predict_classes in score_model.py results in AttributeError about talos

[ackRow]

I think it's a good idea too. I can try to code that and add an error message when using a Functional Model advising you to use the experimental feature.

@bml1g12 predict and predict_classes output different arrays (one the probabilities and the other the predicted label of the object to compare with the actual label) so I don't know why it works for you but I think it's not an optimal solution.

[mikkokotila]

This is now fixed in the current dev branch.

[matthewcarbone]

@bml1g12 Yeah you'll get this error right now if you're not using the Sequential type model. predict_classes only works with that and doesn't work with the functional model. I think what you've done is something of a hot fix that seems to work with some cases but not others. I would be very careful to be sure that predict is doing what you want it to! Otherwise, this should be fine.

[ackRow]

I was trying to use talos with a Keras Functional Model too but surprisingly Keras didn't implemented the predict_classes function on them.

So I come up with a quick fix using argmax(self.keras_model.predict(self.x_val), axis=-1) that worked for me.
(don't forget to import argmax from numpy)

However it seems that it didn't work for everybody, resulting in weird arrays full of zeros.

I finally looked up the code of predict_classes in the Keras repo and it kind of uses argmax the way I proposed.

Therefore I'm waiting for suggestions.

[mikkokotila]

This is related with #39 and #42.

My suggestion is that we add an option for those that want to use the above under a parameter experimental_functional_support=True for Scan() which simply uses the above instead of the one that is being used otherwise. We need to keep in mind it's not tested (as in #42 it becomes apparent how this might not work for all cases).

@x94carbone @ackRow what do you think about this approach?

[matthewcarbone]

I think this is ok. Sort of a use-at-your-own-risk kind of option. It weird because I see all over the internet people recommending this fix, but it seems for certain kinds of data input it totally breaks everything.

It is kinda crazy that Keras hasn't addressed this issue by now if you ask me.

[bml1g12]

y_pred = self.keras_model.predict(self.x_val) seems to be working for me (I ran a keras run with optimal parameters selected by Talos, and got similar results between the two) .

[mikkokotila]

@ackRow that's great, thanks so much. At the moment there are 4 live issues related with this, so it seems like high value target :)