def setUp(self):
super(TrainExperimentTest, self).setUp()
self.hparams = train_experiment.HParams(
# Make the test on CPU.
use_tpu=False,
eval_on_tpu=False,
# Make small batch sizes.
z_dim=4,
train_batch_size=4,
eval_batch_size=16,
predict_batch_size=1,
fake_data=True,
fake_nets=True,
# Make small networks.
gf_dim=2,
df_dim=4,
# Take few steps.
max_number_of_steps=1,
num_eval_steps=1,
continuous_eval_timeout_secs=1,
tpu_iterations_per_loop=1,
model_dir=self.create_tempdir().full_path,
use_tpu_estimator=False,
train_steps_per_eval=1,
generator_lr=1.0,
discriminator_lr=1.0,
beta1=1.0,
shuffle_buffer_size=1,
num_classes=10,
)
def main(_):
hparams = train_experiment.HParams(
train_batch_size=FLAGS.train_batch_size,
eval_batch_size=FLAGS.eval_batch_size,
predict_batch_size=FLAGS.predict_batch_size,
use_tpu=FLAGS.use_tpu,
eval_on_tpu=FLAGS.eval_on_tpu,
generator_lr=FLAGS.generator_lr,
discriminator_lr=FLAGS.discriminator_lr,
beta1=FLAGS.beta1,
gf_dim=FLAGS.gf_dim,
df_dim=FLAGS.df_dim,
num_classes=1000,
shuffle_buffer_size=10000,
z_dim=FLAGS.z_dim,
model_dir=FLAGS.model_dir,
continuous_eval_timeout_secs=FLAGS.continuous_eval_timeout_secs,
use_tpu_estimator=FLAGS.use_tpu_estimator,
max_number_of_steps=FLAGS.max_number_of_steps,
train_steps_per_eval=FLAGS.train_steps_per_eval,
num_eval_steps=FLAGS.num_eval_steps,
fake_nets=FLAGS.fake_nets,
tpu_iterations_per_loop=FLAGS.tpu_iterations_per_loop,
)
if FLAGS.mode == 'train':
train_experiment.run_train(hparams)
elif FLAGS.mode == 'continuous_eval':
train_experiment.run_continuous_eval(hparams)
elif FLAGS.mode == 'train_and_eval' or FLAGS.mode is None:
train_experiment.run_train_and_eval(hparams)
else:
raise ValueError('Mode not recognized: ', FLAGS.mode)
def setUp(self):
super(EstimatorLibTest, self).setUp()
self.hparams = train_experiment.HParams(
train_batch_size=1,
eval_batch_size=32,
predict_batch_size=1,
generator_lr=1.0,
discriminator_lr=1.0,
beta1=1.0,
gf_dim=2,
df_dim=2,
num_classes=10,
shuffle_buffer_size=1,
z_dim=8,
model_dir=None,
max_number_of_steps=None,
train_steps_per_eval=1,
num_eval_steps=1,
debug_params=train_experiment.DebugParams(
use_tpu=False,
eval_on_tpu=False,
fake_nets=None,
fake_data=None,
continuous_eval_timeout_secs=1,
),
tpu_params=train_experiment.TPUParams(
use_tpu_estimator=False,
tpu_location='local',
gcp_project=None,
tpu_zone=None,
tpu_iterations_per_loop=1,
),
)
def test_get_metrics_syntax(self, mock_fid, mock_iscore, use_tpu):
if tf.executing_eagerly():
# tf.metrics.mean is not supported when eager execution is enabled.
return
bs = 40
hparams = train_experiment.HParams(
train_batch_size=1,
eval_batch_size=bs,
predict_batch_size=1,
generator_lr=1.0,
discriminator_lr=1.0,
beta1=1.0,
gf_dim=2,
df_dim=2,
num_classes=10,
shuffle_buffer_size=1,
z_dim=8,
model_dir=None,
max_number_of_steps=None,
train_steps_per_eval=1,
num_eval_steps=1,
debug_params=train_experiment.DebugParams(
use_tpu=use_tpu,
eval_on_tpu=use_tpu,
fake_nets=True,
fake_data=True,
continuous_eval_timeout_secs=1,
),
tpu_params=None,
)
# Fake arguments to pass to `get_metrics`.
fake_noise = tf.zeros([bs, 128])
fake_imgs = tf.zeros([bs, 128, 128, 3])
fake_lbls = tf.zeros([bs])
fake_logits = tf.ones([bs, 1008])
# Mock Inception-inference computations.
mock_iscore.return_value = 1.0
mock_fid.return_value = 0.0
estimator_lib.get_metrics(generator_inputs=fake_noise,
generated_data={
'images': fake_imgs,
'labels': fake_lbls
},
real_data={
'images': fake_imgs,
'labels': fake_lbls
},
discriminator_real_outputs=(fake_logits, ()),
discriminator_gen_outputs=(fake_logits, ()),
hparams=hparams)
def main(_):
from tensorflow_gan.examples.self_attention_estimator import train_experiment
# get TF logger
log = logging.getLogger('tensorflow')
log.setLevel(logging.INFO)
# create formatter and add it to the handlers
formatter = logging.Formatter(
'%(asctime)s - %(name)s - %(levelname)s - %(message)s')
# create file handler
logging_dir = os.environ['HOME'] if FLAGS.use_tpu else FLAGS.model_dir
if not os.path.isdir(logging_dir):
os.makedirs(logging_dir)
fh = logging.FileHandler(logging_dir + '/tensorflow.log')
fh.setLevel(logging.INFO)
fh.setFormatter(formatter)
log.addHandler(fh)
tpu_location = FLAGS.tpu
if FLAGS.use_tpu:
assert ',' not in tpu_location, 'Only using 1 TPU is supported'
hparams = train_experiment.HParams(
train_batch_size=FLAGS.train_batch_size,
eval_batch_size=FLAGS.eval_batch_size,
predict_batch_size=FLAGS.predict_batch_size,
generator_lr=FLAGS.generator_lr,
discriminator_lr=FLAGS.discriminator_lr,
beta1=FLAGS.beta1,
gf_dim=FLAGS.gf_dim,
df_dim=FLAGS.df_dim,
num_classes=FLAGS.num_classes,
shuffle_buffer_size=10000,
z_dim=FLAGS.z_dim,
model_dir=FLAGS.model_dir,
max_number_of_steps=FLAGS.max_number_of_steps,
train_steps_per_eval=FLAGS.train_steps_per_eval,
num_eval_steps=FLAGS.num_eval_steps,
debug_params=train_experiment.DebugParams(
use_tpu=FLAGS.use_tpu,
eval_on_tpu=FLAGS.eval_on_tpu,
fake_nets=False,
fake_data=False,
continuous_eval_timeout_secs=FLAGS.continuous_eval_timeout_secs,
),
tpu_params=train_experiment.TPUParams(
use_tpu_estimator=FLAGS.use_tpu_estimator,
tpu_location=tpu_location,
gcp_project=FLAGS.gcp_project,
tpu_zone=FLAGS.tpu_zone,
tpu_iterations_per_loop=FLAGS.tpu_iterations_per_loop,
),
)
if FLAGS.mode == 'train':
train_experiment.run_train(hparams)
elif FLAGS.mode == 'continuous_eval':
train_experiment.run_continuous_eval(hparams)
elif FLAGS.mode == 'intra_fid_eval':
train_experiment.run_intra_fid_eval(hparams)
elif FLAGS.mode == 'train_and_eval' or FLAGS.mode is None:
train_experiment.run_train_and_eval(hparams)
elif FLAGS.mode == 'gen_images':
train_experiment.gen_images(hparams)
elif FLAGS.mode == 'gen_matrices':
train_experiment.gen_matrices(hparams)
else:
raise ValueError('Mode not recognized: ', FLAGS.mode)
