This is helper function to extract convolutional features of images using pre-trained VGG16 model. This makes it possible to do feature extraction of big size of images on a CPU machine.

extract_vgg_features.py

import numpy as np
import bcolz
import csv
import shutil

def save_array(fname, arr):
    """Helper function to save numpy array arr to fname(.dat)"""
    c=bcolz.carray(arr, rootdir=fname, mode='w')
    c.flush()
    
def load_array(fname):
    """Helper function to load numpy arry from file fnmame(.dat)"""
    return bcolz.open(fname)[:]

def extract_features(path_in, path_out, model):
    """This function extract features using the given convolutional model
    to be used as input features for upper layer models.
    
    Args:
        path_in (str): directory path of input images
        path_out (str): directory path of output dat files 
        model (Model): model with convolutional layers to extract features
        
    Returns:
        Features extracted and labels. Also will save them as dat files
        at ``path_out``
    
    """
    # creates directory `path_out` if not already existed
    if not os.path.exists(path_out):
        os.makedirs(path_out)
    batches = get_batches(path_in, batch_size=1, shuffle=False)
    labels = batches.classes
    n = batches.samples
    save_array(path_out+'labels.dat', labels)
    
    # due to the limited ram space, we will the features of samples one by one
    for i in range(n):
        features = model.predict(next(batches)[0])
        save_array(path_out+'{}.dat'.format(i), features)
    
    # now merge the above files to one array
    features = np.empty_like(features)
    for i in range(n):
        f = load_array(path_out+'{}.dat'.format(i))
        features = np.concatenate((features, f), axis=0)
    
    # features[0] is a placeholder of all zeros, we want to remove it
    # before saving 
    features = np.delete(features, 0, 0)
    save_array(path_out+'features.dat', features)
    
    # delete the individual dat files
    for i in range(n):
        shutil.rmtree(path_out+'{}.dat'.format(i))
        
    return features, labels