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):
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 log_and_get_hooks(eval_batch_size):
"""Convenience function for hook and logger creation."""
# Create hooks that log information about the training and metric values
train_hooks = hooks_helper.get_train_hooks(
FLAGS.hooks,
model_dir=FLAGS.model_dir,
batch_size=FLAGS.batch_size, # for ExamplesPerSecondHook
tensors_to_log={"cross_entropy": "cross_entropy"})
run_params = {
"batch_size": FLAGS.batch_size,
"eval_batch_size": eval_batch_size,
"number_factors": FLAGS.num_factors,
"hr_threshold": FLAGS.hr_threshold,
"train_epochs": FLAGS.train_epochs,
}
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info(model_name="recommendation",
dataset_name=FLAGS.dataset,
run_params=run_params,
test_id=FLAGS.benchmark_test_id)
return benchmark_logger, train_hooks
def main(_):
"""Train NCF model and evaluate its hit rate (HR) metric."""
params = create_params()
if FLAGS.seed is not None:
np.random.seed(FLAGS.seed)
assert params["train_dataset_path"]
assert params["eval_dataset_path"]
assert 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"))
params["num_users"] = input_meta_data["num_users"]
params["num_items"] = input_meta_data["num_items"]
# In PER_HOST_V2 input mode, the number of input batches that TPUEstimator
# uses per step is equal to the number of TPU cores being trained on. More
# over when it initializes the dataset it will ask for the batch size to be
# the per core batchsize (i.e. global batch size / number of TPU cores).
num_train_steps = (int(input_meta_data["num_train_steps"]) //
FLAGS.num_tpu_shards)
num_eval_steps = (int(input_meta_data["num_eval_steps"]) //
FLAGS.num_tpu_shards)
def resize(x):
x["user_id"] = tf.squeeze(x["user_id"], axis=[-1])
x["item_id"] = tf.squeeze(x["item_id"], axis=[-1])
return x
def train_input_fn(params, index):
dataset = ncf_input_pipeline.create_dataset_from_tf_record_files(
params["train_dataset_path"].format(index),
input_meta_data["train_prebatch_size"],
params["batch_size"],
is_training=True)
return dataset.map(resize)
def eval_input_fn(params):
dataset = ncf_input_pipeline.create_dataset_from_tf_record_files(
params["eval_dataset_path"],
input_meta_data["eval_prebatch_size"],
params["batch_size"],
is_training=False)
return dataset.map(resize)
feature_columns = create_feature_columns(params)
model_fn = create_model_fn(feature_columns)
estimator = create_tpu_estimator(model_fn, feature_columns, params)
train_hooks = hooks_helper.get_train_hooks(
["ProfilerHook"],
model_dir=FLAGS.model_dir,
batch_size=FLAGS.batch_size, # for ExamplesPerSecondHook
tensors_to_log={"cross_entropy": "cross_entropy"})
for cycle_index in range(FLAGS.train_epochs):
tf.logging.info("Starting a training cycle: {}/{}".format(
cycle_index + 1, FLAGS.train_epochs))
# pylint: disable=cell-var-from-loop
estimator.train(
input_fn=lambda params: train_input_fn(params, cycle_index),
hooks=train_hooks,
steps=num_train_steps)
tf.logging.info("Beginning evaluation.")
eval_results = estimator.evaluate(eval_input_fn, steps=num_eval_steps)
tf.logging.info("Evaluation complete.")
hr = float(eval_results[rconst.HR_KEY])
ndcg = float(eval_results[rconst.NDCG_KEY])
loss = float(eval_results["loss"])
tf.logging.info(
"Iteration {}: HR = {:.4f}, NDCG = {:.4f}, Loss = {:.4f}".format(
cycle_index + 1, hr, ndcg, loss))
def run_deep_speech(_):
"""Run deep speech training and eval loop."""
tf.set_random_seed(flags_obj.seed)
# Data preprocessing
tf.logging.info("Data preprocessing...")
train_speech_dataset = generate_dataset(flags_obj.train_data_dir)
eval_speech_dataset = generate_dataset(flags_obj.eval_data_dir)
# Number of label classes. Label string is "[a-z]' -"
num_classes = len(train_speech_dataset.speech_labels)
# Use distribution strategy for multi-gpu training
num_gpus = flags_core.get_num_gpus(flags_obj)
distribution_strategy = distribution_utils.get_distribution_strategy(
num_gpus=num_gpus)
run_config = tf.estimator.RunConfig(train_distribute=distribution_strategy)
estimator = tf.estimator.Estimator(model_fn=model_fn,
model_dir=flags_obj.model_dir,
config=run_config,
params={
"num_classes": num_classes,
})
# Benchmark logging
run_params = {
"batch_size": flags_obj.batch_size,
"train_epochs": flags_obj.train_epochs,
"rnn_hidden_size": flags_obj.rnn_hidden_size,
"rnn_hidden_layers": flags_obj.rnn_hidden_layers,
"rnn_type": flags_obj.rnn_type,
"is_bidirectional": flags_obj.is_bidirectional,
"use_bias": flags_obj.use_bias
}
dataset_name = "LibriSpeech"
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info("deep_speech",
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)
per_replica_batch_size = per_device_batch_size(flags_obj.batch_size,
num_gpus)
def input_fn_train():
return dataset.input_fn(per_replica_batch_size, train_speech_dataset)
def input_fn_eval():
return dataset.input_fn(per_replica_batch_size, eval_speech_dataset)
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 + 1,
total_training_cycle)
# Perform batch_wise dataset shuffling
train_speech_dataset.entries = dataset.batch_wise_dataset_shuffle(
train_speech_dataset.entries, cycle_index, flags_obj.sortagrad,
flags_obj.batch_size)
estimator.train(input_fn=input_fn_train, hooks=train_hooks)
# Evaluation
tf.logging.info("Starting to evaluate...")
eval_results = evaluate_model(estimator,
eval_speech_dataset.speech_labels,
eval_speech_dataset.entries,
input_fn_eval)
# Log the WER and CER results.
benchmark_logger.log_evaluation_result(eval_results)
tf.logging.info("Iteration {}: WER = {:.2f}, CER = {:.2f}".format(
cycle_index + 1, eval_results[_WER_KEY], eval_results[_CER_KEY]))
# If some evaluation threshold is met
if model_helpers.past_stop_threshold(flags_obj.wer_threshold,
eval_results[_WER_KEY]):
break
def run_transformer(flags_obj):
"""Create tf.Estimator to train and evaluate transformer model.
Args:
flags_obj: Object containing parsed flag values.
Returns:
Dict of results of the run. Contains the keys `eval_results`,
`train_hooks`, `bleu_cased`, and `bleu_uncased`. `train_hooks` is a list the
instances of hooks used during training.
"""
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["max_length"] = flags_obj.max_length or params["max_length"]
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"]))
total_batch_size = params["batch_size"]
if not params["use_tpu"]:
params["batch_size"] = per_replica_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=total_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)
stats = 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)
return stats
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.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),
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.config.list_physical_devices('GPU')
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.compat.v1.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 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),
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=60 * 60 * 24,
save_checkpoints_steps=None)
# 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=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=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)
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.
dur_list = []
for cycle_index, num_train_epochs in enumerate(schedule):
logging.info('Starting cycle: %d/%d', cycle_index, int(n_loops))
stime = time.time()
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.
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)
dur = time.time() - stime
dur_list.append(dur)
logging.info('DUR FOR EPOCH = %d', dur)
if model_helpers.past_stop_threshold(flags_obj.stop_threshold,
eval_results['accuracy']):
break
for i in dur_list:
logging.info('Time_stat = %d', i)
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
