DOC: Add a page to help users transition from tf.keras.wrappers.scikit_learn

[adriangb]

We should have a page noting the differences (not features) vs. the old wrappers. I think so far we only have:

• fit/predict arguments should be passed to the constructor instead of as kwargs to fit/predict.
• Renamed build_fn to model
• No longer need to declare defaults in build_fn, instead they should be declared in the SciKeras constructor
• To get automatic one-hot encoding with categorical_crossentropy loss="categorical_crossentropy" (or loss=CategoricalCrossentropy, etc.) needs to be passed to the SciKeras constructor (even for user-compiled models) since we no longer introspect compiled models.

[stsievert]

I thought SciKeras aimed to be compatible with tf.keras.wrappers.scikit_learn?

[adriangb]

Yes, these are minor changes. I think all of these we've discussed together previously. An example with all of these changes (copied from #112):

def build_model(optimizer='adam'):
    model = Sequential()
    model.add(Flatten(input_shape=(28, 28)))
    model.add(Dense(512, activation='relu'))
    model.add(Dropout(0.2))
    model.add(Dense(10, activation='softmax'))
    model.compile(optimizer=optimizer,
              loss='categorical_crossentropy',
              metrics=['accuracy'])

    return model
    
model = KerasClassifier(build_fn=build_model)

New:

def build_model(optimizer):
    model = Sequential()
    model.add(Flatten(input_shape=(28, 28)))
    model.add(Dense(512, activation='relu'))
    model.add(Dropout(0.2))
    model.add(Dense(10, activation='softmax'))
    model.compile(optimizer=optimizer,
              loss='categorical_crossentropy',
              metrics=['accuracy'])

    return model
    
model = KerasClassifier(model=build_model, optimizer="rmsprop", loss="categorical_crossentropy")

[stsievert]

To get automatic one-hot encoding with categorical_crossentropy loss="categorical_crossentropy" (or loss=CategoricalCrossentropy, etc.) needs to be passed to the SciKeras constructor (even for user-compiled models) since we no longer introspect compiled models.

Why not specify a default of loss="categorical_crossentropy" for KerasClassifier? That'd also resolve #112, correct? Does it have any side effects?

• fit/predict arguments should be passed to the constructor instead of as kwargs to fit/predict.
• Renamed build_fn to model

I like the implementation around build_fn: it still works, but issues a warning if specified. Could the same be done for parameters passed to fit/predict?

def fit(self, X, y, classes=None, ..., **kwargs):
    for k, v in kwargs.items():
        warn(f"Set fit arguments at initialization (i.e., `BaseWrapper({k}={v})`)")
    self.set_parmas(**kwargs)

[adriangb]

Could the same be done for parameters passed to fit/predict?

def fit(self, X, y, classes=None, ..., **kwargs):
    for k, v in kwargs.items():
        warn(f"Set fit arguments at initialization (i.e., `BaseWrapper({k}={v})`)")
    self.set_parmas(**kwargs)

We could do that, I don't see why not. But I'd like to deprecate that relatively soon.

[adriangb]

def fit(self, X, y, classes=None, ..., **kwargs):
    for k, v in kwargs.items():
        warn(f"Set fit arguments at initialization (i.e., `BaseWrapper({k}={v})`)")
    self.set_params(**kwargs)

The main problem with this approach is that subsequent calls to fit/predict will use the **kwargs, even if they are not passed the second call. But maybe this is enough of an edge case to not worry about.