Removing random_state in generator affects prediction score

[ayhyap]

While trying to stop the generator from shuffling test set images, I found that replacing
df = self.dataset_df.sample(frac=1., random_state=self.random_state)
with
df = self.dataset_df in AugmentedImageSequence.prepare_dataset() changes the performance.

That does not seem normal.

[ayhyap]

After digging a bit, I think #39 addresses this problem.
The AugmentedImageSequence was using the wrong value for 'steps' argument, making the testing process different without shuffling.