Augment.py

import tensorflow as tf

# Define the input layer
input_layer = tf.placeholder(tf.float32, shape=[None, None, None, 3])

# Create a list of rotated images
rotations = [tf.contrib.image.rotate(input_layer, angles=angle) for angle in range(-10, 11, 2)]

# Create a list of random hue adjustments
hue_deltas = [tf.image.random_hue(input_layer, max_delta=0.03 * i) for i in range(0, 5)]

# Create a list of random brightness adjustments
brightness_deltas = [tf.image.random_brightness(input_layer, max_delta=0.1 * i) for i in range(0, 5)]

# Create a list of horizontally flipped images
flips = [tf.image.flip_left_right(input_layer)]

# Initialize an empty list to store the augmentations
augmentations = []

# Iterate over the rotated images
for rotated in rotations:
  # Apply the hue and brightness adjustments and flipping to each rotated image
  for hue in hue_deltas:
    for brightness in brightness_deltas:
      for flip in flips:
        augmentations.append(hue(brightness(flip(rotated))))

# Concatenate the augmentations along the batch dimension
augmented_input = tf.concat(augmentations, axis=0)