from keras.models import Model
from keras.applications.vgg16 import VGG16
from keras.layers import GlobalAveragePooling2D

# Wait for downloading for the 1st time,
# the weights is saved in ~/.keras/models
vgg16 = VGG16(include_top=False,  # don't include 3 fully connected layers
              weights='imagenet', # use pre-trained weights, not random initialization
              pooling=None)       # don't apply any pooling at the last convolutional layer

# Continue building your own model
cst = nn.get_layer('block2_pool').output
cst = GlobalAveragePooling2D()(cst)
custom_vgg = Model(inputs=vgg16.input, outputs=cst)

# Want to infer some image?
# x = read_some_image, RGB, [0, 255], uint8
x = x.astype('float')
x = x.reshape(1, x.shape[0], x.shape[1], x.shape[2])  # to 4D tensor
x = preprocess_input(x)
y = custom_vgg.predict(x)