Gist for mobilenet to use in https://predictiveprogrammer.com/famous-convolutional-neural-network-architectures-2/

mobilenet.py

import keras.layers as ls #import Conv2D, DepthwiseConv2D, BatchNormalization, Dense
from keras.models import Model
from functools import partial

def depthwise_separable_conv(in_tensor, filters_pw, strides):
    l1_1 = ls.DepthwiseConv2D(3, strides=strides,
                            depth_multiplier=1, padding='same')(in_tensor)
    l1_2 = ls.BatchNormalization()(l1_1)
    l1_3 = ls.Activation('relu')(l1_2)

    l2_1 = ls.Conv2D(filters_pw, 1, strides=1, padding='same')(l1_3)
    l2_2 = ls.BatchNormalization()(l2_1)
    l2_3 = ls.Activation('relu')(l2_2)
    return l2_3

def conv(in_tensor, filters, strides=2):
    l1 = ls.Conv2D(filters, 3, strides=strides, padding='same')(in_tensor)
    l2 = ls.BatchNormalization()(l1)
    l3 = ls.Activation('relu')(l2)
    return l3

dws_conv_s1 = partial(depthwise_separable_conv, strides=1)
dws_conv_s2 = partial(depthwise_separable_conv, strides=2)

def mobilenet(in_shape=(224, 224, 3), include_top=True):
    in_ = ls.Input(in_shape)
    conv1 = conv(in_, 32)

    dws_conv1 = dws_conv_s1(conv1, 64)
    dws_conv2 = dws_conv_s2(dws_conv1, 128)
    dws_conv3 = dws_conv_s1(dws_conv2, 128)
    dws_conv4 = dws_conv_s2(dws_conv3, 256)
    dws_conv5 = dws_conv_s1(dws_conv4, 256)
    dws_conv6 = dws_conv_s2(dws_conv5, 512)

    dws_conv11 = dws_conv6

    for _ in range(5):
        dws_conv11 = dws_conv_s1(dws_conv11, 512)

    dws_conv12 = dws_conv_s2(dws_conv11, 1024)
    dws_conv13 = dws_conv_s1(dws_conv12, 1024)

    if include_top:
        pool = ls.GlobalMaxPooling2D()(dws_conv13)
        reshape = ls.Reshape((1, 1, 1024))(pool)
        res = ls.Dense(1000, activation='softmax')(reshape)
    else:
        res = dws_conv13

    model = Model(in_, res)
    return model