Scikeras and Keras Input layer

[phbillet]

I sent the following issue 16495.

Following the advice, I encounter an incompatibility between keras.wrappers.scikit_learn and scikeras.
With keras.wrappers.scikit_learn, I can train a model without Input layer.

With scikeras, it is not possible to do it, the program returns an error :

File ~/anaconda3/envs/EZStacking/lib/python3.8/site-packages/scikeras/wrappers.py:909, in BaseWrapper._fit(self, X, y, sample_weight, warm_start, epochs, initial_epoch, **kwargs)
    906 y = self.target_encoder_.transform(y)
    907 X = self.feature_encoder_.transform(X)
--> 909 self._check_model_compatibility(y)
    911 self._fit_keras_model(
    912     X,
    913     y,
   (...)
    918     **kwargs,
    919 )

File ~/anaconda3/envs/EZStacking/lib/python3.8/site-packages/scikeras/wrappers.py:540, in BaseWrapper._check_model_compatibility(self, y)
    535 # check if this is a multi-output model
    536 if getattr(self, "n_outputs_expected_", None):
    537     # n_outputs_expected_ is generated by data transformers
    538     # we recognize the attribute but do not force it to be
    539     # generated
--> 540     if self.n_outputs_expected_ != len(self.model_.outputs):
    541         raise ValueError(
    542             "Detected a Keras model input of size"
    543             f" {self.n_outputs_expected_ }, but {self.model_} has"
    544             f" {len(self.model_.outputs)} outputs"
    545         )
    546 # check that if the user gave us a loss function it ended up in
    547 # the actual model

TypeError: object of type 'NoneType' has no len()

So I decided to suppress the call self._check_model_compatibility(y) in the function _fit in the file wrappers.py.
Now it works fine, but it is not a good solution.
Is it possible to integrate Keras model without input layer ?

[phbillet]

I would like to use Keras model without Input layer, because I try to build stacked model (with preprocessing pipeline). due to this architecture it is very difficult to determine the input shape, because a preprocessing step can add columns...
It could be interesting to add a new boolean that controls self._check_model_compatibility(y)...

[adriangb]

I'll look into making that backwards compatible but I believe we put that there because sklearn estimators are required to expose their input size.

Could you use the meta parameter in your build_model_fn? https://www.adriangb.com/scikeras/stable/advanced.html#meta

[phbillet]

Here you can find an example of regression.
Could you help me to make it compatible with scikeras ?

[adriangb]

Could you use the meta parameter in your build_model_fn? https://www.adriangb.com/scikeras/stable/advanced.html#meta

Have you had a chance to check the documentation on how to create dynamically sized input layers?

Unfortunately the notebook you linked is a bit too complex and specific to your use case for me to rewrite and understand. If the above doesn't work, could you submit a self contained reproducibile example please?