def main(argv):
del argv # unused arg
if FLAGS.num_cores > 1 or not FLAGS.use_gpu:
raise ValueError('Only single GPU is currently supported.')
tf.enable_v2_behavior()
tf.io.gfile.makedirs(FLAGS.output_dir)
tf.random.set_seed(FLAGS.seed)
dataset_train, ds_info = utils.load_dataset(tfds.Split.TRAIN, with_info=True)
dataset_size = ds_info.splits['train'].num_examples
dataset_train = dataset_train.repeat().shuffle(10 * FLAGS.batch_size).batch(
FLAGS.batch_size)
test_batch_size = 100
validation_steps = ds_info.splits['test'].num_examples // test_batch_size
dataset_test = utils.load_dataset(tfds.Split.TEST)
dataset_test = dataset_test.repeat().batch(test_batch_size)
model = resnet_v1(input_shape=ds_info.features['image'].shape,
depth=20,
num_classes=ds_info.features['label'].num_classes,
batch_norm=FLAGS.batch_norm,
prior_stddev=FLAGS.prior_stddev,
dataset_size=dataset_size)
kl, elbo = get_metrics(model, dataset_size)
model.compile(
tf.keras.optimizers.Adam(FLAGS.init_learning_rate),
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=[
tf.keras.metrics.SparseCategoricalCrossentropy(
name='negative_log_likelihood',
from_logits=True),
tf.keras.metrics.SparseCategoricalAccuracy(name='accuracy'),
elbo,
kl])
logging.info('Model input shape: %s', model.input_shape)
logging.info('Model output shape: %s', model.output_shape)
logging.info('Model number of weights: %s', model.count_params())
tensorboard_cb = tf.keras.callbacks.TensorBoard(log_dir=FLAGS.output_dir,
write_graph=False)
lr_scheduler = utils.make_lr_scheduler(FLAGS.init_learning_rate)
model.fit(dataset_train,
steps_per_epoch=dataset_size // FLAGS.batch_size,
epochs=FLAGS.train_epochs,
validation_data=dataset_test,
validation_steps=validation_steps,
callbacks=[tensorboard_cb, lr_scheduler])
logging.info('Saving model to output_dir.')
model_filename = FLAGS.output_dir + '/model.ckpt'
model.save_weights(model_filename)
def main(argv):
del argv # unused arg
tf.io.gfile.makedirs(FLAGS.output_dir)
tf.random.set_seed(FLAGS.seed)
dataset_train, ds_info = utils.load_dataset(tfds.Split.TRAIN,
with_info=True)
dataset_size = ds_info.splits['train'].num_examples
dataset_train = dataset_train.repeat().shuffle(
10 * FLAGS.batch_size).batch(FLAGS.batch_size)
test_batch_size = 100
validation_steps = ds_info.splits['test'].num_examples // test_batch_size
dataset_test = utils.load_dataset(tfds.Split.TEST)
dataset_test = dataset_test.repeat().batch(test_batch_size)
model = resnet_v1(input_shape=ds_info.features['image'].shape,
depth=20,
num_classes=ds_info.features['label'].num_classes,
batch_norm=FLAGS.batch_norm,
prior_stddev=FLAGS.prior_stddev,
dataset_size=dataset_size)
negative_log_likelihood, accuracy, log_marginal, kl, elbo = get_metrics(
model, dataset_size)
model.compile(tf.keras.optimizers.Adam(FLAGS.init_learning_rate),
loss=negative_log_likelihood,
metrics=[elbo, log_marginal, kl, accuracy])
logging.info('Model input shape: %s', model.input_shape)
logging.info('Model output shape: %s', model.output_shape)
logging.info('Model number of weights: %s', model.count_params())
tensorboard_cb = tf.keras.callbacks.TensorBoard(log_dir=FLAGS.output_dir,
write_graph=False)
lr_scheduler = utils.make_lr_scheduler(FLAGS.init_learning_rate)
model.fit(dataset_train,
steps_per_epoch=dataset_size // FLAGS.batch_size,
epochs=FLAGS.train_epochs,
validation_data=dataset_test,
validation_steps=validation_steps,
callbacks=[tensorboard_cb, lr_scheduler])
logging.info('Saving model to output_dir.')
model_filename = FLAGS.output_dir + '/model.ckpt'
model.save_weights(model_filename)
cfg = options.get_arguments()
EXPERIMENT = f"{cfg.model}_{cfg.experiment}"
MODEL_PATH = f"models/{EXPERIMENT}"
LOG_PATH = f"logs/{EXPERIMENT}"
utils.make_folder(MODEL_PATH)
utils.make_folder(LOG_PATH)
criterions = utils.define_losses()
dataloaders = utils.make_data_novel(cfg)
model = utils.build_structure_generator(cfg).to(cfg.device)
optimizer = utils.make_optimizer(cfg, model)
scheduler = utils.make_lr_scheduler(cfg, optimizer)
logger = utils.make_logger(LOG_PATH)
writer = utils.make_summary_writer(EXPERIMENT)
def on_after_epoch(model, df_hist, images, epoch, saveEpoch):
utils.save_best_model(MODEL_PATH, model, df_hist)
utils.checkpoint_model(MODEL_PATH, model, epoch, saveEpoch)
utils.log_hist(logger, df_hist)
utils.write_on_board_losses_stg2(writer, df_hist)
utils.write_on_board_images_stg2(writer, images, epoch)
if cfg.lrSched is not None:
def on_after_batch(iteration):
utils.write_on_board_lr(writer, scheduler.get_lr(), iteration)
scheduler.step(iteration)