CODEWITHSUNDEEP
pythontensorflowkeras
zcb
zcb

Reputation: 136

How to rescale images from [0,255] to [-1,1] with ImageDataGenerator?

I'm using ImageDataGenerator with flow_from_dataframe to load a dataset.

ImageDataGenerator allows you to specify a rescaling factor like this

datagen = ImageDataGenerator(rescale=1./255)

But what if I want to rescale my images in the range [-1,1]? I should do a subtraction followed by a division

images -= 128.0
images /= 128.0

How can I do these two operations in the rescale of ImageDataGenerator?

Upvotes: 0

Views: 2373

Answers (3)

Gerry P
Gerry P

Reputation: 8112

try using rescale=1/127.5-1 or you can use pre defined processor functions such as

preprocessing_function=tf.keras.applications.mobilenet.preprocess_input

Upvotes: 2

Frightera
Frightera

Reputation: 5079

Keras imagedatagen rescale parameter multiplies the data with a given value. That's why it used as 1/255.0 etc.

When you execute this:

datagen = ImageDataGenerator(rescale=1./255)

Your data's mean will be 0.5 with a range 0 to 1. Now is the tricky part, you want to set your mean to 0 firstly. For that, you can use featurewise_center and samplewise_center. Now, it is done, but your data is scaled between -0.5 and 0.5

To ensure it is between -1 and 1, we will rescale it by 2/255.0 . So the last code should look like this:

datagen = ImageDataGenerator(rescale=2/255.0,
                             featurewise_center=True,
                             samplewise_center=True,)

For more details about _center parameters, you can refer the docs

Upvotes: 0

Dwight Foster
Dwight Foster

Reputation: 352

You could use a preprocessing function like this:

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

and then in the rescale part you can just do this:

datagen = ImageDataGenerator(rescale=rescale_img())

Upvotes: 0

Related Questions