Improve train_test_split

[cancan101]

For NeuralNet:

1. Why use your own implementation of train_test_split rather than using sklearn.cross_validation::train_test_split?
   -- This would fix train_test_split does not work for Pandas Series with non Dense Index #7 because it uses safe_indexing.
2. Allow the user to pass in, through composition (like the batch iterator), a class responsible for performing the train vs valid split. Right now the only option is to subclass NeuralNet which is not ideal.

[cancan101]

I am currently using this:

class NeuralNetFix(NeuralNet):
    def train_test_split(self, X, y, eval_size):
        assert eval_size is None
        X_train = X
        y_train = y
        X_valid = self.X_valid
        y_valid = self.y_valid
        if not self.regression and self.use_label_encoder:
            y_valid = self.enc_.transform(y_valid).astype(np.int32)
        return X_train, X_valid, y_train, y_valid

I then handle the train_test_split myself.

UPDATE: I no longer read X_valid and y_valid from self.

[dnouri]

@cancan101 To answer your first question: You'll notice that NeuralNet.train_test_split uses StratifiedKFold for classification tasks. That's different to what sklearn's train_test_split would give you.

But maybe the stratified split isn't all that important, and we can just use an overridable train_test_split (component) by default.

[dnouri]

I think we can merge this discussion with #42, and close this one. Also consider #45 when thinking about a better way to override the train_test_split method.