def _create_model(img_height, img_width, input_timesteps, compile_params, weights):
if weights == True:
weights = 'imagenet'
channels = 3
else:
weights = None
channels = 1
# Prepair
input_shape = (input_timesteps, img_height, img_width, channels)
# Model architecture
source = Input(
name='seed', shape=input_shape, dtype=tf.float32)
vgg16_encoder_block = vgg16_encoder(input_shape=input_shape[1:], weights=weights)
net = TimeDistributed(vgg16_encoder_block)(source)
net = ConvLSTM2D(filters=128, kernel_size=3, padding='same', return_sequences=True)(net)
net = BatchNormalization()(net)
net = ConvLSTM2D(filters=1, kernel_size=3, padding='same', return_sequences=False)(net)
net = BatchNormalization()(net)
#net = Reshape(target_shape=(64, 64, 1))(net)
vgg16_decoder = resnet_up(net, scale=4)
vgg16_decoder = BatchNormalization()(vgg16_decoder)
vgg16_decoder = resnet_up(vgg16_decoder, scale=4)
vgg16_decoder = BatchNormalization()(vgg16_decoder)
predicted_img = Activation('sigmoid')(vgg16_decoder)
model = Model(inputs=[source], outputs=[predicted_img])
# Compile model
model = compile_model(model, compile_params)
return model
def _create_model(img_height, img_width, input_timesteps, compile_params):
input_shape = (input_timesteps, img_height, img_width, 1)
x = Input(shape=input_shape, name='input')
encoder_input_shape = input_shape[1:]
encode_block = SkipConvLSTMSingleOutput._create_encoder(
encoder_input_shape)
net = TimeDistributed(encode_block)(x)
net = ConvLSTM2D(filters=128,
kernel_size=3,
padding='same',
return_sequences=True)(net)
net = BatchNormalization()(net)
hidden = ConvLSTM2D(filters=80,
kernel_size=3,
padding='same',
return_sequences=False)(net)
hidden = BatchNormalization()(hidden)
decode_block = SkipConvLSTMSingleOutput._create_decoder(
hidden.shape[1:], encode_block)
net = decode_block([hidden, Lambda(lambda x: x[:, -1, :, :, :])(x)])
net = Activation('sigmoid')(net)
model = Model(inputs=x, outputs=net, name='skip_conv_single_output')
model = compile_model(model, compile_params)
return model
def _create_model(img_height, img_width, input_timesteps, compile_params):
# Prepair
input_shape = (input_timesteps, img_height, img_width, 1)
# Model architecture
source = keras.Input(name='seed', shape=input_shape, dtype=tf.float32)
model = conv_lstm_2D(filters=128,
kernel_size=3,
strides=1,
padding='same')(source)
model = BatchNormalization()(model)
model = conv_lstm_2D(filters=128,
kernel_size=3,
strides=1,
padding='same')(model)
model = BatchNormalization()(model)
model = conv_lstm_2D(filters=128,
kernel_size=3,
strides=1,
padding='same')(model)
model = BatchNormalization()(model)
model = conv_lstm_2D(filters=128,
kernel_size=3,
strides=1,
padding='same')(model)
model = BatchNormalization()(model)
model = conv_lstm_2D(filters=128,
kernel_size=3,
strides=1,
padding='same',
return_sequences=False)(model)
model = BatchNormalization()(model)
predict_img = conv_2D(filters=1,
kernel_size=3,
strides=1,
padding='same')(model)
model = keras.Model(inputs=[source], outputs=[predict_img])
# Compile model
model = compile_model(model, compile_params)
return model
def trainByCompileParams(type, lr):
print('Training with {0}, lr = {1}'.format(type, lr))
class MyGenerator(Sequence):
def __init__(self, data_filenames, batch_size):
self.data_filenames = data_filenames
self.batch_size = batch_size
def __len__(self):
return len(self.data_filenames)
def __getitem__(self, idx):
data = restore_data('output/{0}'.format(self.data_filenames[idx]))
i = idx
batch_X = np.expand_dims(np.expand_dims(data[:-1, :, :], axis=0),
axis=-1)
batch_Y = np.expand_dims(np.expand_dims(data[-1, :, :], axis=0),
axis=-1)
__max__ = 17.0
__min__ = -34.0
__range__ = __max__ - __min__
X = (batch_X - __min__) / __range__
Y = (batch_Y - __min__) / __range__
return (X, Y)
listFile = os.listdir('output/')
nTrain = int(len(listFile) * 0.60)
nVal = int(len(listFile) * 0.3)
trainFiles = listFile[:nTrain]
valFiles = listFile[nTrain:nTrain + nVal]
testFiles = listFile[nTrain + nVal:]
trainGenerator = MyGenerator(trainFiles, 1)
valGenerator = MyGenerator(valFiles, 1)
input_timesteps = 12
img_height = 128
img_width = 128
# default: adam
opt = optimizers.Adam(lr=lr)
if (type == 'sgd'):
opt = optimizers.SGD(lr=lr)
input_shape = (input_timesteps, img_height, img_width, 1)
compile_params = {'optimizer': opt, 'loss': 'mse', 'metrics': [PSNRLoss]}
# Model architecture
source = keras.Input(name='seed', shape=input_shape, dtype=tf.float32)
model = conv_lstm_2D(filters=64, kernel_size=3, strides=1,
padding='same')(source)
model = BatchNormalization()(model)
model = conv_lstm_2D(filters=64, kernel_size=3, strides=1,
padding='same')(model)
model = BatchNormalization()(model)
model = conv_lstm_2D(filters=64, kernel_size=3, strides=1,
padding='same')(model)
model = BatchNormalization()(model)
model = conv_lstm_2D(filters=64, kernel_size=3, strides=1,
padding='same')(model)
model = BatchNormalization()(model)
model = conv_lstm_2D(filters=64,
kernel_size=3,
strides=1,
padding='same',
return_sequences=False)(model)
model = BatchNormalization()(model)
predict_img = conv_2D(filters=1, kernel_size=3, strides=1,
padding='same')(model)
model = keras.Model(inputs=[source], outputs=[predict_img])
model = compile_model(model, compile_params)
model.fit_generator(generator=trainGenerator,
steps_per_epoch=nTrain,
epochs=60,
validation_data=valGenerator,
validation_steps=2000)
modelFn = 'model_{0}_{1}.h5'.format(type, str(lr).replace('.', ''))
model.save(modelFn)
print('Trained to {0}'.format(modelFn))
del model