def main(unused_argv):
logging.info('Loading training data')
with utils.read_h5py(FLAGS.input_path) as f:
snapshots = f['v'][...]
equation_kwargs = {k: v.item() for k, v in f.attrs.items()}
logging.info('Inputs have shape %r', snapshots.shape)
if FLAGS.checkpoint_dir:
tf.gfile.MakeDirs(FLAGS.checkpoint_dir)
hparams = training.create_hparams(
FLAGS.equation, equation_kwargs=json.dumps(equation_kwargs))
hparams.parse(FLAGS.hparams)
logging.info('Starting training loop')
metrics_df = training.training_loop(snapshots,
FLAGS.checkpoint_dir,
hparams,
master=FLAGS.master)
if FLAGS.checkpoint_dir:
logging.info('Saving CSV with metrics')
csv_path = os.path.join(FLAGS.checkpoint_dir, 'metrics.csv')
with tf.gfile.GFile(csv_path, 'w') as f:
metrics_df.to_csv(f, index=False)
logging.info('Finished')
def test_training_loop(self, **hparam_values):
with tf.Graph().as_default():
snapshots = np.random.RandomState(0).randn(100, NUM_X_POINTS)
hparams = training.create_hparams(
learning_rates=[1e-3],
learning_stops=[20],
eval_interval=10,
equation_kwargs=json.dumps({'num_points': NUM_X_POINTS}),
**hparam_values)
results = training.training_loop(snapshots, self.tmpdir, hparams)
self.assertIsInstance(results, pd.DataFrame)
self.assertEqual(results.shape[0], 2)
def train(self, hparams):
# train a model on random noise
with tf.Graph().as_default():
snapshots = 0.01 * np.random.RandomState(0).randn(
100, NUM_X_POINTS)
training.training_loop(snapshots, self.checkpoint_dir, hparams)