Thread by @fchollet

A thing I hear sometimes: "what if my loss doesn't match the signature loss = fn(y_true, y_pred)?"

This is not a requirement in Keras -- it's only the default setting. If you have a loss with multiple inputs/targets, here are your options, in order of preference: (a thread)

1. Use a layer that calls `self.add_loss(value)`. Docs:
keras.io/api/losses/#th…

And if you need access to labels, use the "endpoint layer" pattern: keras.io/examples/keras…

2. Override the `train_step` method of your Model class to implement your custom loss. Guide: keras.io/guides/customi…

A few examples:

Metric learning: keras.io/examples/visio…
Vanilla GAN: keras.io/examples/gener…
WGAN-GP: keras.io/examples/gener…
CycleGAN: keras.io/examples/gener…

3. Write a custom training loop from scratch. Guide: keras.io/guides/writing…

Remember that Keras defaults are only meant to cover the most common workflows -- for anything else, you have a range of increasingly advanced options.

We call it "progressive disclosure of complexity"