keras-transform

Library for data augmentation

ANNOUNCEMENT : I won't really work on this library anymore, the recent changes made with Keras 2.2.0 made this library obselete. Please see my blog post : https://dref360.github.io/deterministic-da/

This library provides a data augmentation pipeline for Sequence objects.

Keras-transform allows the user to specify a mask to do data augmentation in a flexible way. This is useful in many tasks like segmentation where we want the ground truth to be augmented. See simple.ipynb.

Keras-transform also works with multiple inputs, outputs by using complex masks. For example, mask=[[True,False],False] would augment the first input but not the second.

keras-transform in 10 lines

from transform.sequences import SequentialTransformer
from transform.sequences import RandomZoomTransformer, RandomVerticalFlipTransformer

seq = ... # A keras.utils.Sequence object that returns a tuple (X,y)
model = ... # A keras Model

"""
A transformer transforms the input. Most data augmentation functions are implemented in transform.sequences.
We can chain transformers together using the SequentialTransformer that takes a list of transformers.
"""
sequence = SequentialTransformer([RandomZoomTransformer(zoom_range=(0.8,1.2)),
                                  RandomVerticalFlipTransformer()])

# To augment X but not y
augmented_sequence = sequence(seq,mask=[True,False])
model.fit_generator(augmented_sequence,steps_per_epoch=len(augmented_sequence))

# To augment X and y
augmented_sequence = sequence(seq,mask=[True,True]) # Alternatively, mask=True would also work.
model.fit_generator(augmented_sequence,steps_per_epoch=len(augmented_sequence))

Contributing

Anyone can contribute by submitting a PR. Any PR that adds a new feature needs to be tested.

Example

Here's an example where X is an image and the ground truth is the grayscale version of the input. The code can be found here.