def main(argv):
if len(argv) > 1:
raise app.UsageError('Too many command-line arguments.')
cfg = hyperparameters.get_config()
train_dataset, data_shapes = datasets.get_sequence_dataset(
data_dir=os.path.join(cfg.data_dir, cfg.train_dir),
batch_size=cfg.batch_size,
num_timesteps=cfg.observed_steps + cfg.predicted_steps)
test_dataset, _ = datasets.get_sequence_dataset(
data_dir=os.path.join(cfg.data_dir, cfg.test_dir),
batch_size=cfg.batch_size,
num_timesteps=cfg.observed_steps + cfg.predicted_steps)
model = build_model(cfg, data_shapes)
optimizer = tf.keras.optimizers.Adam(lr=cfg.learning_rate,
clipnorm=cfg.clipnorm)
model.compile(optimizer)
model.fit(x=train_dataset,
steps_per_epoch=cfg.steps_per_epoch,
epochs=cfg.num_epochs,
validation_data=test_dataset,
validation_steps=1)
def get_dataset(self, batch_size=None, random_offset=True, seed=0):
return datasets.get_sequence_dataset(data_dir=self.data_dir,
file_glob=self.file_glob,
batch_size=batch_size
or self.batch_size,
num_timesteps=self.num_timesteps,
random_offset=random_offset,
seed=seed)
def testAutoencoderTrainingLossGoesDown(self):
"""Tests a minimal Keras training loop for the non-dynamic model parts."""
dataset, data_shapes = datasets.get_sequence_dataset(
data_dir=self.cfg.train_dir,
file_glob='acrobot*',
batch_size=self.cfg.batch_size,
num_timesteps=self.cfg.observed_steps + self.cfg.predicted_steps,
random_offset=True)
autoencoder = Autoencoder(self.cfg, data_shapes)
optimizer = tf.keras.optimizers.Adam(lr=1e-4)
autoencoder.compile(optimizer)
history = autoencoder.fit(dataset, steps_per_epoch=1, epochs=3)
self.assertLess(history.history['loss'][-1], history.history['loss'][0])
