def main(args):
if args.distribute:
distribute.enable_distributed_training()
tf.logging.set_verbosity(tf.logging.INFO)
model_cls = models.get_model(args.model)
params = default_parameters()
# Import and override parameters
# Priorities (low -> high):
# default -> saved -> command
params = merge_parameters(params, model_cls.get_parameters())
params = import_params(args.output, args.model, params)
override_parameters(params, args)
# Export all parameters and model specific parameters
if distribute.rank() == 0:
export_params(params.output, "params.json", params)
export_params(params.output, "%s.json" % args.model,
collect_params(params, model_cls.get_parameters()))
# Build Graph
with tf.Graph().as_default():
if not params.record:
# Build input queue
features = dataset.get_training_input(params.input, params)
else:
features = record.get_input_features(
os.path.join(params.record, "*train*"), "train", params)
# Build model
initializer = get_initializer(params)
regularizer = tf.contrib.layers.l1_l2_regularizer(
scale_l1=params.scale_l1, scale_l2=params.scale_l2)
model = model_cls(params)
# Create global step
global_step = tf.train.get_or_create_global_step()
dtype = tf.float16 if args.half else None
# Multi-GPU setting
sharded_losses = parallel.parallel_model(
model.get_training_func(initializer, regularizer, dtype), features,
params.device_list)
loss = tf.add_n(sharded_losses) / len(sharded_losses)
loss = loss + tf.losses.get_regularization_loss()
if distribute.rank() == 0:
print_variables()
learning_rate = get_learning_rate_decay(params.learning_rate,
global_step, params)
learning_rate = tf.convert_to_tensor(learning_rate, dtype=tf.float32)
tf.summary.scalar("loss", loss)
tf.summary.scalar("learning_rate", learning_rate)
# Create optimizer
if params.optimizer == "Adam":
opt = tf.train.AdamOptimizer(learning_rate,
beta1=params.adam_beta1,
beta2=params.adam_beta2,
epsilon=params.adam_epsilon)
elif params.optimizer == "LazyAdam":
opt = tf.contrib.opt.LazyAdamOptimizer(learning_rate,
beta1=params.adam_beta1,
beta2=params.adam_beta2,
epsilon=params.adam_epsilon)
else:
raise RuntimeError("Optimizer %s not supported" % params.optimizer)
opt = optimizers.MultiStepOptimizer(opt, params.update_cycle)
if args.half:
opt = optimizers.LossScalingOptimizer(opt, params.loss_scale)
# Optimization
grads_and_vars = opt.compute_gradients(
loss, colocate_gradients_with_ops=True)
if params.clip_grad_norm:
grads, var_list = list(zip(*grads_and_vars))
grads, _ = tf.clip_by_global_norm(grads, params.clip_grad_norm)
grads_and_vars = zip(grads, var_list)
train_op = opt.apply_gradients(grads_and_vars, global_step=global_step)
# Validation
if params.validation and params.references[0]:
files = [params.validation] + list(params.references)
eval_inputs = dataset.sort_and_zip_files(files)
eval_input_fn = dataset.get_evaluation_input
else:
eval_input_fn = None
# Hooks
train_hooks = [
tf.train.StopAtStepHook(last_step=params.train_steps),
tf.train.NanTensorHook(loss),
tf.train.LoggingTensorHook(
{
"step": global_step,
"loss": loss,
"source": tf.shape(features["source"]),
"target": tf.shape(features["target"])
},
every_n_iter=1)
]
broadcast_hook = distribute.get_broadcast_hook()
if broadcast_hook:
train_hooks.append(broadcast_hook)
if distribute.rank() == 0:
# Add hooks
save_vars = tf.trainable_variables() + [global_step]
saver = tf.train.Saver(
var_list=save_vars if params.only_save_trainable else None,
max_to_keep=params.keep_checkpoint_max,
sharded=False)
tf.add_to_collection(tf.GraphKeys.SAVERS, saver)
train_hooks.append(
hooks.MultiStepHook(tf.train.CheckpointSaverHook(
checkpoint_dir=params.output,
save_secs=params.save_checkpoint_secs or None,
save_steps=params.save_checkpoint_steps or None,
saver=saver),
step=params.update_cycle))
if eval_input_fn is not None:
train_hooks.append(
hooks.MultiStepHook(hooks.EvaluationHook(
lambda f: inference.create_inference_graph([model], f,
params),
lambda: eval_input_fn(eval_inputs, params),
lambda x: decode_target_ids(x, params),
params.output,
session_config(params),
device_list=params.device_list,
max_to_keep=params.keep_top_checkpoint_max,
eval_secs=params.eval_secs,
eval_steps=params.eval_steps),
step=params.update_cycle))
checkpoint_dir = params.output
else:
checkpoint_dir = None
restore_op = restore_variables(args.checkpoint)
def restore_fn(step_context):
step_context.session.run(restore_op)
# Create session, do not use default CheckpointSaverHook
with tf.train.MonitoredTrainingSession(
checkpoint_dir=checkpoint_dir,
hooks=train_hooks,
save_checkpoint_secs=None,
config=session_config(params)) as sess:
# Restore pre-trained variables
sess.run_step_fn(restore_fn)
while not sess.should_stop():
sess.run(train_op)
def main(args):
if args.distribute:
distribute.enable_distributed_training()
tf.logging.set_verbosity(tf.logging.INFO)
model_cls = models.get_model(args.model)
params = default_parameters()
# Import and override parameters
# Priorities (low -> high):
# default -> saved -> command
params = merge_parameters(params, model_cls.get_parameters())
params = import_params(args.output, args.model, params)
override_parameters(params, args)
# Export all parameters and model specific parameters
if not args.distribute or distribute.rank() == 0:
export_params(params.output, "params.json", params)
export_params(params.output, "%s.json" % args.model,
collect_params(params, model_cls.get_parameters()))
assert 'r2l' in params.input[2]
# Build Graph
use_all_devices(params)
with tf.Graph().as_default():
if not params.record:
# Build input queue
features = dataset.abd_get_training_input(params.input, params)
else:
features = record.get_input_features(
os.path.join(params.record, "*train*"), "train", params)
update_cycle = params.update_cycle
features, init_op = cache.cache_features(features, update_cycle)
# Build model
initializer = get_initializer(params)
regularizer = tf.contrib.layers.l1_l2_regularizer(
scale_l1=params.scale_l1, scale_l2=params.scale_l2)
model = model_cls(params)
# Create global step
global_step = tf.train.get_or_create_global_step()
dtype = tf.float16 if args.fp16 else None
if args.distribute:
training_func = model.get_training_func(initializer, regularizer,
dtype)
loss = training_func(features)
else:
# Multi-GPU setting
sharded_losses = parallel.parallel_model(
model.get_training_func(initializer, regularizer, dtype),
features, params.device_list)
loss = tf.add_n(sharded_losses) / len(sharded_losses)
loss = loss + tf.losses.get_regularization_loss()
# Print parameters
if not args.distribute or distribute.rank() == 0:
print_variables()
learning_rate = get_learning_rate_decay(params.learning_rate,
global_step, params)
learning_rate = tf.convert_to_tensor(learning_rate, dtype=tf.float32)
tf.summary.scalar("learning_rate", learning_rate)
# Create optimizer
if params.optimizer == "Adam":
opt = tf.train.AdamOptimizer(learning_rate,
beta1=params.adam_beta1,
beta2=params.adam_beta2,
epsilon=params.adam_epsilon)
elif params.optimizer == "LazyAdam":
opt = tf.contrib.opt.LazyAdamOptimizer(learning_rate,
beta1=params.adam_beta1,
beta2=params.adam_beta2,
epsilon=params.adam_epsilon)
else:
raise RuntimeError("Optimizer %s not supported" % params.optimizer)
loss, ops = optimize.create_train_op(
loss, opt, global_step,
distribute.all_reduce if args.distribute else None, args.fp16,
params)
restore_op = restore_variables(args.checkpoint)
# Validation
if params.validation and params.references[0]:
files = params.validation + list(params.references)
eval_inputs = dataset.sort_and_zip_files(files)
eval_input_fn = dataset.abd_get_evaluation_input
else:
eval_input_fn = None
# Add hooks
multiplier = tf.convert_to_tensor([update_cycle, 1])
train_hooks = [
tf.train.StopAtStepHook(last_step=params.train_steps),
tf.train.NanTensorHook(loss),
tf.train.LoggingTensorHook(
{
"step": global_step,
"loss": loss,
"source": tf.shape(features["source"]) * multiplier,
"target": tf.shape(features["target"]) * multiplier
},
every_n_iter=1)
]
if args.distribute:
train_hooks.append(distribute.get_broadcast_hook())
config = session_config(params)
if not args.distribute or distribute.rank() == 0:
# Add hooks
save_vars = tf.trainable_variables() + [global_step]
saver = tf.train.Saver(
var_list=save_vars if params.only_save_trainable else None,
max_to_keep=params.keep_checkpoint_max,
sharded=False)
tf.add_to_collection(tf.GraphKeys.SAVERS, saver)
train_hooks.append(
tf.train.CheckpointSaverHook(
checkpoint_dir=params.output,
save_secs=params.save_checkpoint_secs or None,
save_steps=params.save_checkpoint_steps or None,
saver=saver))
if eval_input_fn is not None:
if not args.distribute or distribute.rank() == 0:
train_hooks.append(
hooks.EvaluationHook(
lambda f: inference.create_inference_graph([model], f,
params),
lambda: eval_input_fn(eval_inputs, params),
lambda x: decode_target_ids(x, params),
params.output,
config,
params.keep_top_checkpoint_max,
eval_secs=params.eval_secs,
eval_steps=params.eval_steps))
def restore_fn(step_context):
step_context.session.run(restore_op)
def step_fn(step_context):
# Bypass hook calls
step_context.session.run([init_op, ops["zero_op"]])
for i in range(update_cycle - 1):
step_context.session.run(ops["collect_op"])
return step_context.run_with_hooks(ops["train_op"])
# Create session, do not use default CheckpointSaverHook
if not args.distribute or distribute.rank() == 0:
checkpoint_dir = params.output
else:
checkpoint_dir = None
with tf.train.MonitoredTrainingSession(checkpoint_dir=checkpoint_dir,
hooks=train_hooks,
save_checkpoint_secs=None,
config=config) as sess:
# Restore pre-trained variables
sess.run_step_fn(restore_fn)
while not sess.should_stop():
sess.run_step_fn(step_fn)