mobilenet.py

import keras
from keras.applications.mobilenet import MobileNet
from keras.models import Model
from keras.layers import Activation, GlobalAveragePooling2D, Dropout, Dense, Input


def get_mobilenet(input_size, alpha, weight_decay, dropout):
    input_shape = (input_size, input_size, 3)
    base_model = MobileNet(
        include_top=False, weights='imagenet', 
        input_shape=input_shape, alpha=alpha
    )
    x = base_model.output
    x = GlobalAveragePooling2D()(x)
    x = Dropout(dropout)(x)
    logits = Dense(256, kernel_regularizer=keras.regularizers.l2(weight_decay))(x)
    probabilities = Activation('softmax')(logits)
    model = Model(base_model.input, probabilities)
    
    for layer in model.layers[:-2]:
        layer.trainable = False
    
    return model