A way to wrap a mixed mode model for pipelining purposes?

[florisvdf]

Hi,

I'm looking for a way to include a mixed mode GraphConv network in one of my pipelines where I can flexibly pass hyperparameters. My current approach has been to mostly follow the structure of the MNIST example, but I have not been able to get it to work. Preferably I would like to wrap the model in a scikit-learn classifier wrapper using the tf.keras.wrappers.scikit_learn.KerasClassifier. Is this possible?

I'm suspecting that the use of mixed mode required the custom train and evaluate functions as seen in the MNIST example to support the sparse filter. Is there a straightforward workaround?

My apologies for opening another issue in such a short time window.

Cheers

[danielegrattarola]

HI,

I don't have much experience with the Sklearn wrapper but I guess that a possible workaround would be to extend KerasClassifier and re-write the two methods that you mentioned.
Unfortunately, Keras often assumes that the inputs are dense and with a fairly "strict" x, y input format where all tensors have the same first dimension, which should represent the batch size. There isn't much wiggle room outside of that. The TF2-style definition of models lets you work around those restrictions but it breaks a lot of existing features of Keras, like the wrapper.

I'm not sure if this is the answer you were looking for.
Cheers,
Daniele

[florisvdf]

Hi Daniele,

Thank you for the elaboration. I was expecting as much. If I manage to come up with a solution, I'll make sure to share it in this thread.

Cheers,
Floris