tf.set_random_seed(general_params.seed)
# get training data pipeline. TODO: get also validation data
training_data = Cifar10DataFetcher('TRAIN',
batch_size=hparams.batch_size,
order=image_params.order)
# create hypernet
hnet = Hypernetwork(training_data.image,
training_data.label,
'TRAIN',
hnet_hparams=hparams,
image_params=image_params,
target_hparams=target_hparams)
# protobuf computational graph
with open(os.path.join(output_dir, 'graph.pb'), 'w') as f:
f.write(str(hnet.graph.as_graph_def()))
# summary
writer = tf.summary.FileWriter(output_dir, hnet.graph)
# train
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True,
log_device_placement=True)) as sess:
hnet.Train(sess,
max_steps=1e6,
logger=logger,
writer=writer,
checkpoint_file_name=output_dir)
config.gpu_options.per_process_gpu_memory_fraction = 0.5
with tf.Session(config=config).as_default() as sess:
# get training data pipeline.
training_data = Cifar10DataFetcher('TRAIN', batch_size=hparams.batch_size)
validation_data = Cifar10DataFetcher('VALIDATION',
batch_size=hparams.batch_size)
# create hypernet
hnet = Hypernetwork(training_data.image,
training_data.label,
'TRAIN',
hnet_hparams=hparams)
# protobuf computational graph
with open(os.path.join(output_dir, 'graph.pb'), 'w') as f:
f.write(str(hnet.graph.as_graph_def()))
# summary
writer = tf.summary.FileWriter(output_dir, hnet.graph)
# train
hnet.Train(sess,
validation_data.image,
validation_data.label,
max_steps=1e6,
logger=logger,
writer=writer,
checkpoint_file_name=output_dir,
log_interval=100)