[mhanus]

Keras uses the end of the data set as validation, and only randomizes it if the "shuffle" argument is set to True [1].

[1]: https://keras.io/getting-started/faq/#how-is-the-validation-...

[argonaut]

Except the second half of the data is the flipped of the first half (X = np.concatenate([X, X_flipped]))

[wroscoe]

Well shit. Thanks for pointing that out. I'll revise.