Apart from the ongoing discussion on saving/loading of DL models (issue #3022). The reason current PR fails to even save from self.model_.save() is due to keras.layers.Lambda() which is incompatible with (de)serialization attempts.

I could subclass a layer instead of using a Lambda to avoid this issue(this is the suggested way as per the previous link).

@fkiraly If this seems like a reasonable change, let me know so I could start working on it.

Currently, the two failing tests are being skipped for all DL models until a solution for the save/load issue is found.

You can add the network to the skip list in tests._config.

Quick question: how common is it to install keras-self-attention if you already have tensorflow or keras? Should this be a core dl dependency or just a soft dependency?

I'm aware there are other networks that make use of keras-self-attention in sktime, but I believe we do not need to have it as a core dl dependency yet. We could add it as so if we have more models that make use of it in the future. Personally, I've never used this myself and have seen it only a few times before this.

ok, let's make it a soft dependency then

I see I was missing a warning-level soft dependency check for the network. I've added that, I think it should be good now.

On a side note MatrixProfileTransformer has no explicit error-level check in the constructor. Should it be this way?

On a side note MatrixProfileTransformer has no explicit error-level check in the constructor. Should it be this way?

Yes, it's done by the python_dependencies tag and super.__init__.