Using Keras and pytorch with make_reduction

[HishamSalem]

Hello,

I was wondering if there is a way to utilize sktime's make_reduction with the estimator being a model from keras/pytorch?

if there is another function that does that, I would love to know.

[fkiraly]

@HishamSalem, that should be perfectly possible, although the estimator needs to conform to the sklearn regressor interface (for supervised regressors) or to the sktime regressor interface (for time series regressors).

For keras, you should be able to do this with keras.wrappers.scikit_learn.KerasRegressor, i.e., wrap with that first before passing to make_regressor. Would be interested to hear about your results, e.g., whether it works well with reduction and pipelines.

For pytorch, you can use skorch https://github.com/skorch-dev/skorch, that will conform to the sklearn interface.

[HishamSalem]

@fkiraly , It does work well. A challenge however would have to be how we shape the data in terms of inputs. For example, it works for an MLP however when working with CNN and LSTMS it ends up being more interesting since we have to reshape the layers from the inputs.

I did try it with GRUs but I get a recursion error.

[fkiraly]

hm, interesting - did not see this reply yet, @HishamSalem, apologies.

FYI @AurumnPegasus

[anonymous-engineering]

Any updates on this?

[fkiraly]

it should be possible out of the box, as long as the estimators conform with the sklearn interface.

Regarding @HishamSalem's issue, I can't really say what was going on there due to lack of code - I was hoping for a bug report.

As said, as long as the estimator is sklearn compatible, you should be able to stick it into make_reduction.