What is the primary difference between the reverse() and reverse_sequence() in tensorflow?

Question

I am using simple tensorflow example to flip the image. I had used the reverse_sequence as well as the reverse method and the result is same.If just using reverse() method we can flip the image then why should we use the reverse_sequence() method? I Just want to know that what is the main primary difference between this methods? Thanks in advance :)

import tensorflow as tf
import matplotlib.image as mpimg
import matplotlib.pyplot as plt

# First, load the image again
filename = "MarshOrchid.jpg"
image = mpimg.imread(filename)

# Create a TensorFlow Variable
x = tf.Variable(image, name='x')
height, width, depth = image.shape

model = tf.initialize_all_variables()

with tf.Session() as session:
    # x= tf.reverse(x, dims=[False, True, False],name="reverse")
    x = tf.reverse_sequence(x, [width] * height, 1, batch_dim=0)
    session.run(model)
    result = session.run(x)
    print(session.run(x))

plt.imshow(result)
plt.show()

Answer 1

The tf.reverse_sequence() op is designed to be used on sequential data that has been padded to make a dense tensor. Consider the following matrix, x, in which the non-zero elements appear to be "left-justified":

x = [[1 2 3 4 0 0 0]
     [1 2 3 0 0 0 0]
     [1 2 3 4 5 6 7]]

seq_lens = [4, 3, 7]

Evaluating tf.reverse_sequence(x, seq_lens, seq_dim=1, batch_dim=0) gives:

result = [[4 3 2 1 0 0 0]
          [3 2 1 0 0 0 0]
          [7 6 5 4 3 2 1]]

Note that the result still appears to be "left-justified".

By contrast, if you evaluate tf.reverse(x, [False, True]), the sequence lengths are ignored and you get a "right-justified" result:

result = [[0 0 0 4 3 2 1]
          [0 0 0 0 3 2 1]
          [7 6 5 4 3 2 1]]

Answer 2

Reading the documentation of reverse and reverse_sequence is that the former operates on any axis on the whole axis, while the latter always slices the first dimension (the batch dimension) and then reverses only up to _seq_lens_ elements of that tensor.