# fine tune model from Inception block 6

GAP_LAYER = fixed_layers

for layer in base_model.layers[:GAP_LAYER+1]:

layer.trainable = False

for layer in base_model.layers[GAP_LAYER+1:]:

inception = InceptionV3(weights='imagenet',include_top=False)

inception_out = inception.output

x = layers.GlobalAveragePooling2D()(inception_out)

x = layers.Dense(1024, activation='relu', name='feature_out')(x)

predictions = layers.Dense(2, activation='sigmoid')(x)

model = Model(inputs=inception.input, outputs=predictions)

# transfer_learning_setup(inception)

fine_tune_setup(inception)

lstm_input = layers.Input(shape=(input_dim, feature_dim), name='feature_input')

lstm = layers.LSTM(input_dim)(lstm_input)

cls_out = layers.Dense(output_dim, activation='sigmoid')(lstm)

# lstm.add(layers.Dense(2, activation='sigmoid'))

lstmModel = Model(inputs=lstm_input, outputs=cls_out)

# lstmModel.compile(optimizer=Adam(lr=0.001, decay=1e-6), loss='categorical_crossentropy', metrics=['accuracy'])

frames_input = layers.Input(shape=(None, 299, 299, 3), dtype='float32', name='frame_input')

InceptionModel = Inception_model()

InceptionModel.load_weights('./out/Inception_weight.h5')

feature_model = Model(inputs=InceptionModel.input, outputs=InceptionModel.get_layer('feature_out').output)

# apply feature model to each time step of input

feature_output = layers.TimeDistributed(feature_model, input_shape=(299, 299, 3))(frames_input)

# apply lstm model to extracted feature

lstmModel = LSTM_model(time_step, feature_dim, output_dim)

# mask padding value

# feature_output = layers.Masking(mask_value=0., input_shape=(time_step, feature_dim))(feature_output)

model_output = lstmModel(inputs=feature_output)

end_to_end_model = Model(inputs=frames_input, outputs=model_output)

end_to_end_model._get_distribution_strategy = lambda: None