x = Dropout(0.15)(x)
tStream = Model(inputs=sStream.input, outputs=x)
# Create full 2 stream network
spatial_average = Lambda(function=lambda x: K.mean(x, axis=1),
output_shape=lambda shape:
(shape[0], ) + shape[2:])(sStream.outputs[-1])
spatial_out = Dense(classes, activation='sigmoid')(spatial_average)
temporal_out = Dense(classes, activation='sigmoid')(tStream.outputs[-1])
averaged = Average()([spatial_out, temporal_out])
model = Model(inputs=sStream.input, outputs=averaged)
model.summary()
return model
if __name__ == "__main__":
from training import TrainingSuite
training_suite = TrainingSuite(BATCH_SIZE, EPOCHS, TRAIN_DIR, TEST_DIR,
FRAME_LENGTH, FRAME_WIDTH, FRAME_NUM,
preprocess_input)
model = TS_CNN_LSTM(INPUT_SHAPE, KERNEL_SHAPE, POOL_SHAPE, CLASSES)
training_suite.evaluation(model, WEIGHT_FILE_NAME, gen_logs=False)
model.add(TimeDistributed(Conv2D(filters=64, kernel_size=kernel_shape)))
model.add(TimeDistributed(Activation("relu")))
model.add(TimeDistributed(BatchNormalization()))
model.add(TimeDistributed(MaxPooling2D(pool_size=pool_shape)))
model.add(TimeDistributed(Conv2D(filters=64, kernel_size=kernel_shape)))
model.add(TimeDistributed(Activation("relu")))
model.add(TimeDistributed(BatchNormalization()))
model.add(TimeDistributed(MaxPooling2D(pool_size=pool_shape)))
model.add(TimeDistributed(Flatten()))
model.add(LSTM(512, return_sequences=True))
# Temporal Mean Pooling
model.add(
Lambda(function=lambda x: K.mean(x, axis=1),
output_shape=lambda shape: (shape[0], ) + shape[2:]))
model.add(Dense(classes, name='output', activation='sigmoid'))
model.summary()
return model
if __name__ == "__main__":
from training import TrainingSuite
training_suite = TrainingSuite(BATCH_SIZE, EPOCHS, TRAIN_DIR, TEST_DIR,
FRAME_LENGTH, FRAME_WIDTH, FRAME_NUM)
model = cnn_lstm(INPUT_SHAPE, KERNEL_SHAPE, POOL_SHAPE, CLASSES)
training_suite.evaluation(model, WEIGHT_FILE_NAME)