CODEWITHSUNDEEP
pythontensorflowimage-processingkeras
Stat Tistician
Stat Tistician

Reputation: 883

ImageDataGenerator rescaling to [-1,1] instead of [0,1]

I am using Keras Tensorflow ImageDataGenerator and usually it is used with rescaling factor 1./255 to rescale the initial values from 0 to 255 to 0 to 1 instead. However, I would like to rescale it to -1,1 range.

So instead of:

train_image_generator = ImageDataGenerator(
    rescale=1./255,
)

I tried:

train_image_generator = ImageDataGenerator(
    rescale=((1./127.5)-1)
)

Next this would be applied to a directory:

train_datagen = train_image_generator.flow_from_directory(
  directory=training_dir,
  target_size=(x, y),
  shuffle=True,
  batch_size=x,
  class_mode='binary'
)

Checking some values could be done like this:

train_datagen[1]

But according to the documentation this is a factor we use to multiply the data by the value provided. So it can be only a factor which is used to multiply, so it makes no sense to substract 1 here, because the data is rescaled using the value -0.99215686275 and indeeed when I check the actual values I can see negative values like -130.xx. So this does not work. As I need it to be rescaled to -1,1 and not 0,1 since I want to use the pretrained MobileNet V2 later on, my question is, how can I do this?

I am not talking about a way of avoiding using ImageDataGenerator:

def format_example(image, label):
  image = tf.cast(image, tf.float32)
  image = (image/127.5) - 1
  image = tf.image.resize(image, (IMG_SIZE, IMG_SIZE))
  return image, label

and so on.

So I would like to use ImageDataGenerator. So I could rescale by 1.0/127.5, but still I need to substract 1. Is there a way to later substract 1 from the values in the train_datagen? Something like

train_datagen.actualvalues-1

(I know this is not working.)

Furthermore: I need a solution which works together with image augmentation, so usually I have:

train_image_generator = ImageDataGenerator(
    horizontal_flip=True,
    rotation_range=40,
    width_shift_range=0.2,
    height_shift_range=0.2,
    fill_mode="nearest",
    zoom_range=0.2,
    rescale=((1./127.5)-1) #1./255,
)

Now, my only problem is that rescale=((1./127.5)-1) doesn't work. How can I solve this, how can I rescale to [-1,1] instead of [0,1]?

Upvotes: 1

Views: 2484

Answers (1)

l4morak
l4morak

Reputation: 311

Use preprocessing_function parameter.

def prep_fn(img):
    img = img.astype(np.float32) / 255.0
    img = (img - 0.5) * 2
    return img

gen = ImageDataGenerator(
    preprocessing_function=prep_fn
)

Upvotes: 10

Related Questions