/cross_encoder/CrossEncoder.py

Line 240 in 7373533

loss_value = loss_fct(logits, labels)

i.e. with logits & labels.

• Tom Aarsen

[shangh1]

Thanks for your quick reply! Got it! So the default arguments are logits and labels.