def main(_):
model_helpers.apply_clean(flags.FLAGS)
logdir = '/tmp/logs'
if not os.path.exists(logdir):
os.makedirs(logdir)
logname = 'imagenet_strategy_{}_model_{}_node_{}_gpu_{}_patch_{}_proxy_{}'.format(
flags.FLAGS.autodist_strategy, flags.FLAGS.cnn_model, node_num,
gpu_num, flags.FLAGS.autodist_patch_tf, flags.FLAGS.proxy)
logging.get_absl_handler().use_absl_log_file(logname, logdir)
with logger.benchmark_context(flags.FLAGS):
run(flags.FLAGS)
def run_mnist(flags_obj):
"""Run MNIST training and eval loop.
Args:
flags_obj: An object containing parsed flag values.
"""
model_helpers.apply_clean(flags_obj)
model_function = model_fn
session_config = tf.ConfigProto(
inter_op_parallelism_threads=flags_obj.inter_op_parallelism_threads,
intra_op_parallelism_threads=flags_obj.intra_op_parallelism_threads,
allow_soft_placement=True)
distribution_strategy = distribution_utils.get_distribution_strategy(
flags_core.get_num_gpus(flags_obj), flags_obj.all_reduce_alg)
run_config = tf.estimator.RunConfig(
train_distribute=distribution_strategy,
session_config=session_config,
save_checkpoints_steps=flags_obj.ckpt_steps,
keep_checkpoint_max=flags_obj.max_ckpts,
save_summary_steps=flags_obj.save_summary_steps,
log_step_count_steps=flags_obj.log_step_count_steps
)
data_format = flags_obj.data_format
if data_format is None:
data_format = ('channels_first'
if tf.test.is_built_with_cuda() else 'channels_last')
mnist_classifier = tf.estimator.Estimator(
model_fn=model_function,
model_dir=flags_obj.model_dir,
config=run_config,
params={
'data_format': data_format,
})
# Set up training and evaluation input functions.
def train_input_fn():
"""Prepare data for training."""
# When choosing shuffle buffer sizes, larger sizes result in better
# randomness, while smaller sizes use less memory. MNIST is a small
# enough dataset that we can easily shuffle the full epoch.
ds = dataset.train(flags_obj.data_dir)
ds = ds.cache().shuffle(buffer_size=50000).batch(flags_obj.batch_size)
# Iterate through the dataset a set number (`epochs_between_evals`) of times
# during each training session.
ds = ds.repeat(flags_obj.epochs_between_evals)
return ds
def eval_input_fn():
return dataset.test(flags_obj.data_dir).batch(
flags_obj.batch_size).make_one_shot_iterator().get_next()
# Set up hook that outputs training logs every 100 steps.
train_hooks = hooks_helper.get_train_hooks(
flags_obj.hooks, model_dir=flags_obj.model_dir,
batch_size=flags_obj.batch_size)
train_spec = tf.estimator.TrainSpec(input_fn=train_input_fn, hooks=train_hooks, max_steps=flags_obj.max_steps)
eval_spec = tf.estimator.EvalSpec(input_fn=eval_input_fn, steps=None,
start_delay_secs=10,
throttle_secs=flags_obj.eval_secs)
tf.estimator.train_and_evaluate(mnist_classifier, train_spec, eval_spec)
# Export the model if node is master and export_dir is set and if experiment is multinode - check if its master
if os.environ.get('PS_CONFIG') and os.environ.get('TYPE') != 'master':
tf.logging.debug('No model was exported')
return
if flags_obj.export_dir:
tf.logging.debug('Starting to Export model to {}'.format(str(flags_obj.export_dir)))
image = tf.placeholder(tf.float32, [None, 28, 28])
input_fn = tf.estimator.export.build_raw_serving_input_receiver_fn({
'image': image,
})
mnist_classifier.export_savedmodel(flags_obj.export_dir, input_fn,
strip_default_attrs=True)
tf.logging.debug('Model Exported')
def run_transformer(flags_obj):
"""Create tf.Estimator to train and evaluate transformer model.
Args:
flags_obj: Object containing parsed flag values.
"""
num_gpus = flags_core.get_num_gpus(flags_obj)
# Add flag-defined parameters to params object
params = PARAMS_MAP[flags_obj.param_set]
if num_gpus > 1:
if flags_obj.param_set == "big":
params = model_params.BIG_MULTI_GPU_PARAMS
elif flags_obj.param_set == "base":
params = model_params.BASE_MULTI_GPU_PARAMS
params["data_dir"] = flags_obj.data_dir
params["model_dir"] = flags_obj.model_dir
params["num_parallel_calls"] = flags_obj.num_parallel_calls
params["tpu"] = flags_obj.tpu
params["use_tpu"] = bool(flags_obj.tpu) # was a tpu specified.
params["static_batch"] = flags_obj.static_batch or params["use_tpu"]
params["allow_ffn_pad"] = not params["use_tpu"]
params["use_synthetic_data"] = flags_obj.use_synthetic_data
# Set batch size parameter, which depends on the availability of
# TPU and GPU, and distribution settings.
params["batch_size"] = (flags_obj.batch_size or (
params["default_batch_size_tpu"] if params["use_tpu"]
else params["default_batch_size"]))
if not params["use_tpu"]:
params["batch_size"] = distribution_utils.per_device_batch_size(
params["batch_size"], num_gpus)
schedule_manager = schedule.Manager(
train_steps=flags_obj.train_steps,
steps_between_evals=flags_obj.steps_between_evals,
train_epochs=flags_obj.train_epochs,
epochs_between_evals=flags_obj.epochs_between_evals,
default_train_epochs=DEFAULT_TRAIN_EPOCHS,
batch_size=params["batch_size"],
max_length=params["max_length"],
use_tpu=params["use_tpu"],
num_tpu_shards=flags_obj.num_tpu_shards
)
params["repeat_dataset"] = schedule_manager.repeat_dataset
model_helpers.apply_clean(flags.FLAGS)
# Create hooks that log information about the training and metric values
train_hooks = hooks_helper.get_train_hooks(
flags_obj.hooks,
model_dir=flags_obj.model_dir,
tensors_to_log=TENSORS_TO_LOG, # used for logging hooks
batch_size=schedule_manager.batch_size, # for ExamplesPerSecondHook
use_tpu=params["use_tpu"] # Not all hooks can run with TPUs
)
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info(
model_name="transformer",
dataset_name="wmt_translate_ende",
run_params=params,
test_id=flags_obj.benchmark_test_id)
# Train and evaluate transformer model
estimator = construct_estimator(flags_obj, params, schedule_manager)
run_loop(
estimator=estimator,
# Training arguments
schedule_manager=schedule_manager,
train_hooks=train_hooks,
benchmark_logger=benchmark_logger,
# BLEU calculation arguments
bleu_source=flags_obj.bleu_source,
bleu_ref=flags_obj.bleu_ref,
bleu_threshold=flags_obj.stop_threshold,
vocab_file=flags_obj.vocab_file)
if flags_obj.export_dir and not params["use_tpu"]:
serving_input_fn = export.build_tensor_serving_input_receiver_fn(
shape=[None], dtype=tf.int64, batch_size=None)
# Export saved model, and save the vocab file as an extra asset. The vocab
# file is saved to allow consistent input encoding and output decoding.
# (See the "Export trained model" section in the README for an example of
# how to use the vocab file.)
# Since the model itself does not use the vocab file, this file is saved as
# an extra asset rather than a core asset.
estimator.export_savedmodel(
flags_obj.export_dir, serving_input_fn,
assets_extra={"vocab.txt": flags_obj.vocab_file},
strip_default_attrs=True)
def run_mnist(flags_obj):
"""Run MNIST training and eval loop.
Args:
flags_obj: An object containing parsed flag values.
"""
model_helpers.apply_clean(flags_obj)
model_function = model_fn
session_config = tf.ConfigProto(
inter_op_parallelism_threads=flags_obj.inter_op_parallelism_threads,
intra_op_parallelism_threads=flags_obj.intra_op_parallelism_threads,
allow_soft_placement=True)
distribution_strategy = distribution_utils.get_distribution_strategy(
distribution_strategy=flags_obj.distribution_strategy,
num_gpus=flags_core.get_num_gpus(flags_obj),
all_reduce_alg=flags_obj.all_reduce_alg)
run_config = tf.estimator.RunConfig(train_distribute=distribution_strategy,
session_config=session_config)
data_format = flags_obj.data_format
if data_format is None:
data_format = ('channels_first'
if tf.test.is_built_with_cuda() else 'channels_last')
mnist_classifier = tf.estimator.Estimator(model_fn=model_function,
model_dir=flags_obj.model_dir,
config=run_config,
params={
'data_format': data_format,
})
# Set up training and evaluation input functions.
def train_input_fn():
"""Prepare data for training."""
# When choosing shuffle buffer sizes, larger sizes result in better
# randomness, while smaller sizes use less memory. MNIST is a small
# enough dataset that we can easily shuffle the full epoch.
ds = dataset.train(flags_obj.data_dir)
ds = ds.cache().shuffle(buffer_size=50000).batch(flags_obj.batch_size)
# Iterate through the dataset a set number (`epochs_between_evals`) of times
# during each training session.
ds = ds.repeat(flags_obj.epochs_between_evals)
return ds
def eval_input_fn():
return dataset.test(flags_obj.data_dir).batch(
flags_obj.batch_size).make_one_shot_iterator().get_next()
# Set up hook that outputs training logs every 100 steps.
train_hooks = hooks_helper.get_train_hooks(flags_obj.hooks,
model_dir=flags_obj.model_dir,
batch_size=flags_obj.batch_size)
# Train and evaluate model.
for _ in range(flags_obj.train_epochs // flags_obj.epochs_between_evals):
mnist_classifier.train(input_fn=train_input_fn, hooks=train_hooks)
eval_results = mnist_classifier.evaluate(input_fn=eval_input_fn)
print('\nEvaluation results:\n\t%s\n' % eval_results)
if model_helpers.past_stop_threshold(flags_obj.stop_threshold,
eval_results['accuracy']):
break
# Export the model
if flags_obj.export_dir is not None:
image = tf.placeholder(tf.float32, [None, 28, 28])
input_fn = tf.estimator.export.build_raw_serving_input_receiver_fn({
'image':
image,
})
mnist_classifier.export_savedmodel(flags_obj.export_dir,
input_fn,
strip_default_attrs=True)
def run_loop(name,
train_input_fn,
eval_input_fn,
model_column_fn,
build_estimator_fn,
flags_obj,
tensors_to_log,
early_stop=False):
"""Define training loop."""
model_helpers.apply_clean(flags.FLAGS)
model = build_estimator_fn(model_dir=flags_obj.model_dir,
model_type=flags_obj.model_type,
model_column_fn=model_column_fn,
inter_op=flags_obj.inter_op_parallelism_threads,
intra_op=flags_obj.intra_op_parallelism_threads)
run_params = {
'batch_size': flags_obj.batch_size,
'train_epochs': flags_obj.train_epochs,
'model_type': flags_obj.model_type,
}
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info('wide_deep',
name,
run_params,
test_id=flags_obj.benchmark_test_id)
loss_prefix = LOSS_PREFIX.get(flags_obj.model_type, '')
tensors_to_log = {
k: v.format(loss_prefix=loss_prefix)
for k, v in tensors_to_log.items()
}
train_hooks = hooks_helper.get_train_hooks(flags_obj.hooks,
model_dir=flags_obj.model_dir,
batch_size=flags_obj.batch_size,
tensors_to_log=tensors_to_log)
# Train and evaluate the model every `flags.epochs_between_evals` epochs.
for n in range(flags_obj.train_epochs // flags_obj.epochs_between_evals):
if flags_obj.infer == False:
model.train(input_fn=train_input_fn, hooks=train_hooks)
results = model.evaluate(input_fn=eval_input_fn)
# Display evaluation metrics
tf.logging.info('Results at epoch %d / %d',
(n + 1) * flags_obj.epochs_between_evals,
flags_obj.train_epochs)
tf.logging.info('-' * 60)
for key in sorted(results):
tf.logging.info('%s: %s' % (key, results[key]))
benchmark_logger.log_evaluation_result(results)
if early_stop and model_helpers.past_stop_threshold(
flags_obj.stop_threshold, results['accuracy']):
break
# Export the model
if flags_obj.export_dir is not None:
export_model(model, flags_obj.model_type, flags_obj.export_dir,
model_column_fn)
def resnet_main(flags_obj,
model_function,
input_function,
dataset_name,
shape=None):
model_helpers.apply_clean(flags.FLAGS)
os.environ['TF_ENABLE_WINOGRAD_NONFUSED'] = '1'
session_config = tf.ConfigProto(
inter_op_parallelism_threads=flags_obj.inter_op_parallelism_threads,
intra_op_parallelism_threads=flags_obj.intra_op_parallelism_threads,
allow_soft_placement=True)
distribution_strategy = distribution_utils.get_distribution_strategy(
flags_core.get_num_gpus(flags_obj), flags_obj.all_reduce_alg)
run_config = tf.estimator.RunConfig(train_distribute=distribution_strategy,
session_config=session_config)
if flags_obj.pretrained_model_checkpoint_path is not None:
warm_start_settings = tf.estimator.WarmStartSettings(
flags_obj.pretrained_model_checkpoint_path,
vars_to_warm_start='^(?!.*dense)')
else:
warm_start_settings = None
classifier = tf.estimator.Estimator(
model_fn=model_function,
model_dir=flags_obj.model_dir,
config=run_config,
warm_start_from=warm_start_settings,
params={
'resnet_size': int(flags_obj.resnet_size),
'data_format': flags_obj.data_format,
'batch_size': flags_obj.batch_size,
'resnet_version': int(flags_obj.resnet_version),
'loss_scale': flags_core.get_loss_scale(flags_obj),
'dtype': flags_core.get_tf_dtype(flags_obj),
'fine_tune': flags_obj.fine_tune
})
run_params = {
'batch_size': flags_obj.batch_size,
'dtype': flags_core.get_tf_dtype(flags_obj),
'resnet_size': flags_obj.resnet_size,
'resnet_version': flags_obj.resnet_version,
'synthetic_data': flags_obj.use_synthetic_data,
'train_epochs': flags_obj.train_epochs,
}
if flags_obj.use_synthetic_data:
dataset_name = dataset_name + '-synthetic'
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info('resnet',
dataset_name,
run_params,
test_id=flags_obj.benchmark_test_id)
train_hooks = hooks_helper.get_train_hooks(flags_obj.hooks,
model_dir=flags_obj.model_dir,
batch_size=flags_obj.batch_size)
def input_fn_train(num_epochs):
return input_function(
is_training=True,
data_dir=flags_obj.data_dir,
batch_size=distribution_utils.per_device_batch_size(
flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
num_epochs=num_epochs,
num_gpus=flags_core.get_num_gpus(flags_obj),
dtype=flags_core.get_tf_dtype(flags_obj))
def input_fn_eval():
return input_function(
is_training=False,
data_dir=flags_obj.data_dir,
batch_size=distribution_utils.per_device_batch_size(
flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
num_epochs=1,
dtype=flags_core.get_tf_dtype(flags_obj))
if flags_obj.eval_only or not flags_obj.train_epochs:
schedule, n_loops = [0], 1
else:
n_loops = math.ceil(flags_obj.train_epochs /
flags_obj.epochs_between_evals)
schedule = [
flags_obj.epochs_between_evals for _ in range(int(n_loops))
]
schedule[-1] = flags_obj.train_epochs - sum(
schedule[:-1]) # over counting.
for cycle_index, num_train_epochs in enumerate(schedule):
tf.logging.info('Starting cycle: %d/%d', cycle_index, int(n_loops))
if num_train_epochs:
classifier.train(input_fn=lambda: input_fn_train(num_train_epochs),
hooks=train_hooks,
max_steps=flags_obj.max_train_steps)
tf.logging.info('Starting to evaluate.')
eval_results = classifier.evaluate(input_fn=input_fn_eval,
steps=flags_obj.max_train_steps)
benchmark_logger.log_evaluation_result(eval_results)
if model_helpers.past_stop_threshold(flags_obj.stop_threshold,
eval_results['accuracy']):
break
if flags_obj.export_dir is not None:
dtype = flags_core.get_tf_dtype(flags_obj)
input_receiver_fn = export.build_tensor_serving_input_receiver_fn(
shape, batch_size=flags_obj.batch_size, dtype=dtype)
classifier.export_savedmodel(flags_obj.export_dir, input_receiver_fn)
def resnet_main(flags_obj,
model_function,
input_function,
dataset_name,
shape=None):
"""Shared main loop for ResNet Models.
Args:
flags_obj: An object containing parsed flags. See define_resnet_flags()
for details.
model_function: the function that instantiates the Model and builds the
ops for train/eval. This will be passed directly into the estimator.
input_function: the function that processes the dataset and returns a
dataset that the estimator can train on. This will be wrapped with
all the relevant flags for running and passed to estimator.
dataset_name: the name of the dataset for training and evaluation. This is
used for logging purpose.
shape: list of ints representing the shape of the images used for training.
This is only used if flags_obj.export_dir is passed.
"""
model_helpers.apply_clean(flags.FLAGS)
# Using the Winograd non-fused algorithms provides a small performance boost.
os.environ['TF_ENABLE_WINOGRAD_NONFUSED'] = '1'
# Create session config based on values of inter_op_parallelism_threads and
# intra_op_parallelism_threads. Note that we default to having
# allow_soft_placement = True, which is required for multi-GPU and not
# harmful for other modes.
session_config = tf.ConfigProto(
inter_op_parallelism_threads=flags_obj.inter_op_parallelism_threads,
intra_op_parallelism_threads=flags_obj.intra_op_parallelism_threads,
allow_soft_placement=True)
distribution_strategy = distribution_utils.get_distribution_strategy(
flags_core.get_num_gpus(flags_obj), flags_obj.all_reduce_alg)
run_config = tf.estimator.RunConfig(train_distribute=distribution_strategy,
session_config=session_config)
# initialize our model with all but the dense layer from pretrained resnet
if flags_obj.pretrained_model_checkpoint_path is not None:
warm_start_settings = tf.estimator.WarmStartSettings(
flags_obj.pretrained_model_checkpoint_path,
vars_to_warm_start='^(?!.*dense)')
else:
warm_start_settings = None
classifier = tf.estimator.Estimator(
model_fn=model_function,
model_dir=flags_obj.model_dir,
config=run_config,
warm_start_from=warm_start_settings,
params={
'resnet_size': int(flags_obj.resnet_size),
'data_format': flags_obj.data_format,
'batch_size': flags_obj.batch_size,
'resnet_version': int(flags_obj.resnet_version),
'loss_scale': flags_core.get_loss_scale(flags_obj),
'dtype': flags_core.get_tf_dtype(flags_obj),
'fine_tune': flags_obj.fine_tune
})
run_params = {
'batch_size': flags_obj.batch_size,
'dtype': flags_core.get_tf_dtype(flags_obj),
'resnet_size': flags_obj.resnet_size,
'resnet_version': flags_obj.resnet_version,
'synthetic_data': flags_obj.use_synthetic_data,
'train_epochs': flags_obj.train_epochs,
}
if flags_obj.use_synthetic_data:
dataset_name = dataset_name + '-synthetic'
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info('resnet',
dataset_name,
run_params,
test_id=flags_obj.benchmark_test_id)
train_hooks = hooks_helper.get_train_hooks(flags_obj.hooks,
model_dir=flags_obj.model_dir,
batch_size=flags_obj.batch_size)
def input_fn_train(num_epochs):
return input_function(
is_training=True,
data_dir=flags_obj.data_dir,
batch_size=distribution_utils.per_device_batch_size(
flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
num_epochs=num_epochs,
num_gpus=flags_core.get_num_gpus(flags_obj),
dtype=flags_core.get_tf_dtype(flags_obj))
def input_fn_eval():
return input_function(
is_training=False,
data_dir=flags_obj.data_dir,
batch_size=distribution_utils.per_device_batch_size(
flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
num_epochs=1,
dtype=flags_core.get_tf_dtype(flags_obj))
if flags_obj.eval_only or not flags_obj.train_epochs:
# If --eval_only is set, perform a single loop with zero train epochs.
schedule, n_loops = [0], 1
else:
# Compute the number of times to loop while training. All but the last
# pass will train for `epochs_between_evals` epochs, while the last will
# train for the number needed to reach `training_epochs`. For instance if
# train_epochs = 25 and epochs_between_evals = 10
# schedule will be set to [10, 10, 5]. That is to say, the loop will:
# Train for 10 epochs and then evaluate.
# Train for another 10 epochs and then evaluate.
# Train for a final 5 epochs (to reach 25 epochs) and then evaluate.
n_loops = math.ceil(flags_obj.train_epochs /
flags_obj.epochs_between_evals)
schedule = [
flags_obj.epochs_between_evals for _ in range(int(n_loops))
]
schedule[-1] = flags_obj.train_epochs - sum(
schedule[:-1]) # over counting.
for cycle_index, num_train_epochs in enumerate(schedule):
tf.logging.info('Starting cycle: %d/%d', cycle_index, int(n_loops))
if num_train_epochs:
classifier.train(input_fn=lambda: input_fn_train(num_train_epochs),
hooks=train_hooks,
max_steps=flags_obj.max_train_steps)
tf.logging.info('Starting to evaluate.')
# flags_obj.max_train_steps is generally associated with testing and
# profiling. As a result it is frequently called with synthetic data, which
# will iterate forever. Passing steps=flags_obj.max_train_steps allows the
# eval (which is generally unimportant in those circumstances) to terminate.
# Note that eval will run for max_train_steps each loop, regardless of the
# global_step count.
eval_results = classifier.evaluate(input_fn=input_fn_eval,
steps=flags_obj.max_train_steps)
benchmark_logger.log_evaluation_result(eval_results)
if model_helpers.past_stop_threshold(flags_obj.stop_threshold,
eval_results['accuracy']):
break
if flags_obj.export_dir is not None:
# Exports a saved model for the given classifier.
dtype = flags_core.get_tf_dtype(flags_obj)
input_receiver_fn = export.build_tensor_serving_input_receiver_fn(
shape, batch_size=flags_obj.batch_size, dtype=dtype)
classifier.export_savedmodel(flags_obj.export_dir, input_receiver_fn)
def run_mnist(flags_obj):
"""Run MNIST training and eval loop.
Args:
flags_obj: An object containing parsed flag values.
"""
model_helpers.apply_clean(flags_obj)
model_function = model_fn
# Get number of GPUs as defined by the --num_gpus flags and the number of
# GPUs available on the machine.
num_gpus = flags_core.get_num_gpus(flags_obj)
multi_gpu = num_gpus > 1
if multi_gpu:
# Validate that the batch size can be split into devices.
distribution_utils.per_device_batch_size(flags_obj.batch_size,
num_gpus)
# There are two steps required if using multi-GPU: (1) wrap the model_fn,
# and (2) wrap the optimizer. The first happens here, and (2) happens
# in the model_fn itself when the optimizer is defined.
model_function = tf.contrib.estimator.replicate_model_fn(
model_fn,
loss_reduction=tf.losses.Reduction.MEAN,
devices=["/device:GPU:%d" % d for d in range(num_gpus)])
data_format = flags_obj.data_format
if data_format is None:
data_format = ('channels_first'
if tf.test.is_built_with_cuda() else 'channels_last')
mnist_classifier = tf.estimator.Estimator(model_fn=model_function,
model_dir=flags_obj.model_dir,
params={
'data_format': data_format,
'multi_gpu': multi_gpu
})
# Set up training and evaluation input functions.
def train_input_fn():
"""Prepare data for training."""
# When choosing shuffle buffer sizes, larger sizes result in better
# randomness, while smaller sizes use less memory. MNIST is a small
# enough dataset that we can easily shuffle the full epoch.
ds = dataset.train(flags_obj.data_dir)
ds = ds.cache().shuffle(buffer_size=50000).batch(flags_obj.batch_size)
# Iterate through the dataset a set number (`epochs_between_evals`) of times
# during each training session.
ds = ds.repeat(flags_obj.epochs_between_evals)
return ds
def eval_input_fn():
return dataset.test(flags_obj.data_dir).batch(
flags_obj.batch_size).make_one_shot_iterator().get_next()
# Set up hook that outputs training logs every 100 steps.
train_hooks = hooks_helper.get_train_hooks(flags_obj.hooks,
model_dir=flags_obj.model_dir,
batch_size=flags_obj.batch_size)
# Train and evaluate model.
for _ in range(flags_obj.train_epochs // flags_obj.epochs_between_evals):
mnist_classifier.train(input_fn=train_input_fn, hooks=train_hooks)
eval_results = mnist_classifier.evaluate(input_fn=eval_input_fn)
print('\nEvaluation results:\n\t%s\n' % eval_results)
if model_helpers.past_stop_threshold(flags_obj.stop_threshold,
eval_results['accuracy']):
break
# Export the model
if flags_obj.export_dir is not None:
image = tf.placeholder(tf.float32, [None, 28, 28])
input_fn = tf.estimator.export.build_raw_serving_input_receiver_fn({
'image':
image,
})
mnist_classifier.export_savedmodel(flags_obj.export_dir, input_fn)
def resnet_main(flags_obj,
model_function,
input_function,
dataset_name,
shape=None):
"""Shared main loop for ResNet Models.
Args:
flags_obj: An object containing parsed flags. See define_resnet_flags()
for details.
model_function: the function that instantiates the Model and builds the
ops for train/eval. This will be passed directly into the estimator.
input_function: the function that processes the dataset and returns a
dataset that the estimator can train on. This will be wrapped with
all the relevant flags for running and passed to estimator.
dataset_name: the name of the dataset for training and evaluation. This is
used for logging purpose.
shape: list of ints representing the shape of the images used for training.
This is only used if flags_obj.export_dir is passed.
"""
model_helpers.apply_clean(flags.FLAGS)
# Ensures flag override logic is only executed if explicitly triggered.
if flags_obj.tf_gpu_thread_mode:
override_flags_and_set_envars_for_gpu_thread_pool(flags_obj)
# Creates session config. allow_soft_placement = True, is required for
# multi-GPU and is not harmful for other modes.
session_config = tf.ConfigProto(
inter_op_parallelism_threads=flags_obj.inter_op_parallelism_threads,
intra_op_parallelism_threads=flags_obj.intra_op_parallelism_threads,
allow_soft_placement=True)
distribution_strategy = distribution_utils.get_distribution_strategy(
flags_core.get_num_gpus(flags_obj), flags_obj.all_reduce_alg)
# Creates a `RunConfig` that checkpoints every 24 hours which essentially
# results in checkpoints determined only by `epochs_between_evals`.
run_config = tf.estimator.RunConfig(train_distribute=distribution_strategy,
session_config=session_config,
save_checkpoints_secs=60 * 60 * 24)
# Initializes model with all but the dense layer from pretrained ResNet.
if flags_obj.pretrained_model_checkpoint_path is not None:
if flags_obj.fine_tune:
if string.lower(flags_obj.optimizer) == 'adam':
if flags_obj.no_dense_init:
warm_start_settings = tf.estimator.WarmStartSettings(
flags_obj.pretrained_model_checkpoint_path,
vars_to_warm_start=[
'^(?!.*(resnet_model/dense|beta1_power|beta2_power|Adam|global_step))'
])
# vars_to_warm_start=['^(?!.*(resnet_model/dense|global_step))'])
else:
warm_start_settings = tf.estimator.WarmStartSettings(
flags_obj.pretrained_model_checkpoint_path,
vars_to_warm_start=[
'^(?!.*(resnet_model/dense/kernel/Momentum|resnet_model/dense/bias/Momentum|beta1_power|beta2_power|Adam|global_step))'
])
# vars_to_warm_start=['^(?!.*(resnet_model/dense|global_step))'])
else:
if flags_obj.no_dense_init:
warm_start_settings = tf.estimator.WarmStartSettings(
flags_obj.pretrained_model_checkpoint_path,
vars_to_warm_start=[
'^(?!.*(resnet_model/dense|Momentum|global_step))'
])
else:
warm_start_settings = tf.estimator.WarmStartSettings(
flags_obj.pretrained_model_checkpoint_path,
vars_to_warm_start=[
'^(?!.*(resnet_model/dense/kernel/Momentum|resnet_model/dense/bias/Momentum|global_step))'
])
# vars_to_warm_start=['^(?!.*(resnet_model/dense|global_step))'])
else:
if string.lower(flags_obj.optimizer) == 'adam':
warm_start_settings = tf.estimator.WarmStartSettings(
flags_obj.pretrained_model_checkpoint_path,
vars_to_warm_start=[
'^(?!.*(endecoder|Momentum|beta1_power|beta2_power|global_step))'
])
# vars_to_warm_start='^(?!.*dense)')
else:
warm_start_settings = tf.estimator.WarmStartSettings(
flags_obj.pretrained_model_checkpoint_path,
vars_to_warm_start=['^(?!.*(endecoder|global_step))'])
# vars_to_warm_start='^(?!.*dense)')
else:
warm_start_settings = None
classifier = tf.estimator.Estimator(
model_fn=model_function,
model_dir=flags_obj.model_dir,
config=run_config,
warm_start_from=warm_start_settings,
params={
'resnet_size': int(flags_obj.resnet_size),
'data_format': flags_obj.data_format,
'batch_size': flags_obj.batch_size,
'resnet_version': int(flags_obj.resnet_version),
'loss_scale': flags_core.get_loss_scale(flags_obj),
'dtype': flags_core.get_tf_dtype(flags_obj),
'fine_tune': flags_obj.fine_tune,
'reconst_loss_scale': flags_obj.reconst_loss_scale,
'use_ce': flags_obj.use_ce,
'optimizer': string.lower(flags_obj.optimizer),
'clip_grad': flags_obj.clip_grad,
'spectral_norm': flags_obj.spectral_norm,
'ce_scale': flags_obj.ce_scale,
'sep_grad_nrom': flags_obj.sep_grad_nrom,
'norm_teach_feature': flags_obj.norm_teach_feature,
'no_dense_init': flags_obj.no_dense_init,
'compress_ratio': flags_obj.compress_ratio
})
run_params = {
'batch_size': flags_obj.batch_size,
'dtype': flags_core.get_tf_dtype(flags_obj),
'resnet_size': flags_obj.resnet_size,
'resnet_version': flags_obj.resnet_version,
'synthetic_data': flags_obj.use_synthetic_data,
'train_epochs': flags_obj.train_epochs,
'fine_tune': flags_obj.fine_tune,
'reconst_loss_scale': flags_obj.reconst_loss_scale,
'use_ce': flags_obj.use_ce,
'optimizer': string.lower(flags_obj.optimizer),
'clip_grad': flags_obj.clip_grad,
'spectral_norm': flags_obj.spectral_norm,
'ce_scale': flags_obj.ce_scale,
'sep_grad_nrom': flags_obj.sep_grad_nrom,
'norm_teach_feature': flags_obj.norm_teach_feature,
'no_dense_init': flags_obj.no_dense_init,
'compress_ratio': flags_obj.compress_ratio,
}
if flags_obj.use_synthetic_data:
dataset_name = dataset_name + '-synthetic'
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info('resnet',
dataset_name,
run_params,
test_id=flags_obj.benchmark_test_id)
train_hooks = hooks_helper.get_train_hooks(flags_obj.hooks,
model_dir=flags_obj.model_dir,
batch_size=flags_obj.batch_size)
def input_fn_train(num_epochs):
return input_function(
is_training=True,
data_dir=flags_obj.data_dir,
batch_size=distribution_utils.per_device_batch_size(
flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
num_epochs=num_epochs,
dtype=flags_core.get_tf_dtype(flags_obj),
datasets_num_private_threads=flags_obj.
datasets_num_private_threads,
num_parallel_batches=flags_obj.datasets_num_parallel_batches)
def input_fn_eval():
return input_function(
is_training=False,
data_dir=flags_obj.data_dir,
batch_size=distribution_utils.per_device_batch_size(
flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
num_epochs=1,
dtype=flags_core.get_tf_dtype(flags_obj))
if flags_obj.eval_only or not flags_obj.train_epochs:
# If --eval_only is set, perform a single loop with zero train epochs.
schedule, n_loops = [0], 1
else:
# Compute the number of times to loop while training. All but the last
# pass will train for `epochs_between_evals` epochs, while the last will
# train for the number needed to reach `training_epochs`. For instance if
# train_epochs = 25 and epochs_between_evals = 10
# schedule will be set to [10, 10, 5]. That is to say, the loop will:
# Train for 10 epochs and then evaluate.
# Train for another 10 epochs and then evaluate.
# Train for a final 5 epochs (to reach 25 epochs) and then evaluate.
n_loops = math.ceil(flags_obj.train_epochs /
flags_obj.epochs_between_evals)
schedule = [
flags_obj.epochs_between_evals for _ in range(int(n_loops))
]
schedule[-1] = flags_obj.train_epochs - sum(
schedule[:-1]) # over counting.
print('schedule: ', schedule, flags_obj.epochs_between_evals,
flags_obj.max_train_steps)
for cycle_index, num_train_epochs in enumerate(schedule):
tf.logging.info('Starting cycle: %d/%d', cycle_index, int(n_loops))
if num_train_epochs:
classifier.train(input_fn=lambda: input_fn_train(num_train_epochs),
hooks=train_hooks,
max_steps=flags_obj.max_train_steps)
tf.logging.info('Starting to evaluate.')
# flags_obj.max_train_steps is generally associated with testing and
# profiling. As a result it is frequently called with synthetic data, which
# will iterate forever. Passing steps=flags_obj.max_train_steps allows the
# eval (which is generally unimportant in those circumstances) to terminate.
# Note that eval will run for max_train_steps each loop, regardless of the
# global_step count.
eval_results = classifier.evaluate(input_fn=input_fn_eval,
steps=flags_obj.max_train_steps)
benchmark_logger.log_evaluation_result(eval_results)
if model_helpers.past_stop_threshold(flags_obj.stop_threshold,
eval_results['accuracy']):
break
if flags_obj.export_dir is not None:
# Exports a saved model for the given classifier.
export_dtype = flags_core.get_tf_dtype(flags_obj)
if flags_obj.image_bytes_as_serving_input:
input_receiver_fn = functools.partial(image_bytes_serving_input_fn,
shape,
dtype=export_dtype)
else:
input_receiver_fn = export.build_tensor_serving_input_receiver_fn(
shape, batch_size=flags_obj.batch_size, dtype=export_dtype)
classifier.export_savedmodel(flags_obj.export_dir,
input_receiver_fn,
strip_default_attrs=True)
def run_mnist_eager(flags_obj):
"""
Run MNIST training and eval loop in eager mode.
Args:
flags_obj: An object containing parsed flag values.
"""
# Soft placement
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
tfe.enable_eager_execution(config=config)
model_helpers.apply_clean(flags.FLAGS)
# Automatically determine device and data_format
(device, data_format) = ('/gpu:0', 'channels_first')
if flags_obj.no_gpu or not tf.test.is_gpu_available():
(device, data_format) = ('/cpu:0', 'channels_last')
# If data_format is defined in FLAGS, overwrite automatically set value.
if flags_obj.data_format is not None:
data_format = flags_obj.data_format
print('Using device %s, and data format %s.' % (device, data_format))
# Load the datasets
train_ds = mnist_dataset.train(flags_obj.data_dir).shuffle(60000).batch(
flags_obj.batch_size)
test_ds = mnist_dataset.test(flags_obj.data_dir).batch(
flags_obj.batch_size)
# Create the model and optimizer
model = model_lib.create_model(data_format)
optimizer = tf.train.MomentumOptimizer(flags_obj.lr, flags_obj.momentum)
# Print model summary
print(model.summary())
# Create file writers for writing TensorBoard summaries.
if flags_obj.output_dir:
# Create directories to which summaries will be written
# tensorboard --logdir=<output_dir>
# can then be used to see the recorded summaries.
train_dir = os.path.join(flags_obj.output_dir, 'train')
test_dir = os.path.join(flags_obj.output_dir, 'eval')
tf.gfile.MakeDirs(flags_obj.output_dir)
else:
train_dir = None
test_dir = None
summary_writer = tf.contrib.summary.create_file_writer(train_dir,
flush_millis=10000)
test_summary_writer = tf.contrib.summary.create_file_writer(
test_dir, flush_millis=10000, name='test')
# Create and restore checkpoint (if one exists on the path)
checkpoint_prefix = os.path.join(flags_obj.model_dir, 'ckpt')
step_counter = tf.train.get_or_create_global_step()
checkpoint = tf.train.Checkpoint(model=model,
optimizer=optimizer,
step_counter=step_counter)
# Restore variables on creation if a checkpoint exists.
checkpoint.restore(tf.train.latest_checkpoint(flags_obj.model_dir))
# Train and evaluate for a set number of epochs.
with tf.device(device):
for _ in range(flags_obj.train_epochs):
start = time.time()
with summary_writer.as_default():
train(model, optimizer, train_ds, step_counter,
flags_obj.log_interval)
end = time.time()
# Note time taken
print('\nTrain time for epoch #%d (%d total steps): %f' %
(checkpoint.save_counter.numpy() + 1, step_counter.numpy(),
end - start))
with test_summary_writer.as_default():
test(model, test_ds)
checkpoint.save(checkpoint_prefix)
def run_ncf(FLAGS):
"""Run NCF training and eval with Keras."""
#########################################################################
# Construct AutoDist with ResourceSpec for Different Strategies
resource_spec_file = os.path.join(
os.path.dirname(__file__),
'../resource_spec.yml')
resource_info = yaml.safe_load(open(resource_spec_file, 'r'))
try:
node_num = len(resource_info['nodes'])
except ValueError:
print("nodes need to be set in specficiation file")
try:
gpu_num = len(resource_info['nodes'][0]['gpus'])
except ValueError:
print("gpus need to be set in specficiation file")
if FLAGS.autodist_patch_tf:
os.environ['AUTODIST_PATCH_TF'] = '1'
else:
os.environ['AUTODIST_PATCH_TF'] = '0'
if FLAGS.proxy:
local_proxy_variable = True
else:
local_proxy_variable = False
if FLAGS.autodist_strategy == 'PS':
autodist = AutoDist(
resource_spec_file, PS(
local_proxy_variable=local_proxy_variable))
elif FLAGS.autodist_strategy == 'PSLoadBalancing':
autodist = AutoDist(resource_spec_file, PSLoadBalancing(
local_proxy_variable=local_proxy_variable))
elif FLAGS.autodist_strategy == 'PartitionedPS':
autodist = AutoDist(resource_spec_file, PartitionedPS(
local_proxy_variable=local_proxy_variable))
elif FLAGS.autodist_strategy == 'AllReduce':
autodist = AutoDist(resource_spec_file, AllReduce(chunk_size=256))
elif FLAGS.autodist_strategy == 'Parallax':
autodist = AutoDist(
resource_spec_file,
Parallax(
chunk_size=256,
local_proxy_variable=local_proxy_variable))
else:
raise ValueError(
'the strategy can be only from PS, PSLoadBalancing, PartitionedPS, AllReduce, Parallax')
#########################################################################
if FLAGS.seed is not None:
print("Setting tf seed")
tf.random.set_seed(FLAGS.seed)
model_helpers.apply_clean(FLAGS)
if FLAGS.dtype == "fp16" and FLAGS.fp16_implementation == "keras":
policy = tf.keras.mixed_precision.experimental.Policy(
"mixed_float16", loss_scale=flags_core.get_loss_scale(
FLAGS, default_for_fp16="dynamic"))
tf.keras.mixed_precision.experimental.set_policy(policy)
params = ncf_common.parse_flags(FLAGS)
params["distribute_strategy"] = None
batch_size = params["batch_size"]
time_callback = keras_utils.TimeHistory(batch_size, FLAGS.log_steps)
callbacks = [time_callback]
producer, input_meta_data = None, None
generate_input_online = params["train_dataset_path"] is None
if generate_input_online:
num_users, num_items, _, _, producer = ncf_common.get_inputs(params)
producer.start()
per_epoch_callback = IncrementEpochCallback(producer)
callbacks.append(per_epoch_callback)
else:
assert params["eval_dataset_path"] and params["input_meta_data_path"]
with tf.io.gfile.GFile(params["input_meta_data_path"], "rb") as reader:
input_meta_data = json.loads(reader.read().decode("utf-8"))
num_users = input_meta_data["num_users"]
num_items = input_meta_data["num_items"]
params["num_users"], params["num_items"] = num_users, num_items
if FLAGS.early_stopping:
early_stopping_callback = CustomEarlyStopping(
"val_HR_METRIC", desired_value=FLAGS.hr_threshold)
callbacks.append(early_stopping_callback)
with tf.Graph().as_default(), autodist.scope():
(train_input_dataset, eval_input_dataset, num_train_steps, num_eval_steps) = (
ncf_input_pipeline.create_ncf_input_data(params, producer, input_meta_data, None))
steps_per_epoch = None if generate_input_online else num_train_steps
keras_model = _get_keras_model(params)
if FLAGS.optimizer == 'adam':
optimizer = tf.keras.optimizers.Adam(
learning_rate=params["learning_rate"],
beta_1=params["beta1"],
beta_2=params["beta2"],
epsilon=params["epsilon"])
elif FLAGS.optimizer == 'sgd':
optimizer = tf.keras.optimizers.SGD(
learning_rate=params["learning_rate"])
elif FLAGS.optimizer == 'lazyadam':
optimizer = LazyAdam(
learning_rate=params["learning_rate"],
beta_1=params["beta1"],
beta_2=params["beta2"],
epsilon=params["epsilon"])
else:
raise ValueError('Do not support other optimizers...')
if FLAGS.fp16_implementation == "graph_rewrite":
optimizer = tf.compat.v1.train.experimental.enable_mixed_precision_graph_rewrite(
optimizer, loss_scale=flags_core.get_loss_scale(FLAGS, default_for_fp16="dynamic"))
elif FLAGS.dtype == "fp16" and params["keras_use_ctl"]:
optimizer = tf.keras.mixed_precision.experimental.LossScaleOptimizer(
optimizer, tf.keras.mixed_precision.experimental.global_policy().loss_scale)
return run_ncf_custom_training(
params,
autodist,
keras_model,
optimizer,
callbacks,
train_input_dataset,
eval_input_dataset,
num_train_steps,
num_eval_steps,
generate_input_online=generate_input_online,
return_simulation=FLAGS.simulation_strategy_id is not None)
def resnet_main(flags_obj,
model_function,
input_function,
dataset_name,
shape=None):
"""Shared main loop for ResNet Models.
Args:
flags_obj: An object containing parsed flags. See define_resnet_flags()
for details.
model_function: the function that instantiates the Model and builds the
ops for train/eval. This will be passed directly into the estimator.
input_function: the function that processes the dataset and returns a
dataset that the estimator can train on. This will be wrapped with
all the relevant flags for running and passed to estimator.
dataset_name: the name of the dataset for training and evaluation. This is
used for logging purpose.
shape: list of ints representing the shape of the images used for training.
This is only used if flags_obj.export_dir is passed.
"""
model_helpers.apply_clean(flags.FLAGS)
# Using the Winograd non-fused algorithms provides a small performance boost.
os.environ['TF_ENABLE_WINOGRAD_NONFUSED'] = '1'
# Create session config based on values of inter_op_parallelism_threads and
# intra_op_parallelism_threads. Note that we default to having
# allow_soft_placement = True, which is required for multi-GPU and not
# harmful for other modes.
session_config = tf.ConfigProto(
inter_op_parallelism_threads=flags_obj.inter_op_parallelism_threads,
intra_op_parallelism_threads=flags_obj.intra_op_parallelism_threads,
allow_soft_placement=True)
distribution_strategy = distribution_utils.get_distribution_strategy(
flags_core.get_num_gpus(flags_obj), flags_obj.all_reduce_alg)
run_config = tf.estimator.RunConfig(train_distribute=distribution_strategy,
session_config=session_config)
classifier = tf.estimator.Estimator(
model_fn=model_function,
model_dir=flags_obj.model_dir,
config=run_config,
params={
'resnet_size': int(flags_obj.resnet_size),
'data_format': flags_obj.data_format,
'batch_size': flags_obj.batch_size,
'resnet_version': int(flags_obj.resnet_version),
'loss_scale': flags_core.get_loss_scale(flags_obj),
'dtype': flags_core.get_tf_dtype(flags_obj)
})
run_params = {
'batch_size': flags_obj.batch_size,
'dtype': flags_core.get_tf_dtype(flags_obj),
'resnet_size': flags_obj.resnet_size,
'resnet_version': flags_obj.resnet_version,
'synthetic_data': flags_obj.use_synthetic_data,
'train_epochs': flags_obj.train_epochs,
}
if flags_obj.use_synthetic_data:
dataset_name = dataset_name + '-synthetic'
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info('resnet',
dataset_name,
run_params,
test_id=flags_obj.benchmark_test_id)
train_hooks = hooks_helper.get_train_hooks(flags_obj.hooks,
model_dir=flags_obj.model_dir,
batch_size=flags_obj.batch_size)
def input_fn_train():
return input_function(
is_training=True,
data_dir=flags_obj.data_dir,
batch_size=distribution_utils.per_device_batch_size(
flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
num_epochs=flags_obj.epochs_between_evals,
num_gpus=flags_core.get_num_gpus(flags_obj))
def input_fn_eval():
return input_function(
is_training=False,
data_dir=flags_obj.data_dir,
batch_size=distribution_utils.per_device_batch_size(
flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
num_epochs=1)
total_training_cycle = (flags_obj.train_epochs //
flags_obj.epochs_between_evals)
for cycle_index in range(total_training_cycle):
tf.logging.info('Starting a training cycle: %d/%d', cycle_index,
total_training_cycle)
classifier.train(input_fn=input_fn_train,
hooks=train_hooks,
max_steps=flags_obj.max_train_steps)
tf.logging.info('Starting to evaluate.')
# flags_obj.max_train_steps is generally associated with testing and
# profiling. As a result it is frequently called with synthetic data, which
# will iterate forever. Passing steps=flags_obj.max_train_steps allows the
# eval (which is generally unimportant in those circumstances) to terminate.
# Note that eval will run for max_train_steps each loop, regardless of the
# global_step count.
eval_results = classifier.evaluate(input_fn=input_fn_eval,
steps=flags_obj.max_train_steps)
benchmark_logger.log_evaluation_result(eval_results)
if model_helpers.past_stop_threshold(flags_obj.stop_threshold,
eval_results['accuracy']):
break
if flags_obj.export_dir is not None:
# Exports a saved model for the given classifier.
input_receiver_fn = export.build_tensor_serving_input_receiver_fn(
shape, batch_size=flags_obj.batch_size)
classifier.export_savedmodel(flags_obj.export_dir, input_receiver_fn)
def resnet_main(flags_obj,
model_function,
input_function,
dataset_name,
shape=None):
"""Shared main loop for ResNet Models.
Args:
flags_obj: An object containing parsed flags. See define_resnet_flags()
for details.
model_function: the function that instantiates the Model and builds the
ops for train/eval. This will be passed directly into the estimator.
input_function: the function that processes the dataset and returns a
dataset that the estimator can train on. This will be wrapped with
all the relevant flags for running and passed to estimator.
dataset_name: the name of the dataset for training and evaluation. This is
used for logging purpose.
shape: list of ints representing the shape of the images used for training.
This is only used if flags_obj.export_dir is passed.
Returns:
Dict of results of the run. Contains the keys `eval_results` and
`train_hooks`. `eval_results` contains accuracy (top_1) and accuracy_top_5.
`train_hooks` is a list the instances of hooks used during training.
"""
model_helpers.apply_clean(flags.FLAGS)
# Ensures flag override logic is only executed if explicitly triggered.
if flags_obj.tf_gpu_thread_mode:
override_flags_and_set_envars_for_gpu_thread_pool(flags_obj)
# Configures cluster spec for distribution strategy.
num_workers = distribution_utils.configure_cluster(flags_obj.worker_hosts,
flags_obj.task_index)
# Creates session config. allow_soft_placement = True, is required for
# multi-GPU and is not harmful for other modes.
session_config = tf.compat.v1.ConfigProto(
inter_op_parallelism_threads=flags_obj.inter_op_parallelism_threads,
intra_op_parallelism_threads=flags_obj.intra_op_parallelism_threads,
allow_soft_placement=True)
distribution_strategy = distribution_utils.get_distribution_strategy(
distribution_strategy=flags_obj.distribution_strategy,
num_gpus=flags_core.get_num_gpus(flags_obj),
num_workers=num_workers,
all_reduce_alg=flags_obj.all_reduce_alg,
num_packs=flags_obj.num_packs)
# Creates a `RunConfig` that checkpoints every 24 hours which essentially
# results in checkpoints determined only by `epochs_between_evals`.
run_config = tf.estimator.RunConfig(train_distribute=distribution_strategy,
session_config=session_config,
save_checkpoints_secs=None,
save_checkpoints_steps=2000)
# Initializes model with all but the dense layer from pretrained ResNet.
if flags_obj.pretrained_model_checkpoint_path is not None:
warm_start_settings = tf.estimator.WarmStartSettings(
flags_obj.pretrained_model_checkpoint_path,
vars_to_warm_start='^(?!.*dense)')
else:
warm_start_settings = None
classifier = tf.estimator.Estimator(model_fn=model_function,
model_dir=flags_obj.model_dir,
config=run_config,
warm_start_from=warm_start_settings,
params={
'resnet_size':
int(flags_obj.resnet_size),
'data_format':
flags_obj.data_format,
'batch_size':
flags_obj.batch_size,
'resnet_version':
int(flags_obj.resnet_version),
'loss_scale':
flags_core.get_loss_scale(
flags_obj,
default_for_fp16=128),
'dtype':
flags_core.get_tf_dtype(flags_obj),
'fine_tune':
flags_obj.fine_tune,
'num_workers':
num_workers,
})
run_params = {
'batch_size': flags_obj.batch_size,
'dtype': flags_core.get_tf_dtype(flags_obj),
'resnet_size': flags_obj.resnet_size,
'resnet_version': flags_obj.resnet_version,
'synthetic_data': flags_obj.use_synthetic_data,
'train_epochs': flags_obj.train_epochs,
'num_workers': num_workers,
}
if flags_obj.use_synthetic_data:
dataset_name = dataset_name + '-synthetic'
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info('resnet',
dataset_name,
run_params,
test_id=flags_obj.benchmark_test_id)
train_hooks = hooks_helper.get_train_hooks(flags_obj.hooks,
model_dir=flags_obj.model_dir,
batch_size=flags_obj.batch_size)
def input_fn_train(num_epochs, input_context=None):
return input_function(
is_training=True,
data_dir=flags_obj.data_dir,
batch_size=distribution_utils.per_replica_batch_size(
flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
num_epochs=num_epochs,
dtype=flags_core.get_tf_dtype(flags_obj),
datasets_num_private_threads=flags_obj.
datasets_num_private_threads,
input_context=input_context)
def input_fn_eval():
return input_function(
is_training=False,
data_dir=flags_obj.data_dir,
batch_size=distribution_utils.per_replica_batch_size(
flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
num_epochs=1,
dtype=flags_core.get_tf_dtype(flags_obj))
train_epochs = (0 if flags_obj.eval_only or not flags_obj.train_epochs else
flags_obj.train_epochs)
use_train_and_evaluate = flags_obj.use_train_and_evaluate or num_workers > 1
if use_train_and_evaluate:
train_spec = tf.estimator.TrainSpec(
input_fn=lambda input_context=None: input_fn_train(
train_epochs, input_context=input_context),
hooks=train_hooks,
max_steps=flags_obj.max_train_steps)
eval_spec = tf.estimator.EvalSpec(input_fn=input_fn_eval)
tf.compat.v1.logging.info('Starting to train and evaluate.')
tf.estimator.train_and_evaluate(classifier, train_spec, eval_spec)
# tf.estimator.train_and_evalute doesn't return anything in multi-worker
# case.
eval_results = {}
else:
if train_epochs == 0:
# If --eval_only is set, perform a single loop with zero train epochs.
schedule, n_loops = [0], 1
else:
# Compute the number of times to loop while training. All but the last
# pass will train for `epochs_between_evals` epochs, while the last will
# train for the number needed to reach `training_epochs`. For instance if
# train_epochs = 25 and epochs_between_evals = 10
# schedule will be set to [10, 10, 5]. That is to say, the loop will:
# Train for 10 epochs and then evaluate.
# Train for another 10 epochs and then evaluate.
# Train for a final 5 epochs (to reach 25 epochs) and then evaluate.
n_loops = math.ceil(train_epochs / flags_obj.epochs_between_evals)
schedule = [
flags_obj.epochs_between_evals for _ in range(int(n_loops))
]
schedule[-1] = train_epochs - sum(schedule[:-1]) # over counting.
for cycle_index, num_train_epochs in enumerate(schedule):
tf.compat.v1.logging.info('Starting cycle: %d/%d', cycle_index,
int(n_loops))
if num_train_epochs:
# Since we are calling classifier.train immediately in each loop, the
# value of num_train_epochs in the lambda function will not be changed
# before it is used. So it is safe to ignore the pylint error here
# pylint: disable=cell-var-from-loop
classifier.train(
input_fn=lambda input_context=None: input_fn_train(
num_train_epochs, input_context=input_context),
hooks=train_hooks,
max_steps=flags_obj.max_train_steps)
# flags_obj.max_train_steps is generally associated with testing and
# profiling. As a result it is frequently called with synthetic data,
# which will iterate forever. Passing steps=flags_obj.max_train_steps
# allows the eval (which is generally unimportant in those circumstances)
# to terminate. Note that eval will run for max_train_steps each loop,
# regardless of the global_step count.
tf.compat.v1.logging.info('Starting to evaluate.')
eval_results = classifier.evaluate(input_fn=input_fn_eval,
steps=flags_obj.max_train_steps)
benchmark_logger.log_evaluation_result(eval_results)
if model_helpers.past_stop_threshold(flags_obj.stop_threshold,
eval_results['accuracy']):
break
if flags_obj.export_dir is not None:
# Exports a saved model for the given classifier.
export_dtype = flags_core.get_tf_dtype(flags_obj)
if flags_obj.image_bytes_as_serving_input:
input_receiver_fn = functools.partial(image_bytes_serving_input_fn,
shape,
dtype=export_dtype)
else:
input_receiver_fn = export.build_tensor_serving_input_receiver_fn(
shape, batch_size=flags_obj.batch_size, dtype=export_dtype)
classifier.export_savedmodel(flags_obj.export_dir,
input_receiver_fn,
strip_default_attrs=True)
stats = {}
stats['eval_results'] = eval_results
stats['train_hooks'] = train_hooks
return stats
def run_transformer(flags_obj):
"""Create tf.Estimator to train and evaluate transformer model.
Args:
flags_obj: Object containing parsed flag values.
"""
num_gpus = flags_core.get_num_gpus(flags_obj)
# Add flag-defined parameters to params object
params = PARAMS_MAP[flags_obj.param_set]
if num_gpus > 1:
if flags_obj.param_set == "big":
params = model_params.BIG_MULTI_GPU_PARAMS
elif flags_obj.param_set == "base":
params = model_params.BASE_MULTI_GPU_PARAMS
params["data_dir"] = flags_obj.data_dir
params["model_dir"] = flags_obj.model_dir
params["num_parallel_calls"] = flags_obj.num_parallel_calls
params["tpu"] = flags_obj.tpu
params["use_tpu"] = bool(flags_obj.tpu) # was a tpu specified.
params["static_batch"] = flags_obj.static_batch or params["use_tpu"]
params["allow_ffn_pad"] = not params["use_tpu"]
params["use_synthetic_data"] = flags_obj.use_synthetic_data
params["worker_hosts"] = flags_obj.worker_hosts
params["task_index"] = flags_obj.task_index
params["server_protocol"] = flags_obj.server_protocol
# Set batch size parameter, which depends on the availability of
# TPU and GPU, and distribution settings.
params["batch_size"] = (
flags_obj.batch_size
or (params["default_batch_size_tpu"]
if params["use_tpu"] else params["default_batch_size"]))
if not params["use_tpu"]:
params["batch_size"] = distribution_utils.per_device_batch_size(
params["batch_size"], num_gpus)
print("============== Batch Size for each GPU ==============")
print("Batch Size for each GPU", params["batch_size"])
print("============== Batch Size for each GPU ==============")
schedule_manager = schedule.Manager(
train_steps=flags_obj.train_steps,
steps_between_evals=flags_obj.steps_between_evals,
train_epochs=flags_obj.train_epochs,
epochs_between_evals=flags_obj.epochs_between_evals,
default_train_epochs=DEFAULT_TRAIN_EPOCHS,
batch_size=params["batch_size"],
max_length=params["max_length"],
use_tpu=params["use_tpu"],
num_tpu_shards=flags_obj.num_tpu_shards)
params["repeat_dataset"] = schedule_manager.repeat_dataset
model_helpers.apply_clean(flags.FLAGS)
print("============== Train Hooks ==============")
print(flags_obj.hooks)
print("============== Train Hooks ==============")
# Create hooks that log information about the training and metric values
train_hooks = hooks_helper.get_train_hooks(
flags_obj.hooks,
model_dir=flags_obj.model_dir,
save_steps=5000,
tensors_to_log=TENSORS_TO_LOG, # used for logging hooks
batch_size=schedule_manager.batch_size # for ExamplesPerSecondHook
)
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info(model_name="transformer",
dataset_name="wmt_translate_ende",
run_params=params,
test_id=flags_obj.benchmark_test_id)
# Train and evaluate transformer model
network = construct_network(num_gpus, flags_obj, params, schedule_manager)
run_loop(
network=network,
# Training arguments
schedule_manager=schedule_manager,
train_hooks=train_hooks,
benchmark_logger=benchmark_logger,
# BLEU calculation arguments
bleu_source=flags_obj.bleu_source,
bleu_ref=flags_obj.bleu_ref,
bleu_threshold=flags_obj.stop_threshold,
vocab_file=flags_obj.vocab_file)
if flags_obj.export_dir and not params["use_tpu"]:
serving_input_fn = export.build_tensor_serving_input_receiver_fn(
shape=[None], dtype=tf.int64, batch_size=None)
