tf_cpu_gpu_hue_adjust.py

import numpy as np
import tensorflow as tf

from time import time

num_iter = 1000 
batch_size = 64
height = 256 
width = 256
channels = 3
total_images = num_iter * batch_size

seed = 42
np.random.seed(seed)
tf.set_random_seed(seed)

imgs = np.random.randint(256, size = (batch_size, height, width, channels), dtype = np.uint8)

def body(dev):
  def inner(x, y):
    with tf.device(dev):
      adjusted = tf.image.adjust_hue(y, 0.1)
    return (tf.add(x, 1), adjusted)
  return inner

cond = lambda x, y: tf.less(x, num_iter)

x = tf.Variable(tf.constant(0))

init_op = tf.global_variables_initializer()
init_op_2 = tf.local_variables_initializer()

with tf.Session() as sess:
  sess.run(init_op)
  sess.run(init_op_2)
  
  def time_it(dev_type, dev_number):
    dev = "/" + dev_type + ":" + str(dev_number)
    start = time()
    result = tf.while_loop(cond, body(dev), [x, imgs])
    sess.run(result)
    total_time = time() - start
    time_per_image = total_time / total_images
    images_per_sec = 1.0 / time_per_image
    print("\n\nDevice: %s" % dev)
    print("Total images: %d" % total_images)
    print("Total time: %f seconds\n" % total_time)
    print("Time per image (us): %.16f" % (time_per_image * 1e6))
    print("Images/sec: %.2f\n" % images_per_sec)
    return images_per_sec

  im_sec_gpu = time_it("gpu", 0)
  im_sec_cpu = time_it("cpu", 0)
  print("\n\nCPU/GPU acceleration factor: %.2f\n\n" % (im_sec_gpu / im_sec_cpu))

  hue_delta = 0.1

  # Numeric test
  with tf.device("/cpu:0"):
    cpu_result = tf.image.adjust_hue(imgs, hue_delta).eval()
  with tf.device("/gpu:0"):
    gpu_result = tf.image.adjust_hue(imgs, hue_delta).eval()

  print("\nCPU result (hue delta: %d):\n\n%s\n" % (hue_delta, cpu_result[0:10, 0:2, 0]))
  print("\nGPU result (hue delta: %d):\n\n%s\n" % (hue_delta, gpu_result[0:10, 0:2, 0]))
  print("\nDifference:\n\n%s\n" % (cpu_result[0:10, 0:2, 0] - gpu_result[0:10, 0:2, 0]))
  print("\nreduce_sum of difference (all pixels):\n\n%s\n" % (tf.reduce_sum(cpu_result - gpu_result).eval()))
  print("max difference (all pixels): %s\n\n" % (tf.reduce_max(cpu_result - gpu_result).eval()))

  new_delta = 0.5

  # Random hue adjustment
  with tf.device("/cpu:0"):
    cpu_result = tf.image.random_hue(imgs, new_delta).eval()
  with tf.device("/gpu:0"):
    gpu_result = tf.image.random_hue(imgs, new_delta).eval()

  rand_diff_cpu = imgs - cpu_result
  rand_diff_gpu = imgs - gpu_result

  print("CPU random hue vs original:\n%s\n\n" % rand_diff_cpu[0:5, 0:5, 0])
  print("GPU random hue vs original:\n%s\n\n" % rand_diff_gpu[0:5, 0:5, 0])