Changes to Dataset in for loop don't work

Question

I’m trying to augment my dataset by randomly changing the hue of its images within a for loop but the changes do not persist outside of the loop. I imported the dataset with tf.keras.utils.image_dataset_from_directory. The rest of the code looks as follows:


def augment(image, label, counter):
  randNr = tf.random.uniform(shape=(), minval=-1, maxval=1, dtype=tf.dtypes.float32)
  image = tf.image.adjust_hue(image, delta=randNr)
  #desplay some values
  if(counter<1):
    print(randNr)
    plt.figure()
    plt.imshow(image[0].numpy().astype("uint8"))
    plt.show()
  return image, label

temp1 = 0
for image, label in v_dataset:
  image, label = augment(image, label, temp1)
  #desplay some values
  if(temp1<1):
    plt.figure()
    plt.imshow(image[0].numpy().astype("uint8"))
    plt.show()
  temp1 += 1

#display some values
plt.figure(figsize=(10, 10))
for images, labels in v_dataset.take(1):
    print("images shape: ", np.shape(images))
    for i in range(9):
        ax = plt.subplot(3, 3, i + 1)
        plt.imshow(images[i].numpy().astype("uint8"))
        plt.title(int(labels[i]))
        plt.axis("off")
plt.show()

When I print an image the first two times, the hue has changed as intended. When I print out more images later, however, none of them have a variation in hue. Any Ideas on why this occurs and how to fix it?

Asked By: Rumpel Stilzchen

||

Source

Answer 1

You can augment the image inside the model:

inputs = Input(shape=(...))
x = augment_hue(inputs)
...

If you use a Sequential model, use a Lambda layer:

Lambda(lambda x: augment_hue(x))

Answered By: AndrzejO

Answer 2

I did not find a working solution for this problem. In the end I just modified the data before importing it, using the keras ImageDataGenerator.

Answered By: Rumpel Stilzchen

Answer 3

You could modify dataset by calling dataset.map(...). However, everything inside .map(...) will be compiled, so using random number generator from python will just yield the same result every time (the one which was traced). Your options are:

use tf.py_function
wrap your function in tf.autograph.experimental.do_not_convert
use keras preprocessing layer e.g. https://www.tensorflow.org/api_docs/python/tf/keras/preprocessing/image/random_channel_shift

Answered By: Goofy_Goof

Changes to Dataset in for loop don't work

Question:

Answers: