CODEWITHSUNDEEP
kerassyntax
Arkleseisure
Arkleseisure

Reputation: 438

What is the syntax for passing arguments into keras loss functions?

I'm trying to give my keras neural network categorical crossentropy loss with from_logits=True. However, I'm not sure how to pass this into the code, as it asks me to specify the target and output.

Normally I can use:

network.compile(sgd, loss='categorical_crossentropy'),

but now I'm having to try this:

network.compile(sgd, loss=categorical_crossentropy(from_logits=True))

which gives me an error:

TypeError: categorical_crossentropy() missing 2 required positional arguments: 'target' and 'output'

The best I can come up with is:

network.compile(sgd, loss=categorical_crossentropy(y_true, network.output, from_logits=True))

I don't have any idea what to put for y_true however as this is not part of the network. I've had a look around online but haven't come across anything which specifies how to do this, including, weirdly, the keras documentation.

Upvotes: 1

Views: 1104

Answers (1)

Daniel Möller
Daniel Möller

Reputation: 86630

Keras losses need strictly two arguments: y_true (ground truth data) and y_pred (model's output).

If you want to use a function with a different signature, you must wrap it to follow the correct signature.

import keras.backend as K

def cc_from_logits(y_true, y_pred):
    return K.categorical_crossentropy(y_true, y_pred, from_logits=True, axis=-1)

model.compile(loss=cc_from_logits)

I'm quite convinced that cc_with_logits brings the exact same results as softmax + 'categorical_crossentropy'.

Upvotes: 2

Related Questions