MN vs MN+LSTM

gistfile1.txt

import numpy as np
import keras
import time

from keras.layers import  TimeDistributed, GlobalAveragePooling2D, Dense, LSTM, Input

from keras.models import Model

def inference_model():
    model = keras.applications.MobileNetV2(include_top=True, weights=None)
    x = model.output
    x = keras.layers.Dense(130, activation='relu')(x)
    x = keras.layers.Dense(1, activation='sigmoid')(x)
    model = keras.models.Model(model.input, x)

    return model


def mobilenetv2lstm():
    backbone_model = keras.applications.MobileNetV2(include_top=False,
                                     weights='imagenet',
                                     alpha=0.5)

    image_input = Input(shape=[None, 224, 224, 3])

    backbone_feature_layer = TimeDistributed(backbone_model)(image_input)
    x = TimeDistributed(GlobalAveragePooling2D())(backbone_feature_layer)
    x = TimeDistributed(Dense(130, activation='elu'))(x)

    x = LSTM(16, activation='elu', return_sequences=True)(x)

    output = TimeDistributed(Dense(1, activation='sigmoid', name='predictions'))(x)
    model = Model(image_input, output)

    return model

if __name__ == '__main__':
     batch_size = 1
     seq_len = 8

     model = mobilenetv2lstm()
     input = np.random.rand(batch_size, seq_len, 224, 224, 3)

     # model = inference_model()
     # input = np.random.rand(batch_size * seq_len, 224, 224, 3)

     print('\n input shape', input.shape)

     for i in range(0, 20):
        start_time = time.time()
        model.predict(input)
        print(time.time() - start_time)