Improve data augmentation with shifted images

[wroscoe]

Here is a great example: https://chatbotslife.com/using-augmentation-to-mimic-human-driving-496b569760a9#.b6jyrypqu

[alanswx]

I don't know if this improves things, or if this augmentation is quite right, but you can do a lot of it with one call:

         X_train, X_validation, y_train, y_validation = train_test_split(X, Y, test_size=0.20, random_state=42)
         datagen = ImageDataGenerator(
            rotation_range=1,
                width_shift_range=0.2,
                height_shift_range=0.2,
                shear_range=0.2,
                zoom_range=0.02,
                horizontal_flip=False)


         # compute quantities required for featurewise normalization
         # (std, mean, and principal components if ZCA whitening is applied)
         datagen.fit(X_train)

         # fits the model on batches with real-time data augmentation:
         m.fit_generator(datagen.flow(X_train, y_train, batch_size=32, shuffle=True),
                samples_per_epoch=len(X_train)*3, nb_epoch=20,callbacks=callbacks_list,
                validation_data=(X_validation, y_validation)
                )

[wroscoe]

Parts of that will work. We'll want to validate on non-augmented data.

[alanswx]

That snippet does validate on non-augmented data. (I believe)

[wroscoe]

You're right @alanswx but it doesn't switch the steering angles when you flip the image horizontally.

[alanswx]

that is why i turned off the flipping code -- once i get my car driving i will check in a class that does all the augmentation

[wroscoe]

I've moved the augmentation code to the donkeybarn repo. https://github.com/autorope/donkeybarn/blob/master/donkeybarn/cv/augment.py