Keras: Keras Custom Loss Function

[Erstwild]

I am working on a port of this project: https://github.com/jaredleekatzman/DeepSurv to use Keras instead of Lasagne.

My question centers upon how to best integrate a new custom loss function (cox proportional hazards regression for example) into the Keras paradigm:

def negative_log_likelihood(ytime, ystatus):
    LL_i = T.switch(T.eq(ystatus[i],1), self.theta - T.log(T.sum(self.exp_theta * T.gt(ytime, ytime[i]))),0)

Any pointers to get started would be of great assistance and sincerely appreciated!

[kuza55]

You probably want to do something like

model.compile(optimizer, negative_log_likelihood, ...)

[JonnoFTW]

This doesn't work if you want to save and reload a model.

[Erstwild]

Interesting. What would be the best way to approach this to still use the standard model persistence options in Keras?

[JonnoFTW]

It could be fixed easily by allowing the load_model function to take kwargs that match the model.compile arguments and override those from the deserialized model.

[cicobalico]

To reload the model saved with custom loss function try

python load_model(model_path, custom_objects={"negative_log_likelihood":negative_log_likelihood}) ```

[Erstwild]

I'll certainly investigate the custom objects approach and get back to you all when I get the chance.

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs, but feel free to re-open it if needed.

[tomcat123a]

This is also a problem for many other machine learning problems where the loss function could not be written as sum of separable parts and each of the components contains only one data sample.