Grabbing the Batch Size on a Loss Function?

fastai
1

So quick question, I am using a loss function that inherits from the batch_size used in the data. How do I pass it in? I know I can implement it as a callback but I’m not quite there yet to convert it over.

Eg:

def total_loss(out, label):
  bs = learn.data.batch_size

Can I pass it into the loss function? But how do I know then that the model will actually pass it through if I do:

def total_loss(out, label, batch_size)

Thanks for any clarification you guys could add for me!

2

In general, the (mini)batch size is found as the first dimension of out.

To understand the scope of variables used in functions, you will need to study “closures” in Python.

HTH, Malcolm

3

By mini-batch do you mean if I specify for instance bs=12 when I make my databunch, that size?

4

mini-batch size == bs in fastai.

It’s easy to see whether bs ends up in out’s dimension 0. Just try it: run a batch through your model and check the dimensions of out.

If not, understanding “partial function” will help you.

1 Like
5

Got it. I’ll look into partial functions! But it makes sense! Thank you very much!