How to use featurewise_center=True together with flow_from_directory in ImageDataGenerator?

Question

I set the featurewise_center = True and then use flow_from_directory to set up my training and validation data in keras. However, i got the error

UserWarning: This ImageDataGenerator specifies `featurewise_center`, 
but it hasn't been fit on any training data. Fit it first by calling `.fit(n
numpy_data)`

Is there any means I can use flow_from_directory and then to fit the data as required ?

Answer 1

featurewise_center transforms the images to 0 mean. This is done by using the formulae

X = X - mean(X)

But for the ImageDataGenerator to do this transformation it needs to know the mean of the dataset and fit method on the ImageDataGenerator does exactly this operation of calculating these statistics.

As the keras docs explain

Fits the data generator to some sample data. This computes the internal data stats related to the data-dependent transformations, based on an array of sample data.

If the dataset can be fully loaded into the memory, we can do so by loading all the images into a numpy array and running the fit on it.

Sample code (RGB images of 256x256) :

from keras.layers import Input, Dense, Flatten, Conv2D
from keras.models import Sequential
from keras.preprocessing.image import ImageDataGenerator
import numpy as np
from pathlib import Path
from PIL import Image

height = width = 256 

def read_pil_image(img_path, height, width):
        with open(img_path, 'rb') as f:
            return np.array(Image.open(f).convert('RGB').resize((width, height)))

def load_all_images(dataset_path, height, width, img_ext='png'):
    return np.array([read_pil_image(str(p), height, width) for p in 
                                    Path(dataset_path).rglob("*."+img_ext)]) 

train_datagen = ImageDataGenerator(featurewise_center=True)
train_datagen.fit(load_all_images('./images/', height, width))

train_generator = train_datagen.flow_from_directory(
        './images/',
        target_size=(height, width),
        batch_size=32,
        class_mode='binary',
        color_mode='rgb')

model = Sequential()
model.add(Conv2D(1,(3,3), input_shape=(height,width,3)))
model.add(Flatten())
model.add(Dense(1))
model.compile('adam', 'binary_crossentropy')

model.fit_generator(train_generator)

But what if the data cannot be fully loaded into memory ? One approach is to sample the images randomly from the dataset.

Normally we use mean of training data only to do mean normalization and use the same mean for normalization validation/test data. It will be bit tricky to do the same via the datagenerator.