def test_estimator(self, hparams_string, batch_size=1):
"""Structural test to make sure Estimator Builder works."""
seed = 42
# Set up and clean test directory.
model_dir = os.path.join(flags.FLAGS.test_tmpdir,
"AdanetImproveNasBuilderTest")
if tf.gfile.Exists(model_dir):
tf.gfile.DeleteRecursively(model_dir)
tf.gfile.MkDir(model_dir)
data_provider = fake_data.FakeImageProvider(seed=seed)
estimator_builder = adanet_improve_nas.Builder()
hparams = estimator_builder.hparams(default_batch_size=3,
hparams_string=hparams_string)
run_config = tf.estimator.RunConfig(tf_random_seed=seed,
model_dir=model_dir)
_ = data_provider.get_input_fn("train",
tf.estimator.ModeKeys.TRAIN,
batch_size=batch_size)
test_input_fn = data_provider.get_input_fn("test",
tf.estimator.ModeKeys.EVAL,
batch_size=batch_size)
estimator = estimator_builder.estimator(data_provider=data_provider,
run_config=run_config,
hparams=hparams,
train_steps=10,
seed=seed)
eval_metrics = estimator.evaluate(input_fn=test_input_fn, steps=1)
self.assertGreater(eval_metrics["loss"], 0.0)
def main(argv):
del argv
run_config = make_run_config()
estimator_builder = adanet_improve_nas.Builder()
hparams = estimator_builder.hparams(FLAGS.batch_size, FLAGS.hparams)
tf.logging.info("Running Experiment with HParams: %s", hparams)
if FLAGS.dataset == "cifar10":
data_provider = cifar10.Provider()
elif FLAGS.dataset == "cifar100":
data_provider = cifar100.Provider()
elif FLAGS.dataset == "fake":
data_provider = fake_data.FakeImageProvider(
num_examples=10,
num_classes=10,
image_dim=32,
channels=3,
seed=42)
else:
raise ValueError("Invalid dataset")
estimator = estimator_builder.estimator(
data_provider=data_provider,
run_config=run_config,
hparams=hparams,
train_steps=FLAGS.train_steps)
train_spec = tf.estimator.TrainSpec(
input_fn=data_provider.get_input_fn(
partition="train",
mode=tf.estimator.ModeKeys.TRAIN,
batch_size=FLAGS.batch_size),
max_steps=FLAGS.train_steps
)
eval_spec = tf.estimator.EvalSpec(
input_fn=data_provider.get_input_fn(
partition="test",
mode=tf.estimator.ModeKeys.EVAL,
batch_size=FLAGS.batch_size),
steps=FLAGS.eval_steps,
start_delay_secs=10,
throttle_secs=1800
)
tf.logging.info("Training!")
tf.estimator.train_and_evaluate(estimator, train_spec, eval_spec)
tf.logging.info("Done training!")