CODEWITHSUNDEEP
tensorflowgoogle-colaboratorytf.keras
kawingkelvin
kawingkelvin

Reputation: 3971

tf.keras.preprocessing.image.ImageDataGenerator random_transform does not perform rescale

I am using colab and tf version is '2.2.0-rc2'. Using tf.keras.preprocessing.image.ImageDataGenerator to perform a single random transform using .random_transform on a single image does not rescale it (with 1/255.) Code snippet:

import tensorflow as tf
import PIL

img_gen = tf.keras.preprocessing.image.ImageDataGenerator(rescale=1./255)
img = PIL.Image.open("my_image.jpg")
img = img.resize((224, 224), PIL.Image.BICUBIC)
img_t = img_gen.random_transform(np.array(img))

img_t is still a tensor with range [0, 255] instead of [0.0, 1.0].

Is this by design or this is a bug? I don't remember the behavior before tensorflow 2.0 (i probably never use it that way).

Note: I don't see any issue if I use .flow or .flow_from_directory and their generators.

Upvotes: 0

Views: 905

Answers (1)

thushv89
thushv89

Reputation: 11333

It looks to me like does make sense. flow or flow_from_* methods and random_transform() transforms data differently.

  • The flow() methods will take the configuration/transformations provided to ImageDataGenerator and spit out data according to those parameters

  • random_transform() comes up with a random transformation configuration (based on the parameters you've setup in __init__ of your ImageDataGenerator and apply that to the data. This means your rescale parameter might be overridden by default.

More info on random_transform()

I haven't used random_tranform() myself before. But it seems this is how it works.

Let's say we define the following ImageDataGenerator

img_gen = tf.keras.preprocessing.image.ImageDataGenerator(
    rescale=1./255, vertical_flip=True, brightness_range=(-0.1,0.1)
)

Now you can see what kind of random transformations are available using,

print(img_gen.get_random_transform((1,224,224,3)))

which will give different random values for vertical_flip and brightness. As you can see rescale is apparently not a random parameter they consider. This explains why you didn't get the expected result.

Upvotes: 1

Related Questions