TensorFlow image operations for batches

Question

There are a number of image operations in TensorFlow used for distorting input images during training, e.g. tf.image.random_flip_left_right(image, seed=None) and tf.image.random_brightness(image, max_delta, seed=None) and several others.

These functions are made for single images (i.e. 3-D tensors with shape [height, width, color-channel]). How can I make them work on a batch of images (i.e. 4-D tensors with shape [batch, height, width, color-channel])?

A working example would be greatly appreciated!

Answer 1

One possibility is to use the recently added tf.map_fn() to apply the single-image operator to each element of the batch.

result = tf.map_fn(lambda img: tf.image.random_flip_left_right(img), images)

This effectively builds the same graph as keveman suggests building, but it can be more efficient for larger batch sizes, by using TensorFlow's support for loops.

Answer 2

You can call the image operation in a loop and concatenate the result. For example :

transformed_images = []
for i in range(batch_size):
  transformed_images.append(tf.image.random_flip_left_right(image[i, :, :, :]))
retsult = tf.stack(transformed_images)