CODEWITHSUNDEEP
image-processingkerasgenerative-adversarial-network
Okafor Zuruoke
Okafor Zuruoke

Reputation: 137

Can Keras ImageDataGenerator rescale an image between -1 and 1

I'm currently building a GAN for images on my local directory. So I'm using the Keras.ImageDataGenerator flow_from_dir constructor.

I'd want to normalize my images between -1 and 1 that's the convention for GANs cause of the tanh activation.

I'm having problem in rescaling the image like implementing 1/127.5 -1 in the rescale argument.

    from tensorflow.keras.preprocessing.image import ImageDataGenerator
    fid = drive.ListFile({'q':"title='NM_cycleGAN.zip'"}).GetList()[0]['id']
    f = drive.CreateFile({'id': fid})
    f.GetContentFile('NM_cycleGAN.zip')
    PATH = '/content/NM_CycleGAN'
    train_A_dir = os.path.join(PATH, 'Train_A')
    train_A = os.path.join(train_A_dir, 'Negroid')
    trainA_image_generator = ImageDataGenerator(rescale=1./127.5 - 1)

    train_A = trainA_image_generator.flow_from_directory(batch_size=batch_size,
                                                       directory=train_A_dir,
                                                       shuffle=True, seed=1,
                                                       target_size=(128, 128),
                                                       class_mode=None)
    inpA = next(iter(train_A))


    print(inpA[0].min())


   -253

Upvotes: 0

Views: 2054

Answers (1)

Zabir Al Nazi Nabil
Zabir Al Nazi Nabil

Reputation: 11218

First of rescale parameter multiplies the data with a scalar, usually we use 1/255.

This makes the mean of the data at 0.5 with range 0. to 1.

After that, we can use featurewise_center and samplewise_center to make sure our mean is at 0., which will ensure the data has range -0.5 to 0.5

Now, with these in mind, you can choose your proper scaling factor.

tf.keras.preprocessing.image.ImageDataGenerator(
    featurewise_center=True, samplewise_center=True,
    rescale = 2/255.
)

With these parameter you'll get the desired behaviour. A small snippet running the datagen:

x = np.random.randint(0,255, (10,224,224,3))

datagen = tf.keras.preprocessing.image.ImageDataGenerator(
    featurewise_center=True, samplewise_center=True,
    rescale = 2/255.
)

for x_i in datagen.flow(x, batch_size = 5):
  print(np.mean(x_i))
  print(np.min(x_i))
  print(np.max(x_i))

  break

Out:

-1.0709576e-08
-0.9992176
0.9986379

As you can see the mean is 0, the min is -1 and max is +1.

ref: https://www.tensorflow.org/api_docs/python/tf/keras/preprocessing/image/ImageDataGenerator

Upvotes: 4

Related Questions