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))
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))
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 construct_estimator(flags_obj, params, schedule_manager):
"""Construct an estimator from either Estimator or TPUEstimator.
Args:
flags_obj: The FLAGS object parsed from command line.
params: A dict of run specific parameters.
schedule_manager: A schedule.Manager object containing the run schedule.
Returns:
An estimator object to be used for training and eval.
"""
if not params["use_tpu"]:
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)
return tf.estimator.Estimator(
model_fn=model_fn, model_dir=flags_obj.model_dir, params=params,
config=tf.estimator.RunConfig(train_distribute=distribution_strategy))
tpu_cluster_resolver = tf.contrib.cluster_resolver.TPUClusterResolver(
tpu=flags_obj.tpu,
zone=flags_obj.tpu_zone,
project=flags_obj.tpu_gcp_project
)
tpu_config = tf.contrib.tpu.TPUConfig(
iterations_per_loop=schedule_manager.single_iteration_train_steps,
num_shards=flags_obj.num_tpu_shards)
run_config = tf.contrib.tpu.RunConfig(
cluster=tpu_cluster_resolver,
model_dir=flags_obj.model_dir,
session_config=tf.ConfigProto(
allow_soft_placement=True, log_device_placement=True),
tpu_config=tpu_config)
return tf.contrib.tpu.TPUEstimator(
model_fn=model_fn,
use_tpu=params["use_tpu"] and flags_obj.tpu != tpu_util.LOCAL,
train_batch_size=schedule_manager.batch_size,
eval_batch_size=schedule_manager.batch_size,
params={
# TPUEstimator needs to populate batch_size itself due to sharding.
key: value for key, value in params.items() if key != "batch_size"},
config=run_config)
def parse_flags(flags_obj):
"""Convenience function to turn flags into params."""
num_gpus = flags_core.get_num_gpus(flags_obj)
num_devices = FLAGS.num_tpu_shards if FLAGS.tpu else num_gpus or 1
batch_size = (flags_obj.batch_size + num_devices - 1) // num_devices
eval_divisor = (rconst.NUM_EVAL_NEGATIVES + 1) * num_devices
eval_batch_size = flags_obj.eval_batch_size or flags_obj.batch_size
eval_batch_size = ((eval_batch_size + eval_divisor - 1) //
eval_divisor * eval_divisor // num_devices)
return {
"train_epochs": flags_obj.train_epochs,
"batches_per_step": num_devices,
"use_seed": flags_obj.seed is not None,
"batch_size": batch_size,
"eval_batch_size": eval_batch_size,
"learning_rate": flags_obj.learning_rate,
"mf_dim": flags_obj.num_factors,
"model_layers": [int(layer) for layer in flags_obj.layers],
"mf_regularization": flags_obj.mf_regularization,
"mlp_reg_layers": [float(reg) for reg in flags_obj.mlp_regularization],
"num_neg": flags_obj.num_neg,
"num_gpus": num_gpus,
"use_tpu": flags_obj.tpu is not None,
"tpu": flags_obj.tpu,
"tpu_zone": flags_obj.tpu_zone,
"tpu_gcp_project": flags_obj.tpu_gcp_project,
"beta1": flags_obj.beta1,
"beta2": flags_obj.beta2,
"epsilon": flags_obj.epsilon,
"match_mlperf": flags_obj.ml_perf,
"use_xla_for_gpu": flags_obj.use_xla_for_gpu,
"clone_model_in_keras_dist_strat":
flags_obj.clone_model_in_keras_dist_strat,
"epochs_between_evals": FLAGS.epochs_between_evals,
"turn_off_distribution_strategy": FLAGS.turn_off_distribution_strategy,
}
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_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,
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)
def invert(image, label):
return (image * -1.0) + 1.0, label
def brightness(image, label):
return tf.image.random_brightness(image, max_delta=0.2), label
if INVERT:
inverted = ds.map(invert)
ds = ds.concatenate(inverted)
if BRIGHTNESS:
ds = ds.concatenate(ds.map(brightness))
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,
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 our_test_fn():
images = []
for i in list(range(1,10)) + ['dog']:
images.append(np.array(imageio.imread('{}.png'.format(i)).ravel()/255.0, dtype='float32'))
images = np.array(images)
return tf.convert_to_tensor(images)
# Check our own examples
predictions = mnist_classifier.predict(input_fn=our_test_fn)
table = []
for i in list(range(1, 10)) + ['dog']:
prediction = next(predictions)
if i == 'dog':
print("{}. CNN thinks it's a {} ({:.1f}%)".format(i, prediction['classes'], prediction['probabilities'][prediction['classes']]*100))
else:
print("{} at {:.1f}. CNN thinks it's a {} ({:.1f}%)".format(i, prediction['probabilities'][i]*100, prediction['classes'], prediction['probabilities'][prediction['classes']]*100))
table.append((i, prediction['probabilities']))
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.
"""
# 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,
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 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] # 设置网络规模的种类，基础版还是高级版
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 # 什么叫使用合成数据？
# 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( # 用来管理训练进度的实例，例如steps，验证间隔步数等
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) # 清理一下数据和之前保存的模型，但是需要在参数中指定允许清理
# Create hooks that log information about the training and metric values
train_hooks = hooks_helper.get_train_hooks( # 好像是输出日志的时候需要这个实例
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() # 还是用来输出日志的
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) # 返回一个tf.estimator.Estimator用来训练和验证模型
run_loop(
estimator=estimator, # “估计器”，用来帮助训练和验证模型
# Training arguments
schedule_manager=schedule_manager, # 用来管理训练过程的，训练多少steps，多久验证一次等
train_hooks=train_hooks, # 打日志的？
benchmark_logger=benchmark_logger, # 打日志的？
# BLEU calculation arguments
bleu_source=flags_obj.bleu_source, # 3个关于bleu的文件
bleu_ref=flags_obj.bleu_ref,
bleu_threshold=flags_obj.stop_threshold,
vocab_file=flags_obj.vocab_file) # 词表文件
if flags_obj.export_dir:
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})
def input_fn_eval():
return input_function(
is_training=False, data_dir=flags_obj.data_dir,
batch_size=per_device_batch_size(
flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
num_epochs=1)
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)
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_device_batch_size = distribution_utils.per_device_batch_size(
flags_obj.batch_size, num_gpus)
def input_fn_train():
return dataset.input_fn(
per_device_batch_size, train_speech_dataset)
def input_fn_eval():
return dataset.input_fn(
per_device_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_ncf(_):
"""Run NCF training and eval loop."""
if FLAGS.download_if_missing and not FLAGS.use_synthetic_data:
movielens.download(FLAGS.dataset, FLAGS.data_dir)
if FLAGS.seed is not None:
np.random.seed(FLAGS.seed)
num_gpus = flags_core.get_num_gpus(FLAGS)
batch_size = distribution_utils.per_device_batch_size(
int(FLAGS.batch_size), num_gpus)
total_training_cycle = FLAGS.train_epochs // FLAGS.epochs_between_evals
eval_per_user = rconst.NUM_EVAL_NEGATIVES + 1
eval_batch_size = int(FLAGS.eval_batch_size or
max([FLAGS.batch_size, eval_per_user]))
if eval_batch_size % eval_per_user:
eval_batch_size = eval_batch_size // eval_per_user * eval_per_user
tf.logging.warning(
"eval examples per user does not evenly divide eval_batch_size. "
"Overriding to {}".format(eval_batch_size))
if FLAGS.use_synthetic_data:
ncf_dataset = None
cleanup_fn = lambda: None
num_users, num_items = data_preprocessing.DATASET_TO_NUM_USERS_AND_ITEMS[
FLAGS.dataset]
num_train_steps = data_preprocessing.SYNTHETIC_BATCHES_PER_EPOCH
num_eval_steps = data_preprocessing.SYNTHETIC_BATCHES_PER_EPOCH
else:
ncf_dataset, cleanup_fn = data_preprocessing.instantiate_pipeline(
dataset=FLAGS.dataset, data_dir=FLAGS.data_dir,
batch_size=batch_size,
eval_batch_size=eval_batch_size,
num_neg=FLAGS.num_neg,
epochs_per_cycle=FLAGS.epochs_between_evals,
num_cycles=total_training_cycle,
match_mlperf=FLAGS.ml_perf,
deterministic=FLAGS.seed is not None,
use_subprocess=FLAGS.use_subprocess,
cache_id=FLAGS.cache_id)
num_users = ncf_dataset.num_users
num_items = ncf_dataset.num_items
num_train_steps = int(np.ceil(
FLAGS.epochs_between_evals * ncf_dataset.num_train_positives *
(1 + FLAGS.num_neg) / FLAGS.batch_size))
num_eval_steps = int(np.ceil((1 + rconst.NUM_EVAL_NEGATIVES) *
ncf_dataset.num_users / eval_batch_size))
model_helpers.apply_clean(flags.FLAGS)
params = {
"use_seed": FLAGS.seed is not None,
"hash_pipeline": FLAGS.hash_pipeline,
"batch_size": batch_size,
"eval_batch_size": eval_batch_size,
"learning_rate": FLAGS.learning_rate,
"num_users": num_users,
"num_items": num_items,
"mf_dim": FLAGS.num_factors,
"model_layers": [int(layer) for layer in FLAGS.layers],
"mf_regularization": FLAGS.mf_regularization,
"mlp_reg_layers": [float(reg) for reg in FLAGS.mlp_regularization],
"num_neg": FLAGS.num_neg,
"use_tpu": FLAGS.tpu is not None,
"tpu": FLAGS.tpu,
"tpu_zone": FLAGS.tpu_zone,
"tpu_gcp_project": FLAGS.tpu_gcp_project,
"beta1": FLAGS.beta1,
"beta2": FLAGS.beta2,
"epsilon": FLAGS.epsilon,
"match_mlperf": FLAGS.ml_perf,
"use_xla_for_gpu": FLAGS.use_xla_for_gpu,
"use_estimator": FLAGS.use_estimator,
}
if FLAGS.use_estimator:
train_estimator, eval_estimator = construct_estimator(
num_gpus=num_gpus, model_dir=FLAGS.model_dir,
iterations=num_train_steps, params=params,
batch_size=flags.FLAGS.batch_size, eval_batch_size=eval_batch_size)
else:
runner = model_runner.NcfModelRunner(ncf_dataset, params, num_train_steps,
num_eval_steps, FLAGS.use_while_loop)
# 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)
eval_input_fn = None
target_reached = False
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.TRAIN_LOOP)
for cycle_index in range(total_training_cycle):
assert FLAGS.epochs_between_evals == 1 or not mlperf_helper.LOGGER.enabled
tf.logging.info("Starting a training cycle: {}/{}".format(
cycle_index + 1, total_training_cycle))
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.TRAIN_EPOCH,
value=cycle_index)
# Train the model
if FLAGS.use_estimator:
train_input_fn, train_record_dir, batch_count = \
data_preprocessing.make_input_fn(
ncf_dataset=ncf_dataset, is_training=True)
if batch_count != num_train_steps:
raise ValueError(
"Step counts do not match. ({} vs. {}) The async process is "
"producing incorrect shards.".format(batch_count, num_train_steps))
train_estimator.train(input_fn=train_input_fn, hooks=train_hooks,
steps=num_train_steps)
if train_record_dir:
tf.gfile.DeleteRecursively(train_record_dir)
tf.logging.info("Beginning evaluation.")
if eval_input_fn is None:
eval_input_fn, _, eval_batch_count = data_preprocessing.make_input_fn(
ncf_dataset=ncf_dataset, is_training=False)
if eval_batch_count != num_eval_steps:
raise ValueError(
"Step counts do not match. ({} vs. {}) The async process is "
"producing incorrect shards.".format(
eval_batch_count, num_eval_steps))
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_START,
value=cycle_index)
eval_results = eval_estimator.evaluate(eval_input_fn,
steps=num_eval_steps)
tf.logging.info("Evaluation complete.")
else:
runner.train()
tf.logging.info("Beginning evaluation.")
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_START,
value=cycle_index)
eval_results = runner.eval()
tf.logging.info("Evaluation complete.")
hr = float(eval_results[rconst.HR_KEY])
ndcg = float(eval_results[rconst.NDCG_KEY])
mlperf_helper.ncf_print(
key=mlperf_helper.TAGS.EVAL_TARGET,
value={"epoch": cycle_index, "value": FLAGS.hr_threshold})
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_ACCURACY,
value={"epoch": cycle_index, "value": hr})
mlperf_helper.ncf_print(
key=mlperf_helper.TAGS.EVAL_HP_NUM_NEG,
value={"epoch": cycle_index, "value": rconst.NUM_EVAL_NEGATIVES})
# Logged by the async process during record creation.
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_HP_NUM_USERS,
deferred=True)
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_STOP, value=cycle_index)
# Benchmark the evaluation results
benchmark_logger.log_evaluation_result(eval_results)
# Log the HR and NDCG results.
tf.logging.info(
"Iteration {}: HR = {:.4f}, NDCG = {:.4f}".format(
cycle_index + 1, hr, ndcg))
# If some evaluation threshold is met
if model_helpers.past_stop_threshold(FLAGS.hr_threshold, hr):
target_reached = True
break
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.RUN_STOP,
value={"success": target_reached})
cleanup_fn() # Cleanup data construction artifacts and subprocess.
# Clear the session explicitly to avoid session delete error
tf.keras.backend.clear_session()
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.RUN_FINAL)
def run_ncf(_):
"""Run NCF training and eval loop."""
if FLAGS.download_if_missing:
movielens.download(FLAGS.dataset, FLAGS.data_dir)
movielens_dataset.construct_train_eval_csv(data_dir=FLAGS.data_dir,
dataset=FLAGS.dataset)
tf.logging.info("Data preprocessing...")
ncf_dataset = movielens_dataset.data_preprocessing(FLAGS.data_dir,
FLAGS.dataset,
FLAGS.num_neg)
model_helpers.apply_clean(flags.FLAGS)
# Create NeuMF model and convert it to Estimator
tf.logging.info("Creating Estimator from Keras model...")
layers = [int(layer) for layer in FLAGS.layers]
mlp_regularization = [float(reg) for reg in FLAGS.mlp_regularization]
keras_model = neumf_model.NeuMF(ncf_dataset.num_users,
ncf_dataset.num_items, FLAGS.num_factors,
layers, FLAGS.batch_size,
FLAGS.mf_regularization,
mlp_regularization)
num_gpus = flags_core.get_num_gpus(FLAGS)
estimator = convert_keras_to_estimator(keras_model, num_gpus,
FLAGS.model_dir)
# 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
)
run_params = {
"batch_size": FLAGS.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)
# Training and evaluation cycle
def get_train_input_fn():
return movielens_dataset.get_input_fn(
True,
distribution_utils.per_device_batch_size(FLAGS.batch_size,
num_gpus), ncf_dataset,
FLAGS.data_dir, FLAGS.dataset, FLAGS.epochs_between_evals)
def get_pred_input_fn():
return movielens_dataset.get_input_fn(
False,
distribution_utils.per_device_batch_size(FLAGS.batch_size,
num_gpus), ncf_dataset,
FLAGS.data_dir, FLAGS.dataset, 1)
total_training_cycle = FLAGS.train_epochs // FLAGS.epochs_between_evals
for cycle_index in range(total_training_cycle):
tf.logging.info("Starting a training cycle: {}/{}".format(
cycle_index + 1, total_training_cycle))
# Train the model
estimator.train(input_fn=get_train_input_fn(), hooks=train_hooks)
# Evaluate the model
eval_results = evaluate_model(estimator, FLAGS.batch_size, num_gpus,
ncf_dataset, get_pred_input_fn())
# Benchmark the evaluation results
benchmark_logger.log_evaluation_result(eval_results)
# Log the HR and NDCG results.
hr = eval_results[_HR_KEY]
ndcg = eval_results[_NDCG_KEY]
tf.logging.info("Iteration {}: HR = {:.4f}, NDCG = {:.4f}".format(
cycle_index + 1, hr, ndcg))
# If some evaluation threshold is met
if model_helpers.past_stop_threshold(FLAGS.hr_threshold, hr):
break
# Clear the session explicitly to avoid session delete error
tf.keras.backend.clear_session()
def run_keras_model_benchmark(_):
"""Run the benchmark on keras model."""
# Ensure a valid model name was supplied via command line argument
if FLAGS.model not in MODELS.keys():
raise AssertionError("The --model command line argument should "
"be a key in the `MODELS` dictionary.")
# Check if eager execution is enabled
if FLAGS.eager:
tf.logging.info("Eager execution is enabled...")
tf.enable_eager_execution()
# Load the model
tf.logging.info("Benchmark on {} model...".format(FLAGS.model))
keras_model = MODELS[FLAGS.model]
# Get dataset
dataset_name = "ImageNet"
if FLAGS.use_synthetic_data:
tf.logging.info("Using synthetic dataset...")
dataset_name += "_Synthetic"
train_dataset = dataset.generate_synthetic_input_dataset(
FLAGS.model, FLAGS.batch_size)
val_dataset = dataset.generate_synthetic_input_dataset(
FLAGS.model, FLAGS.batch_size)
model = keras_model(weights=None)
else:
tf.logging.info("Using CIFAR-10 dataset...")
dataset_name = "CIFAR-10"
ds = dataset.Cifar10Dataset(FLAGS.batch_size)
train_dataset = ds.train_dataset
val_dataset = ds.test_dataset
model = keras_model(weights=None,
input_shape=ds.input_shape,
classes=ds.num_classes)
num_gpus = flags_core.get_num_gpus(FLAGS)
distribution = None
# Use distribution strategy
if FLAGS.dist_strat:
distribution = distribution_utils.get_distribution_strategy(
num_gpus=num_gpus)
elif num_gpus > 1:
# Run with multi_gpu_model
# If eager execution is enabled, only one GPU is utilized even if multiple
# GPUs are provided.
if FLAGS.eager:
tf.logging.warning(
"{} GPUs are provided, but only one GPU is utilized as "
"eager execution is enabled.".format(num_gpus))
model = tf.keras.utils.multi_gpu_model(model, gpus=num_gpus)
# Adam optimizer and some other optimizers doesn't work well with
# distribution strategy (b/113076709)
# Use GradientDescentOptimizer here
optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.001)
model.compile(loss="categorical_crossentropy",
optimizer=optimizer,
metrics=["accuracy"],
distribute=distribution)
# Create benchmark logger for benchmark logging
run_params = {
"batch_size": FLAGS.batch_size,
"synthetic_data": FLAGS.use_synthetic_data,
"train_epochs": FLAGS.train_epochs,
"num_train_images": FLAGS.num_train_images,
"num_eval_images": FLAGS.num_eval_images,
}
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info(model_name=FLAGS.model,
dataset_name=dataset_name,
run_params=run_params,
test_id=FLAGS.benchmark_test_id)
# Create callbacks that log metric values about the training and evaluation
callbacks = model_callbacks.get_model_callbacks(
FLAGS.callbacks,
batch_size=FLAGS.batch_size,
metric_logger=benchmark_logger)
# Train and evaluate the model
history = model.fit(train_dataset,
epochs=FLAGS.train_epochs,
callbacks=callbacks,
validation_data=val_dataset,
steps_per_epoch=int(
np.ceil(FLAGS.num_train_images /
FLAGS.batch_size)),
validation_steps=int(
np.ceil(FLAGS.num_eval_images / FLAGS.batch_size)))
tf.logging.info("Logging the evaluation results...")
for epoch in range(FLAGS.train_epochs):
eval_results = {
"accuracy":
history.history["val_acc"][epoch],
"loss":
history.history["val_loss"][epoch],
tf.GraphKeys.GLOBAL_STEP:
(epoch + 1) * np.ceil(FLAGS.num_eval_images / FLAGS.batch_size)
}
benchmark_logger.log_evaluation_result(eval_results)
# Clear the session explicitly to avoid session delete error
tf.keras.backend.clear_session()
def run_keras_model_benchmark(_):
"""Run the benchmark on keras model."""
# Ensure a valid model name was supplied via command line argument
if FLAGS.model not in MODELS.keys():
raise AssertionError("The --model command line argument should "
"be a key in the `MODELS` dictionary.")
# Load the model
tf.logging.info("Benchmark on {} model...".format(FLAGS.model))
keras_model = MODELS[FLAGS.model]
model = keras_model(weights=None)
# Get dataset
dataset_name = "ImageNet"
if FLAGS.use_synthetic_data:
tf.logging.info("Using synthetic dataset...")
dataset_name += "_Synthetic"
train_num_images = FLAGS.batch_size
val_num_images = FLAGS.batch_size
train_dataset = dataset.generate_synthetic_input_dataset(
FLAGS.model, train_num_images)
val_dataset = dataset.generate_synthetic_input_dataset(
FLAGS.model, val_num_images)
else:
raise ValueError("Only synthetic dataset is supported!")
# If run with multiple GPUs
num_gpus = flags_core.get_num_gpus(FLAGS)
if num_gpus > 0:
model = tf.keras.utils.multi_gpu_model(model, gpus=num_gpus)
# Configure the model
model.compile(loss="categorical_crossentropy",
optimizer="sgd",
metrics=["accuracy"])
# Create benchmark logger for benchmark logging
run_params = {
"batch_size": FLAGS.batch_size,
"synthetic_data": FLAGS.use_synthetic_data,
"train_epochs": FLAGS.train_epochs
}
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info(
model_name=FLAGS.model,
dataset_name=dataset_name,
run_params=run_params,
test_id=FLAGS.benchmark_test_id)
# Create callbacks that log metric values about the training and evaluation
callbacks = model_callbacks.get_model_callbacks(
FLAGS.callbacks,
batch_size=FLAGS.batch_size,
metric_logger=benchmark_logger)
# Train and evaluate the model
history = model.fit(
train_dataset,
epochs=FLAGS.train_epochs,
callbacks=callbacks,
validation_data=val_dataset,
steps_per_epoch=int(np.ceil(train_num_images / FLAGS.batch_size)),
validation_steps=int(np.ceil(val_num_images / FLAGS.batch_size))
)
tf.logging.info("Logging the evaluation results...")
for epoch in range(FLAGS.train_epochs):
eval_results = {
"accuracy": history.history["val_acc"][epoch],
"loss": history.history["val_loss"][epoch],
tf.GraphKeys.GLOBAL_STEP: (epoch + 1) * np.ceil(
train_num_images/FLAGS.batch_size)
}
benchmark_logger.log_evaluation_result(eval_results)
# Clear the session explicitly to avoid session delete error
tf.keras.backend.clear_session()
def run_ncf(_):
"""Run NCF training and eval loop."""
if FLAGS.download_if_missing:
movielens.download(FLAGS.dataset, FLAGS.data_dir)
if FLAGS.seed is not None:
np.random.seed(FLAGS.seed)
num_gpus = flags_core.get_num_gpus(FLAGS)
batch_size = distribution_utils.per_device_batch_size(
int(FLAGS.batch_size), num_gpus)
eval_per_user = rconst.NUM_EVAL_NEGATIVES + 1
eval_batch_size = int(FLAGS.eval_batch_size
or max([FLAGS.batch_size, eval_per_user]))
if eval_batch_size % eval_per_user:
eval_batch_size = eval_batch_size // eval_per_user * eval_per_user
tf.logging.warning(
"eval examples per user does not evenly divide eval_batch_size. "
"Overriding to {}".format(eval_batch_size))
ncf_dataset, cleanup_fn = data_preprocessing.instantiate_pipeline(
dataset=FLAGS.dataset,
data_dir=FLAGS.data_dir,
batch_size=batch_size,
eval_batch_size=eval_batch_size,
num_neg=FLAGS.num_neg,
epochs_per_cycle=FLAGS.epochs_between_evals,
match_mlperf=FLAGS.ml_perf,
deterministic=FLAGS.seed is not None)
model_helpers.apply_clean(flags.FLAGS)
train_estimator, eval_estimator = construct_estimator(
num_gpus=num_gpus,
model_dir=FLAGS.model_dir,
params={
"use_seed": FLAGS.seed is not None,
"hash_pipeline": FLAGS.hash_pipeline,
"batch_size": batch_size,
"learning_rate": FLAGS.learning_rate,
"num_users": ncf_dataset.num_users,
"num_items": ncf_dataset.num_items,
"mf_dim": FLAGS.num_factors,
"model_layers": [int(layer) for layer in FLAGS.layers],
"mf_regularization": FLAGS.mf_regularization,
"mlp_reg_layers": [float(reg) for reg in FLAGS.mlp_regularization],
"num_neg": FLAGS.num_neg,
"use_tpu": FLAGS.tpu is not None,
"tpu": FLAGS.tpu,
"tpu_zone": FLAGS.tpu_zone,
"tpu_gcp_project": FLAGS.tpu_gcp_project,
"beta1": FLAGS.beta1,
"beta2": FLAGS.beta2,
"epsilon": FLAGS.epsilon,
"match_mlperf": FLAGS.ml_perf,
},
batch_size=flags.FLAGS.batch_size,
eval_batch_size=eval_batch_size)
# 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)
approx_train_steps = int(ncf_dataset.num_train_positives *
(1 + FLAGS.num_neg) // FLAGS.batch_size)
pred_input_fn = data_preprocessing.make_pred_input_fn(
ncf_dataset=ncf_dataset)
total_training_cycle = FLAGS.train_epochs // FLAGS.epochs_between_evals
for cycle_index in range(total_training_cycle):
tf.logging.info("Starting a training cycle: {}/{}".format(
cycle_index + 1, total_training_cycle))
# Train the model
train_input_fn, train_record_dir, batch_count = \
data_preprocessing.make_train_input_fn(ncf_dataset=ncf_dataset)
if np.abs(approx_train_steps - batch_count) > 1:
tf.logging.warning(
"Estimated ({}) and reported ({}) number of batches differ by more "
"than one".format(approx_train_steps, batch_count))
train_estimator.train(input_fn=train_input_fn,
hooks=train_hooks,
steps=batch_count)
tf.gfile.DeleteRecursively(train_record_dir)
tf.logging.info("Beginning evaluation.")
eval_results = eval_estimator.evaluate(pred_input_fn)
tf.logging.info("Evaluation complete.")
# Benchmark the evaluation results
benchmark_logger.log_evaluation_result(eval_results)
# Log the HR and NDCG results.
hr = eval_results[rconst.HR_KEY]
ndcg = eval_results[rconst.NDCG_KEY]
tf.logging.info("Iteration {}: HR = {:.4f}, NDCG = {:.4f}".format(
cycle_index + 1, hr, ndcg))
# If some evaluation threshold is met
if model_helpers.past_stop_threshold(FLAGS.hr_threshold, hr):
break
cleanup_fn() # Cleanup data construction artifacts and subprocess.
# Clear the session explicitly to avoid session delete error
tf.keras.backend.clear_session()
def main(_):
# Data preprocessing
# The file name of training and test dataset
train_fname = os.path.join(
FLAGS.data_dir, FLAGS.dataset + "-" + constants.TRAIN_RATINGS_FILENAME)
test_fname = os.path.join(
FLAGS.data_dir, FLAGS.dataset + "-" + constants.TEST_RATINGS_FILENAME)
neg_fname = os.path.join(FLAGS.data_dir,
FLAGS.dataset + "-" + constants.TEST_NEG_FILENAME)
assert os.path.exists(train_fname), (
"Run data_download.py first to download and extract {} dataset".format(
FLAGS.dataset))
tf.logging.info("Data preprocessing...")
ncf_dataset = dataset.data_preprocessing(train_fname, test_fname,
neg_fname, FLAGS.num_neg)
# Create NeuMF model and convert it to Estimator
tf.logging.info("Creating Estimator from Keras model...")
layers = [int(layer) for layer in FLAGS.layers]
mlp_regularization = [float(reg) for reg in FLAGS.mlp_regularization]
keras_model = neumf_model.NeuMF(ncf_dataset.num_users,
ncf_dataset.num_items, FLAGS.num_factors,
layers, FLAGS.batch_size,
FLAGS.mf_regularization,
mlp_regularization)
num_gpus = flags_core.get_num_gpus(FLAGS)
estimator = convert_keras_to_estimator(keras_model, num_gpus,
FLAGS.model_dir)
# Create hooks that log information about the training and metric values
train_hooks = hooks_helper.get_train_hooks(
FLAGS.hooks,
batch_size=FLAGS.batch_size # for ExamplesPerSecondHook
)
run_params = {
"batch_size": FLAGS.batch_size,
"number_factors": FLAGS.num_factors,
"hr_threshold": FLAGS.hr_threshold,
"train_epochs": FLAGS.train_epochs,
}
benchmark_logger = logger.config_benchmark_logger(FLAGS)
benchmark_logger.log_run_info(model_name="recommendation",
dataset_name=FLAGS.dataset,
run_params=run_params)
# Training and evaluation cycle
def train_input_fn():
return dataset.input_fn(
True, per_device_batch_size(FLAGS.batch_size, num_gpus),
ncf_dataset, FLAGS.epochs_between_evals)
total_training_cycle = FLAGS.train_epochs // FLAGS.epochs_between_evals
for cycle_index in range(total_training_cycle):
tf.logging.info("Starting a training cycle: {}/{}".format(
cycle_index + 1, total_training_cycle))
# Train the model
estimator.train(input_fn=train_input_fn, hooks=train_hooks)
# Evaluate the model
eval_results = evaluate_model(estimator, FLAGS.batch_size, num_gpus,
ncf_dataset)
# Benchmark the evaluation results
benchmark_logger.log_evaluation_result(eval_results)
# Log the HR and NDCG results.
hr = eval_results[_HR_KEY]
ndcg = eval_results[_NDCG_KEY]
tf.logging.info("Iteration {}: HR = {:.4f}, NDCG = {:.4f}".format(
cycle_index + 1, hr, ndcg))
# If some evaluation threshold is met
if model_helpers.past_stop_threshold(FLAGS.hr_threshold, hr):
break
# Clear the session explicitly to avoid session delete error
tf.keras.backend.clear_session()
def run_deep_speech(_):
"""Run deep speech training and eval loop."""
# Data preprocessing
# The file name of training and test dataset
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)
# Input shape of each data example:
# [time_steps (T), feature_bins(F), channel(C)]
# Channel is set as 1 by default.
input_shape = (None, train_speech_dataset.num_feature_bins, 1)
# Create deep speech model and convert it to Estimator
tf.logging.info("Creating Estimator from Keras model...")
keras_model = deep_speech_model.DeepSpeech(
input_shape, flags_obj.rnn_hidden_layers, flags_obj.rnn_type,
flags_obj.is_bidirectional, flags_obj.rnn_hidden_size,
flags_obj.rnn_activation, num_classes, flags_obj.use_bias)
# Convert to estimator
num_gpus = flags_core.get_num_gpus(flags_obj)
estimator = convert_keras_to_estimator(keras_model, num_gpus)
# 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_activation": flags_obj.rnn_activation,
"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,
batch_size=flags_obj.batch_size)
per_device_batch_size = distribution_utils.per_device_batch_size(
flags_obj.batch_size, num_gpus)
def input_fn_train():
return dataset.input_fn(per_device_batch_size, train_speech_dataset)
def input_fn_eval(): # #pylint: disable=unused-variable
return dataset.input_fn(per_device_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)
estimator.train(input_fn=input_fn_train, hooks=train_hooks)
# Evaluate (TODO)
# tf.logging.info("Starting to evaluate.")
# eval_results = evaluate_model(
# estimator, keras_model, data_set.speech_labels, [], input_fn_eval)
# benchmark_logger.log_evaluation_result(eval_results)
# If some evaluation threshold is met
# Log the HR and NDCG results.
# wer = eval_results[_WER_KEY]
# cer = eval_results[_CER_KEY]
# tf.logging.info(
# "Iteration {}: WER = {:.2f}, CER = {:.2f}".format(
# cycle_index + 1, wer, cer))
# if model_helpers.past_stop_threshold(FLAGS.wer_threshold, wer):
# break
# Clear the session explicitly to avoid session delete error
tf.keras.backend.clear_session()
def model_fn(features, labels, mode, params):
"""Defines how to train, evaluate and predict from the transformer model."""
num_devices = flags_core.get_num_gpus(flags_obj)
consolidation_device = 'gpu:0'
# feature_shards, label_shards = replicate_model_fn._split_batch(features, labels, num_devices, device=consolidation_device)
tower_losses = []
tower_gradvars = []
tower_preds = []
for i in range(num_devices):
worker_device = '/{}:{}'.format('gpu', i)
device_setter = local_device_setter(
ps_device_type='gpu',
worker_device=worker_device,
ps_strategy=tf.contrib.training.GreedyLoadBalancingStrategy(
num_devices, tf.contrib.training.byte_size_load_fn))
with tf.variable_scope('model', reuse=bool(i != 0)):
with tf.name_scope('tower_%d' % i) as name_scope:
with tf.device(device_setter):
# Create model and get output logits.
model = transformer.Transformer(
params, mode == tf.estimator.ModeKeys.TRAIN)
#logits = model(features, labels)
loss, gradvars, preds = _tower_fn(model,
features,
labels,
params=params)
tower_losses.append(loss)
tower_gradvars.append(gradvars)
tower_preds.append(preds)
# Compute global loss and gradients
gradvars = []
with tf.name_scope('gradient_averaging'):
all_grads = {}
for grad, var in itertools.chain(*tower_gradvars):
if grad is not None:
all_grads.setdefault(var, []).append(grad)
for var, grads in six.iteritems(all_grads):
with tf.device(var.device):
if len(grads) == 1:
avg_grad = grads[0]
else:
# for a in range(len(grads)):
# if len(grads[a]) > 1:
# avg_grad = tf.multiply(tf.add_n(grads[a]), 1. / len(grads[a]))
# gradvars.append((avg_grad, var))
avg_grad = tf.multiply(tf.add_n(grads),
1. / len(grads))
# print("AVG_GRAD: ", avg_grad, "VAR: ", var)
gradvars.append((avg_grad, var))
with tf.device(consolidation_device):
loss = tf.reduce_mean(tower_losses, name='loss')
tf.identity(loss, "cross_entropy")
logits = tf.reduce_mean(tower_preds, axis=0)
# logits = tf.concat([l for l in tower_preds], axis=0)
if mode == tf.estimator.ModeKeys.PREDICT:
return tf.estimator.EstimatorSpec(
tf.estimator.ModeKeys.PREDICT,
predictions=logits,
export_outputs={
"translate": tf.estimator.export.PredictOutput(logits)
})
if mode == tf.estimator.ModeKeys.TRAIN:
with tf.variable_scope("get_train_op"):
print("in get_train_op")
learning_rate = get_learning_rate(
learning_rate=params["learning_rate"],
hidden_size=params["hidden_size"],
learning_rate_warmup_steps=params[
"learning_rate_warmup_steps"])
optimizer = tf.contrib.opt.LazyAdamOptimizer(
learning_rate,
beta1=params["optimizer_adam_beta1"],
beta2=params["optimizer_adam_beta2"],
epsilon=params["optimizer_adam_epsilon"])
optimizer = tf.train.SyncReplicasOptimizer(
optimizer, replicas_to_aggregate=num_devices)
sync_hook = optimizer.make_session_run_hook(is_chief)
# update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
global_step = tf.train.get_global_step()
update_ops = tf.assign(global_step,
global_step + 1,
name='update_global_step')
minimize_op = optimizer.apply_gradients(
gradvars, global_step=tf.train.get_global_step())
train_op = tf.group(minimize_op, update_ops)
#train_op = [optimizer.apply_gradients(gradvars, global_step=tf.train.get_global_step())]
metric_dict = {"learning_rate": learning_rate}
metric_dict["minibatch_loss"] = loss
record_scalars(metric_dict)
return tf.estimator.EstimatorSpec(
mode=mode,
loss=loss,
training_hooks=[sync_hook],
train_op=train_op)
elif mode == tf.estimator.ModeKeys.EVAL:
return tf.estimator.EstimatorSpec(
mode=mode,
loss=loss,
predictions={"predictions": logits},
eval_metric_ops=metrics.get_eval_metrics(
logits, labels, params))
def run(flags_obj):
"""Run ResNet ImageNet training and eval loop using native Keras APIs.
Args:
flags_obj: An object containing parsed flag values.
Raises:
ValueError: If fp16 is passed as it is not currently supported.
"""
if flags_obj.enable_eager:
tf.enable_eager_execution()
dtype = flags_core.get_tf_dtype(flags_obj)
if dtype == 'fp16':
raise ValueError(
'dtype fp16 is not supported in Keras. Use the default '
'value(fp32).')
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')
tf.keras.backend.set_image_data_format(data_format)
per_device_batch_size = distribution_utils.per_device_batch_size(
flags_obj.batch_size, flags_core.get_num_gpus(flags_obj))
# pylint: disable=protected-access
if flags_obj.use_synthetic_data:
input_fn = keras_common.get_synth_input_fn(
height=imagenet_main.DEFAULT_IMAGE_SIZE,
width=imagenet_main.DEFAULT_IMAGE_SIZE,
num_channels=imagenet_main.NUM_CHANNELS,
num_classes=imagenet_main.NUM_CLASSES,
dtype=flags_core.get_tf_dtype(flags_obj))
else:
input_fn = imagenet_main.input_fn
train_input_dataset = input_fn(is_training=True,
data_dir=flags_obj.data_dir,
batch_size=per_device_batch_size,
num_epochs=flags_obj.train_epochs,
parse_record_fn=parse_record_keras)
eval_input_dataset = input_fn(is_training=False,
data_dir=flags_obj.data_dir,
batch_size=per_device_batch_size,
num_epochs=flags_obj.train_epochs,
parse_record_fn=parse_record_keras)
strategy = distribution_utils.get_distribution_strategy(
flags_obj.num_gpus, flags_obj.turn_off_distribution_strategy)
strategy_scope = keras_common.get_strategy_scope(strategy)
with strategy_scope:
optimizer = keras_common.get_optimizer()
model = resnet_model.resnet50(num_classes=imagenet_main.NUM_CLASSES)
model.compile(loss='sparse_categorical_crossentropy',
optimizer=optimizer,
metrics=['sparse_categorical_accuracy'])
time_callback, tensorboard_callback, lr_callback = keras_common.get_callbacks(
learning_rate_schedule, imagenet_main.NUM_IMAGES['train'])
train_steps = imagenet_main.NUM_IMAGES['train'] // flags_obj.batch_size
train_epochs = flags_obj.train_epochs
if flags_obj.train_steps:
train_steps = min(flags_obj.train_steps, train_steps)
train_epochs = 1
num_eval_steps = (imagenet_main.NUM_IMAGES['validation'] //
flags_obj.batch_size)
validation_data = eval_input_dataset
if flags_obj.skip_eval:
# Only build the training graph. This reduces memory usage introduced by
# control flow ops in layers that have different implementations for
# training and inference (e.g., batch norm).
tf.keras.backend.set_learning_phase(1)
num_eval_steps = None
validation_data = None
history = model.fit(
train_input_dataset,
epochs=train_epochs,
steps_per_epoch=train_steps,
callbacks=[time_callback, lr_callback, tensorboard_callback],
validation_steps=num_eval_steps,
validation_data=validation_data,
verbose=1)
eval_output = None
if not flags_obj.skip_eval:
eval_output = model.evaluate(eval_input_dataset,
steps=num_eval_steps,
verbose=1)
stats = keras_common.build_stats(history, eval_output, time_callback)
return stats
def run(flags_obj):
"""Run ResNet Cifar-10 training and eval loop using native Keras APIs.
Args:
flags_obj: An object containing parsed flag values.
Raises:
ValueError: If fp16 is passed as it is not currently supported.
Returns:
Dictionary of training and eval stats.
"""
if flags_obj.enable_eager:
tf.enable_eager_execution()
dtype = flags_core.get_tf_dtype(flags_obj)
if dtype == 'fp16':
raise ValueError(
'dtype fp16 is not supported in Keras. Use the default '
'value(fp32).')
per_device_batch_size = distribution_utils.per_device_batch_size(
flags_obj.batch_size, flags_core.get_num_gpus(flags_obj))
if flags_obj.use_synthetic_data:
input_fn = keras_common.get_synth_input_fn(
height=cifar_main.HEIGHT,
width=cifar_main.WIDTH,
num_channels=cifar_main.NUM_CHANNELS,
num_classes=cifar_main.NUM_CLASSES,
dtype=flags_core.get_tf_dtype(flags_obj))
else:
input_fn = cifar_main.input_fn
train_input_dataset = input_fn(is_training=True,
data_dir=flags_obj.data_dir,
batch_size=per_device_batch_size,
num_epochs=flags_obj.train_epochs,
parse_record_fn=parse_record_keras)
eval_input_dataset = input_fn(is_training=False,
data_dir=flags_obj.data_dir,
batch_size=per_device_batch_size,
num_epochs=flags_obj.train_epochs,
parse_record_fn=parse_record_keras)
optimizer = keras_common.get_optimizer()
strategy = distribution_utils.get_distribution_strategy(
flags_obj.num_gpus, flags_obj.turn_off_distribution_strategy)
model = resnet_cifar_model.resnet56(classes=cifar_main.NUM_CLASSES)
model.compile(loss='categorical_crossentropy',
optimizer=optimizer,
metrics=['categorical_accuracy'],
distribute=strategy)
time_callback, tensorboard_callback, lr_callback = keras_common.get_callbacks(
learning_rate_schedule, cifar_main.NUM_IMAGES['train'])
train_steps = cifar_main.NUM_IMAGES['train'] // flags_obj.batch_size
train_epochs = flags_obj.train_epochs
if flags_obj.train_steps:
train_steps = min(flags_obj.train_steps, train_steps)
train_epochs = 1
num_eval_steps = (cifar_main.NUM_IMAGES['validation'] //
flags_obj.batch_size)
validation_data = eval_input_dataset
if flags_obj.skip_eval:
num_eval_steps = None
validation_data = None
history = model.fit(
train_input_dataset,
epochs=train_epochs,
steps_per_epoch=train_steps,
callbacks=[time_callback, lr_callback, tensorboard_callback],
validation_steps=num_eval_steps,
validation_data=validation_data,
verbose=1)
eval_output = None
if not flags_obj.skip_eval:
eval_output = model.evaluate(eval_input_dataset,
steps=num_eval_steps,
verbose=1)
stats = keras_common.build_stats(history, eval_output)
return stats
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.
"""
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:
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)
train_hooks = list(train_hooks) + lottery.hooks_from_flags(
flags_obj.flag_values_dict())
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.lth_generate_predictions:
ckpt = tf.train.latest_checkpoint(flags_obj.model_dir)
if flags_obj.lth_no_pruning:
m_hooks = []
else:
m_hooks = lottery.hooks_from_flags(flags_obj.flag_values_dict())
eval_results = classifier.predict(
input_fn=input_fn_eval,
checkpoint_path=ckpt,
hooks=m_hooks,
)
assert flags_obj.lth_prediction_result_dir
with tf.gfile.Open(os.path.join(flags_obj.data_dir, 'test_batch.bin'),
'rb') as f:
labels = list(f.read()[::32 * 32 * 3 + 1])
eval_results = list(eval_results)
if not tf.gfile.Exists(flags_obj.lth_prediction_result_dir):
tf.gfile.MakeDirs(flags_obj.lth_prediction_result_dir)
with tf.gfile.Open(
os.path.join(flags_obj.lth_prediction_result_dir,
'predictions'), 'wb') as f:
for label, res in zip(labels, eval_results):
res['label'] = label
pickle.dump(eval_results, f)
return
try:
cpr = tf.train.NewCheckpointReader(
tf.train.latest_checkpoint(flags_obj.model_dir))
current_step = cpr.get_tensor('global_step')
except:
current_step = 0
while current_step < flags_obj.max_train_steps:
next_checkpoint = min(current_step + 10000, flags_obj.max_train_steps)
classifier.train(input_fn=lambda: input_fn_train(1000),
hooks=train_hooks,
max_steps=next_checkpoint)
current_step = next_checkpoint
tf.logging.info('Starting to evaluate.')
eval_results = classifier.evaluate(input_fn=input_fn_eval)
benchmark_logger.log_evaluation_result(eval_results)
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_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]
### yr
params['vocab_size_in']=6100
params['vocab_size_out'] = 25
#params['vocab_size']='s'
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)
mode = tf.estimator.ModeKeys.TRAIN
## build ph
inputs_ph = tf.placeholder(tf.int32, shape=(None, None), name='inputs')
targets_ph = tf.placeholder(tf.int32, shape=(None, None), name='targets')
targets_ph_2 = tf.placeholder(tf.int32, shape=(None, None), name='targets2')
loss_M, train_op_M = model_fn(inputs_ph, targets_ph, targets_ph_2, mode, params)
print('Using GPU in Decoder')
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.9)
sess = tf.Session(
config=tf.ConfigProto(allow_soft_placement=True, log_device_placement=False, gpu_options=gpu_options))
with sess.as_default():
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
if os.path.exists(os.path.join(FLAGS.model_dir, "checkpoint")):
saver.restore(sess, tf.train.latest_checkpoint(FLAGS.model_dir))
cnt=0
for ii in xrange(schedule_manager.train_eval_iterations):
tf.logging.info("Starting iteration %d" % (ii + 1))
## get data
random.shuffle(datall)
for data in datall:
feed = {inputs_ph:data['inputs'],
targets_ph: data['targets'],
targets_ph_2:data['targets2']}
loss, train_op = sess.run([loss_M, train_op_M], feed_dict=feed)
cnt+=1
print loss
##
if cnt%100==0:
print 'loss at %d'%cnt,loss
if cnt%2000==0:
filename = os.path.join(
FLAGS.model_dir, "model_{}.ckpt".format(cnt))
saver.save(sess, filename)
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
# added below to override the learning rate, warmup steps, max_length, vocab_size in parameter file.
# params would now have data passed as a flag
params["learning_rate"] = flags_obj.learning_rate
params["learning_rate_warmup_steps"] = flags_obj.learning_rate_warmup_steps
params["max_length"] = flags_obj.max_length
params["vocab_size"] = flags_obj.vocab_size
# added for selecting the learning rate scheme
params["lr_scheme"] = flags_obj.lr_scheme
params["warmup_init_lr"] = flags_obj.warmup_init_lr
# added for selecting the optimizer algorithm
params["opt_alg"] = flags_obj.opt_alg
# added to provide optimizer parameters
params["optimizer_sgd_momentum"] = flags_obj.optimizer_sgd_momentum
params["optimizer_rms_decay"] = flags_obj.optimizer_rms_decay
params["optimizer_rms_momentum"] = flags_obj.optimizer_rms_momentum
params["optimizer_rms_epsilon"] = flags_obj.optimizer_rms_epsilon
# added to overide layer_postprocess_dropout value
params["layer_postprocess_dropout"] = flags_obj.layer_postprocess_dropout
# 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"]))
# commented below to remove distribution strategy
"""
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
)
# added for horovod
bcast_hook = hvd.BroadcastGlobalVariablesHook(0)
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)
# added for horovod
train_hooks.append(bcast_hook)
# 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})
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.
"""
# 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)
if flags_core.get_num_gpus(flags_obj) == 0:
distribution = tf.contrib.distribute.OneDeviceStrategy('device:CPU:0')
elif flags_core.get_num_gpus(flags_obj) == 1:
distribution = tf.contrib.distribute.OneDeviceStrategy('device:GPU:0')
else:
distribution = tf.contrib.distribute.MirroredStrategy(
num_gpus=flags_core.get_num_gpus(flags_obj)
)
run_config = tf.estimator.RunConfig(train_distribute=distribution,
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,
}
benchmark_logger = logger.config_benchmark_logger(flags_obj)
benchmark_logger.log_run_info('resnet', dataset_name, run_params)
train_hooks = hooks_helper.get_train_hooks(
flags_obj.hooks,
batch_size=flags_obj.batch_size)
def input_fn_train():
return input_function(
is_training=True, data_dir=flags_obj.data_dir,
batch_size=per_device_batch_size(
flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
num_epochs=flags_obj.epochs_between_evals)
def input_fn_eval():
return input_function(
is_training=False, data_dir=flags_obj.data_dir,
batch_size=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 run_ncf(_):
"""Run NCF training and eval loop."""
if FLAGS.download_if_missing:
movielens.download(FLAGS.dataset, FLAGS.data_dir)
num_gpus = flags_core.get_num_gpus(FLAGS)
batch_size = distribution_utils.per_device_batch_size(
int(FLAGS.batch_size), num_gpus)
eval_batch_size = int(FLAGS.eval_batch_size or FLAGS.batch_size)
ncf_dataset = data_preprocessing.instantiate_pipeline(
dataset=FLAGS.dataset,
data_dir=FLAGS.data_dir,
batch_size=batch_size,
eval_batch_size=eval_batch_size,
num_neg=FLAGS.num_neg,
epochs_per_cycle=FLAGS.epochs_between_evals,
match_mlperf=FLAGS.ml_perf)
model_helpers.apply_clean(flags.FLAGS)
train_estimator, eval_estimator = construct_estimator(
num_gpus=num_gpus,
model_dir=FLAGS.model_dir,
params={
"batch_size": batch_size,
"learning_rate": FLAGS.learning_rate,
"num_users": ncf_dataset.num_users,
"num_items": ncf_dataset.num_items,
"mf_dim": FLAGS.num_factors,
"model_layers": [int(layer) for layer in FLAGS.layers],
"mf_regularization": FLAGS.mf_regularization,
"mlp_reg_layers": [float(reg) for reg in FLAGS.mlp_regularization],
"use_tpu": FLAGS.tpu is not None,
"tpu": FLAGS.tpu,
"tpu_zone": FLAGS.tpu_zone,
"tpu_gcp_project": FLAGS.tpu_gcp_project,
},
batch_size=flags.FLAGS.batch_size,
eval_batch_size=eval_batch_size)
# 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
)
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)
approx_train_steps = int(ncf_dataset.num_train_positives *
(1 + FLAGS.num_neg) // FLAGS.batch_size)
pred_input_fn = data_preprocessing.make_pred_input_fn(
ncf_dataset=ncf_dataset)
total_training_cycle = 1 if FLAGS.inference_only else FLAGS.train_epochs // FLAGS.epochs_between_evals
for cycle_index in range(total_training_cycle):
tf.logging.info("Starting a training cycle: {}/{}".format(
cycle_index + 1, total_training_cycle))
if not FLAGS.inference_only:
# Train the model
train_input_fn, train_record_dir, batch_count = \
data_preprocessing.make_train_input_fn(ncf_dataset=ncf_dataset)
if np.abs(approx_train_steps - batch_count) > 1:
tf.logging.warning(
"Estimated ({}) and reported ({}) number of batches differ by more "
"than one".format(approx_train_steps, batch_count))
train_estimator.train(input_fn=train_input_fn,
hooks=train_hooks,
steps=batch_count)
tf.gfile.DeleteRecursively(train_record_dir)
# Evaluate the model
eval_results = evaluate_model(eval_estimator, ncf_dataset,
pred_input_fn)
# Benchmark the evaluation results
benchmark_logger.log_evaluation_result(eval_results)
# Log the HR and NDCG results.
hr = eval_results[_HR_KEY]
ndcg = eval_results[_NDCG_KEY]
tf.logging.fatal("Iteration {}: HR = {:.4f}, NDCG = {:.4f}".format(
cycle_index + 1, hr, ndcg))
# Export SavedModel
if FLAGS.export_savedmodel:
eval_estimator.export_savedmodel(FLAGS.model_dir,
serving_input_receiver_fn)
print("SavedModel successfully exported to: {}/<timestamp>".format(
FLAGS.model_dir))
# Some of the NumPy vector math can be quite large and likes to stay in
# memory for a while.
gc.collect()
# If some evaluation threshold is met
if model_helpers.past_stop_threshold(FLAGS.hr_threshold, hr):
break
# Clear the session explicitly to avoid session delete error
tf.keras.backend.clear_session()
def run_keras_model_benchmark(_):
"""Run the benchmark on keras model."""
# Ensure a valid model name was supplied via command line argument
if FLAGS.model not in MODELS.keys():
raise AssertionError("The --model command line argument should "
"be a key in the `MODELS` dictionary.")
# Check if eager execution is enabled
if FLAGS.eager:
tf.logging.info("Eager execution is enabled...")
tf.enable_eager_execution()
# Load the model
tf.logging.info("Benchmark on {} model...".format(FLAGS.model))
keras_model = MODELS[FLAGS.model]
model = keras_model(weights=None)
# Get dataset
dataset_name = "ImageNet"
if FLAGS.use_synthetic_data:
tf.logging.info("Using synthetic dataset...")
dataset_name += "_Synthetic"
train_dataset = dataset.generate_synthetic_input_dataset(
FLAGS.model, FLAGS.batch_size)
val_dataset = dataset.generate_synthetic_input_dataset(
FLAGS.model, FLAGS.batch_size)
else:
raise ValueError("Only synthetic dataset is supported!")
num_gpus = flags_core.get_num_gpus(FLAGS)
distribution = None
# Use distribution strategy
if FLAGS.dist_strat:
distribution = distribution_utils.get_distribution_strategy(
num_gpus=num_gpus)
elif num_gpus > 1:
# Run with multi_gpu_model
# If eager execution is enabled, only one GPU is utilized even if multiple
# GPUs are provided.
if FLAGS.eager:
tf.logging.warning(
"{} GPUs are provided, but only one GPU is utilized as "
"eager execution is enabled.".format(num_gpus))
model = tf.keras.utils.multi_gpu_model(model, gpus=num_gpus)
# Adam optimizer and some other optimizers doesn't work well with
# distribution strategy (b/113076709)
# Use GradientDescentOptimizer here
optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.001)
model.compile(loss="categorical_crossentropy",
optimizer=optimizer,
metrics=["accuracy"],
distribute=distribution)
# Create benchmark logger for benchmark logging
run_params = {
"batch_size": FLAGS.batch_size,
"synthetic_data": FLAGS.use_synthetic_data,
"train_epochs": FLAGS.train_epochs,
"num_train_images": FLAGS.num_train_images,
"num_eval_images": FLAGS.num_eval_images,
}
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info(
model_name=FLAGS.model,
dataset_name=dataset_name,
run_params=run_params,
test_id=FLAGS.benchmark_test_id)
# Create callbacks that log metric values about the training and evaluation
callbacks = model_callbacks.get_model_callbacks(
FLAGS.callbacks,
batch_size=FLAGS.batch_size,
metric_logger=benchmark_logger)
# Train and evaluate the model
history = model.fit(
train_dataset,
epochs=FLAGS.train_epochs,
callbacks=callbacks,
validation_data=val_dataset,
steps_per_epoch=int(np.ceil(FLAGS.num_train_images / FLAGS.batch_size)),
validation_steps=int(np.ceil(FLAGS.num_eval_images / FLAGS.batch_size))
)
tf.logging.info("Logging the evaluation results...")
for epoch in range(FLAGS.train_epochs):
eval_results = {
"accuracy": history.history["val_acc"][epoch],
"loss": history.history["val_loss"][epoch],
tf.GraphKeys.GLOBAL_STEP: (epoch + 1) * np.ceil(
FLAGS.num_eval_images/FLAGS.batch_size)
}
benchmark_logger.log_evaluation_result(eval_results)
# Clear the session explicitly to avoid session delete error
tf.keras.backend.clear_session()
def __init__(self, flags_obj):
"""Init function
Args:
flags_obj: Object containing parsed flag values, i.e., FLAGS.
"""
self.flags_obj = flags_obj
# Add flag-defined parameters to params object
num_gpus = flags_core.get_num_gpus(flags_obj)
self.params = params = misc.get_model_params(flags_obj.param_set,
num_gpus)
params["num_gpus"] = num_gpus
params["data_dir"] = flags_obj.data_dir
params["val_data_dir"] = flags_obj.val_data_dir
params["model_dir"] = flags_obj.model_dir
params["static_batch"] = flags_obj.static_batch
params["max_input_length"] = flags_obj.max_input_length
params["max_target_length"] = flags_obj.max_target_length
params["decode_batch_size"] = flags_obj.decode_batch_size
params["decode_max_length"] = flags_obj.decode_max_length
params["padded_decode"] = flags_obj.padded_decode
params["num_parallel_calls"] = (flags_obj.num_parallel_calls
or tf.data.experimental.AUTOTUNE)
params["use_synthetic_data"] = flags_obj.use_synthetic_data
params["batch_size"] = flags_obj.batch_size * max(num_gpus, 1)
logging.info('actual batch_size = {} * {}'.format(
flags_obj.batch_size, max(num_gpus, 1)))
params["repeat_dataset"] = None
params["dtype"] = flags_core.get_tf_dtype(flags_obj)
params[
"enable_metrics_in_training"] = flags_obj.enable_metrics_in_training
params["num_hashes"] = flags_obj.num_hashes
params["test_num_hashes"] = flags_obj.test_num_hashes
params[
"use_full_attention_in_reformer"] = flags_obj.use_full_attention_in_reformer
params["bucket_size"] = flags_obj.bucket_size
if flags_obj.one_dropout is not None:
params['layer_postprocess_dropout'] = flags_obj.one_dropout
params['attention_dropout'] = flags_obj.one_dropout
params['relu_dropout'] = flags_obj.one_dropout
if flags_obj.attention_dropout is not None:
params['attention_dropout'] = flags_obj.attention_dropout
params['lsh_attention_dropout'] = params['attention_dropout'] if params[
"use_full_attention_in_reformer"] else flags_obj.lsh_attention_dropout
logging.info(
f'dropouts (postprocess, attention, lsh_attention, relu) = {[params[k] for k in ["layer_postprocess_dropout", "attention_dropout", "lsh_attention_dropout", "relu_dropout"]]}'
)
logging.info(
f'attention_padding_strategy = {flags_obj.attention_padding_strategy}'
)
assert self.flags_obj.vocab_file, 'vocab file is None'
self.tokenizer = tfds.features.text.SubwordTextEncoder.load_from_file(
self.flags_obj.vocab_file)
self.EOS_id = self.tokenizer.encode('<EOS>')[0]
params["vocab_size"] = self.tokenizer.vocab_size
logging.info(
'loaded vocab from {}, vocab_size={} and EOS_id={}'.format(
self.flags_obj.vocab_file, self.tokenizer.vocab_size,
self.EOS_id))
logging.info(f'training_schema = [{self.flags_obj.training_schema}]')
if params["dtype"] == tf.float16:
# TODO(reedwm): It's pretty ugly to set the global policy in a constructor
# like this. What if multiple instances of Seq2SeqTask are created?
# We should have a better way in the tf.keras.mixed_precision API of doing
# this.
loss_scale = flags_core.get_loss_scale(flags_obj,
default_for_fp16="dynamic")
policy = tf.compat.v2.keras.mixed_precision.experimental.Policy(
"mixed_float16", loss_scale=loss_scale)
tf.compat.v2.keras.mixed_precision.experimental.set_policy(policy)
self.distribution_strategy = distribution_utils.get_distribution_strategy(
distribution_strategy=flags_obj.distribution_strategy,
num_gpus=num_gpus,
tpu_address=flags_obj.tpu or "")
logging.info("Running dtitle model with num_gpus = %d", num_gpus)
if self.distribution_strategy:
logging.info("For training, using distribution strategy: %s",
self.distribution_strategy)
else:
logging.info("Not using any distribution strategy.")
def resnet_main(flags_obj,
model_function,
input_function,
dataset_name,
percent,
model_class,
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=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
params = {
'resnet_size': int(flags_obj.resnet_size),
'data_format': 'channels_last',
'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,
'adv_train': False,
'attack': False,
}
classifier = tf.compat.v1.estimator.Estimator(
model_fn=model_function,
model_dir=flags_obj.model_dir,
config=run_config,
warm_start_from=warm_start_settings,
params=params)
params['adv_train'] = True
classifier_adv = tf.compat.v1.estimator.Estimator(
model_fn=model_function,
model_dir=flags_obj.model_dir,
config=run_config,
warm_start_from=warm_start_settings,
params=params)
params['adv_train'] = False
params['attack'] = True
classifier_attack = tf.compat.v1.estimator.Estimator(
model_fn=model_function,
model_dir=flags_obj.model_dir,
config=run_config,
warm_start_from=warm_start_settings,
params=params)
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,
percent=percent,
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,
percent=0,
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))
def input_fn_eval_attack():
return input_function(
is_training=False,
percent=100,
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)
# tf.compat.v1.logging.info(tf.global_variables())
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
if flags_obj.adv_train:
classifier_adv.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)
else:
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 clean.')
eval_results = classifier.evaluate(input_fn=input_fn_eval,
steps=flags_obj.max_train_steps)
tf.compat.v1.logging.info('Starting to evaluate adv.')
eval_results_adv = classifier_adv.evaluate(
input_fn=input_fn_eval, steps=flags_obj.max_train_steps)
tf.compat.v1.logging.info('Starting to evaluate attack.')
eval_results_attack = classifier_attack.evaluate(
input_fn=input_fn_eval_attack, steps=flags_obj.max_train_steps)
print(
'########################## clean #############################'
)
benchmark_logger.log_evaluation_result(eval_results)
print(
'########################## adv #############################')
benchmark_logger.log_evaluation_result(eval_results_adv)
print(
'########################## attack #############################'
)
benchmark_logger.log_evaluation_result(eval_results_attack)
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['eval_atttack_results'] = eval_results_attack
stats['eval_adv_results'] = eval_results_adv
stats['train_hooks'] = train_hooks
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
# 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 __init__(self, flags_obj):
"""Init function of TransformerMain.
Args:
flags_obj: Object containing parsed flag values, i.e., FLAGS.
Raises:
ValueError: if not using static batch for input data on TPU.
"""
self.flags_obj = flags_obj
self.predict_model = None
# Add flag-defined parameters to params object
num_gpus = flags_core.get_num_gpus(flags_obj)
self.params = params = misc.get_model_params(flags_obj.param_set,
num_gpus)
params["num_gpus"] = num_gpus
params["use_ctl"] = flags_obj.use_ctl
params["data_dir"] = flags_obj.data_dir
params["model_dir"] = flags_obj.model_dir
params["static_batch"] = flags_obj.static_batch
params["max_length"] = flags_obj.max_length
params["decode_batch_size"] = flags_obj.decode_batch_size
params["decode_max_length"] = flags_obj.decode_max_length
params["padded_decode"] = flags_obj.padded_decode
params["num_parallel_calls"] = (flags_obj.num_parallel_calls
or tf.data.experimental.AUTOTUNE)
params["use_synthetic_data"] = flags_obj.use_synthetic_data
params["batch_size"] = flags_obj.batch_size or params[
"default_batch_size"]
params["repeat_dataset"] = None
params["dtype"] = flags_core.get_tf_dtype(flags_obj)
params["enable_tensorboard"] = flags_obj.enable_tensorboard
params[
"enable_metrics_in_training"] = flags_obj.enable_metrics_in_training
params["steps_between_evals"] = flags_obj.steps_between_evals
params["enable_checkpointing"] = flags_obj.enable_checkpointing
self.distribution_strategy = distribution_utils.get_distribution_strategy(
distribution_strategy=flags_obj.distribution_strategy,
num_gpus=num_gpus,
all_reduce_alg=flags_obj.all_reduce_alg,
num_packs=flags_obj.num_packs,
tpu_address=flags_obj.tpu or "")
if self.use_tpu:
params[
"num_replicas"] = self.distribution_strategy.num_replicas_in_sync
if not params["static_batch"]:
raise ValueError("TPU requires static batch for input data.")
else:
logging.info("Running transformer with num_gpus = %d", num_gpus)
if self.distribution_strategy:
logging.info("For training, using distribution strategy: %s",
self.distribution_strategy)
else:
logging.info("Not using any distribution strategy.")
performance.set_mixed_precision_policy(
params["dtype"],
flags_core.get_loss_scale(flags_obj, default_for_fp16="dynamic"))
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_ncf(_):
"""Run NCF training and eval loop."""
if FLAGS.download_if_missing:
movielens.download(FLAGS.dataset, FLAGS.data_dir)
num_gpus = flags_core.get_num_gpus(FLAGS)
batch_size = distribution_utils.per_device_batch_size(
int(FLAGS.batch_size), num_gpus)
eval_batch_size = int(FLAGS.eval_batch_size or FLAGS.batch_size)
ncf_dataset = data_preprocessing.instantiate_pipeline(
dataset=FLAGS.dataset, data_dir=FLAGS.data_dir,
batch_size=batch_size,
eval_batch_size=eval_batch_size,
num_neg=FLAGS.num_neg,
epochs_per_cycle=FLAGS.epochs_between_evals,
match_mlperf=FLAGS.ml_perf)
model_helpers.apply_clean(flags.FLAGS)
train_estimator, eval_estimator = construct_estimator(
num_gpus=num_gpus, model_dir=FLAGS.model_dir, params={
"batch_size": batch_size,
"learning_rate": FLAGS.learning_rate,
"num_users": ncf_dataset.num_users,
"num_items": ncf_dataset.num_items,
"mf_dim": FLAGS.num_factors,
"model_layers": [int(layer) for layer in FLAGS.layers],
"mf_regularization": FLAGS.mf_regularization,
"mlp_reg_layers": [float(reg) for reg in FLAGS.mlp_regularization],
"use_tpu": FLAGS.tpu is not None,
"tpu": FLAGS.tpu,
"tpu_zone": FLAGS.tpu_zone,
"tpu_gcp_project": FLAGS.tpu_gcp_project,
}, batch_size=flags.FLAGS.batch_size, eval_batch_size=eval_batch_size)
# 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
)
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)
approx_train_steps = int(ncf_dataset.num_train_positives
* (1 + FLAGS.num_neg) // FLAGS.batch_size)
pred_input_fn = data_preprocessing.make_pred_input_fn(ncf_dataset=ncf_dataset)
total_training_cycle = FLAGS.train_epochs // FLAGS.epochs_between_evals
for cycle_index in range(total_training_cycle):
tf.logging.info("Starting a training cycle: {}/{}".format(
cycle_index + 1, total_training_cycle))
# Train the model
train_input_fn, train_record_dir, batch_count = \
data_preprocessing.make_train_input_fn(ncf_dataset=ncf_dataset)
if np.abs(approx_train_steps - batch_count) > 1:
tf.logging.warning(
"Estimated ({}) and reported ({}) number of batches differ by more "
"than one".format(approx_train_steps, batch_count))
train_estimator.train(input_fn=train_input_fn, hooks=train_hooks,
steps=batch_count)
tf.gfile.DeleteRecursively(train_record_dir)
# Evaluate the model
eval_results = evaluate_model(
eval_estimator, ncf_dataset, pred_input_fn)
# Benchmark the evaluation results
benchmark_logger.log_evaluation_result(eval_results)
# Log the HR and NDCG results.
hr = eval_results[_HR_KEY]
ndcg = eval_results[_NDCG_KEY]
tf.logging.info(
"Iteration {}: HR = {:.4f}, NDCG = {:.4f}".format(
cycle_index + 1, hr, ndcg))
# Some of the NumPy vector math can be quite large and likes to stay in
# memory for a while.
gc.collect()
# If some evaluation threshold is met
if model_helpers.past_stop_threshold(FLAGS.hr_threshold, hr):
break
# Clear the session explicitly to avoid session delete error
tf.keras.backend.clear_session()
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"] = distribution_utils.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
# 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 run_transformer(flags_obj):
print("run_transformer")
"""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["static_batch"] = flags_obj.static_batch
params["allow_ffn_pad"] = True
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"])
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=False,
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=False # 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)
os.environ['TF_SYNC_ON_FINISH'] = '0'
os.environ['TF_ENABLE_WINOGRAD_NONFUSED'] = '1'
sess_config = tf.ConfigProto(
allow_soft_placement=True,
log_device_placement=False,
intra_op_parallelism_threads=0,
gpu_options=tf.GPUOptions(force_gpu_compatible=True))
print("SESS_CONFIG: ", sess_config)
config = RunConfig(session_config=sess_config,
model_dir=params["model_dir"])
variable_strategy = 'GPU'
use_distortion_for_training = True
experiment_fn = get_experiment_fn(config.is_chief, flags_obj, params,
schedule_manager, num_gpus,
variable_strategy,
use_distortion_for_training)
#tf.contrib.learn.learn_runner.run(experiment_fn, run_config=config, hparams=tf.contrib.training.HParams(is_chief=config.is_chief, **hparams))
tf.contrib.learn.learn_runner.run(experiment_fn,
run_config=config,
hparams=tf.contrib.training.HParams(
is_chief=config.is_chief, **params))
'''
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:
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_ncf(_):
"""Run NCF training and eval loop."""
if FLAGS.download_if_missing and not FLAGS.use_synthetic_data:
movielens.download(FLAGS.dataset, FLAGS.data_dir)
if FLAGS.seed is not None:
np.random.seed(FLAGS.seed)
num_gpus = flags_core.get_num_gpus(FLAGS)
batch_size = distribution_utils.per_device_batch_size(
int(FLAGS.batch_size), num_gpus)
eval_per_user = rconst.NUM_EVAL_NEGATIVES + 1
eval_batch_size = int(FLAGS.eval_batch_size
or max([FLAGS.batch_size, eval_per_user]))
if eval_batch_size % eval_per_user:
eval_batch_size = eval_batch_size // eval_per_user * eval_per_user
tf.logging.warning(
"eval examples per user does not evenly divide eval_batch_size. "
"Overriding to {}".format(eval_batch_size))
if FLAGS.use_synthetic_data:
ncf_dataset = None
cleanup_fn = lambda: None
num_users, num_items = data_preprocessing.DATASET_TO_NUM_USERS_AND_ITEMS[
FLAGS.dataset]
num_train_steps = data_preprocessing.SYNTHETIC_BATCHES_PER_EPOCH
num_eval_steps = data_preprocessing.SYNTHETIC_BATCHES_PER_EPOCH
else:
ncf_dataset, cleanup_fn = data_preprocessing.instantiate_pipeline(
dataset=FLAGS.dataset,
data_dir=FLAGS.data_dir,
batch_size=batch_size,
eval_batch_size=eval_batch_size,
num_neg=FLAGS.num_neg,
epochs_per_cycle=FLAGS.epochs_between_evals,
match_mlperf=FLAGS.ml_perf,
deterministic=FLAGS.seed is not None,
use_subprocess=FLAGS.use_subprocess,
cache_id=FLAGS.cache_id)
num_users = ncf_dataset.num_users
num_items = ncf_dataset.num_items
num_train_steps = int(
np.ceil(FLAGS.epochs_between_evals *
ncf_dataset.num_train_positives * (1 + FLAGS.num_neg) /
FLAGS.batch_size))
num_eval_steps = int(
np.ceil((1 + rconst.NUM_EVAL_NEGATIVES) * ncf_dataset.num_users /
eval_batch_size))
model_helpers.apply_clean(flags.FLAGS)
train_estimator, eval_estimator = construct_estimator(
num_gpus=num_gpus,
model_dir=FLAGS.model_dir,
params={
"use_seed": FLAGS.seed is not None,
"hash_pipeline": FLAGS.hash_pipeline,
"batch_size": batch_size,
"eval_batch_size": eval_batch_size,
"learning_rate": FLAGS.learning_rate,
"num_users": num_users,
"num_items": num_items,
"mf_dim": FLAGS.num_factors,
"model_layers": [int(layer) for layer in FLAGS.layers],
"mf_regularization": FLAGS.mf_regularization,
"mlp_reg_layers": [float(reg) for reg in FLAGS.mlp_regularization],
"num_neg": FLAGS.num_neg,
"use_tpu": FLAGS.tpu is not None,
"tpu": FLAGS.tpu,
"tpu_zone": FLAGS.tpu_zone,
"tpu_gcp_project": FLAGS.tpu_gcp_project,
"beta1": FLAGS.beta1,
"beta2": FLAGS.beta2,
"epsilon": FLAGS.epsilon,
"match_mlperf": FLAGS.ml_perf,
"use_xla_for_gpu": FLAGS.use_xla_for_gpu,
},
batch_size=flags.FLAGS.batch_size,
eval_batch_size=eval_batch_size)
# 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)
pred_input_fn = None
total_training_cycle = FLAGS.train_epochs // FLAGS.epochs_between_evals
target_reached = False
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.TRAIN_LOOP)
for cycle_index in range(total_training_cycle):
assert FLAGS.epochs_between_evals == 1 or not mlperf_helper.LOGGER.enabled
tf.logging.info("Starting a training cycle: {}/{}".format(
cycle_index + 1, total_training_cycle))
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.TRAIN_EPOCH,
value=cycle_index)
# Train the model
train_input_fn, train_record_dir, batch_count = \
data_preprocessing.make_input_fn(
ncf_dataset=ncf_dataset, is_training=True)
if batch_count != num_train_steps:
raise ValueError(
"Step counts do not match. ({} vs. {}) The async process is "
"producing incorrect shards.".format(batch_count,
num_train_steps))
train_estimator.train(input_fn=train_input_fn,
hooks=train_hooks,
steps=num_train_steps)
if train_record_dir:
tf.gfile.DeleteRecursively(train_record_dir)
tf.logging.info("Beginning evaluation.")
if pred_input_fn is None:
pred_input_fn, _, eval_batch_count = data_preprocessing.make_input_fn(
ncf_dataset=ncf_dataset, is_training=False)
if eval_batch_count != num_eval_steps:
raise ValueError(
"Step counts do not match. ({} vs. {}) The async process is "
"producing incorrect shards.".format(
eval_batch_count, num_eval_steps))
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_START,
value=cycle_index)
eval_results = eval_estimator.evaluate(pred_input_fn,
steps=num_eval_steps)
hr = float(eval_results[rconst.HR_KEY])
ndcg = float(eval_results[rconst.NDCG_KEY])
tf.logging.info("Evaluation complete.")
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_TARGET,
value={
"epoch": cycle_index,
"value": FLAGS.hr_threshold
})
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_ACCURACY,
value={
"epoch": cycle_index,
"value": hr
})
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_HP_NUM_NEG,
value={
"epoch": cycle_index,
"value": rconst.NUM_EVAL_NEGATIVES
})
# Logged by the async process during record creation.
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_HP_NUM_USERS,
deferred=True)
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_STOP,
value=cycle_index)
# Benchmark the evaluation results
benchmark_logger.log_evaluation_result(eval_results)
# Log the HR and NDCG results.
tf.logging.info("Iteration {}: HR = {:.4f}, NDCG = {:.4f}".format(
cycle_index + 1, hr, ndcg))
# If some evaluation threshold is met
if model_helpers.past_stop_threshold(FLAGS.hr_threshold, hr):
target_reached = True
break
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.RUN_STOP,
value={"success": target_reached})
cleanup_fn() # Cleanup data construction artifacts and subprocess.
# Clear the session explicitly to avoid session delete error
tf.keras.backend.clear_session()
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.RUN_FINAL)
def __init__(self, flags_obj):
"""Init function of TransformerMain.
Args:
flags_obj: Object containing parsed flag values, i.e., FLAGS.
Raises:
ValueError: if not using static batch for input data on TPU.
"""
self.flags_obj = flags_obj
self.predict_model = None
# Add flag-defined parameters to params object
num_gpus = flags_core.get_num_gpus(flags_obj)
self.params = params = misc.get_model_params(flags_obj.param_set, num_gpus)
params["num_gpus"] = num_gpus
params["use_ctl"] = flags_obj.use_ctl
params["data_dir"] = flags_obj.data_dir
params["model_dir"] = flags_obj.model_dir
params["static_batch"] = flags_obj.static_batch
params["max_length"] = flags_obj.max_length
params["decode_batch_size"] = flags_obj.decode_batch_size
params["decode_max_length"] = flags_obj.decode_max_length
params["padded_decode"] = flags_obj.padded_decode
params["num_parallel_calls"] = (
flags_obj.num_parallel_calls or tf.data.experimental.AUTOTUNE)
params["use_synthetic_data"] = flags_obj.use_synthetic_data
params["batch_size"] = flags_obj.batch_size or params["default_batch_size"]
params["repeat_dataset"] = None
params["dtype"] = flags_core.get_tf_dtype(flags_obj)
params["enable_metrics_in_training"] = flags_obj.enable_metrics_in_training
if params["dtype"] == tf.float16:
# TODO(reedwm): It's pretty ugly to set the global policy in a constructor
# like this. What if multiple instances of TransformerTask are created?
# We should have a better way in the tf.keras.mixed_precision API of doing
# this.
loss_scale = flags_core.get_loss_scale(flags_obj,
default_for_fp16="dynamic")
policy = tf.compat.v2.keras.mixed_precision.experimental.Policy(
"mixed_float16", loss_scale=loss_scale)
tf.compat.v2.keras.mixed_precision.experimental.set_policy(policy)
if params["dtype"] == tf.bfloat16:
policy = tf.compat.v2.keras.mixed_precision.experimental.Policy(
"mixed_bfloat16")
tf.compat.v2.keras.mixed_precision.experimental.set_policy(policy)
self.distribution_strategy = distribution_utils.get_distribution_strategy(
distribution_strategy=flags_obj.distribution_strategy,
num_gpus=num_gpus,
tpu_address=flags_obj.tpu or "")
if self.use_tpu:
params["num_replicas"] = self.distribution_strategy.num_replicas_in_sync
if not params["static_batch"]:
raise ValueError("TPU requires static batch for input data.")
else:
logging.info("Running transformer with num_gpus = %d", num_gpus)
if self.distribution_strategy:
logging.info("For training, using distribution strategy: %s",
self.distribution_strategy)
else:
logging.info("Not using any distribution strategy.")
def run_keras_model_benchmark(_):
new_job_thread = threading.Thread(target=receive,
args=(
FLAGS.server_address.split(':')[0],
FLAGS.port,
),
daemon=True)
new_job_thread.start()
"""Run the benchmark on keras model."""
# Ensure a valid model name was supplied via command line argument
if FLAGS.model not in MODELS.keys():
raise AssertionError("The --model command line argument should "
"be a key in the `MODELS` dictionary.")
# print(FLAGS.gpus_list)
# exit()
# Check if eager execution is enabled
if FLAGS.eager:
tf.logging.info("Eager execution is enabled...")
tf.enable_eager_execution()
# Load the model
tf.logging.info("Benchmark on {} model...".format(FLAGS.model))
keras_model = MODELS[FLAGS.model]
model = keras_model(weights=None)
# Get dataset
dataset_name = "ImageNet"
if FLAGS.use_synthetic_data:
tf.logging.info("Using synthetic dataset...")
dataset_name += "_Synthetic"
train_dataset = dataset.generate_synthetic_input_dataset(
FLAGS.model, FLAGS.batch_size)
val_dataset = dataset.generate_synthetic_input_dataset(
FLAGS.model, FLAGS.batch_size)
else:
raise ValueError("Only synthetic dataset is supported!")
num_gpus = flags_core.get_num_gpus(FLAGS)
distribution = None
# Use distribution strategy
if FLAGS.dist_strat:
distribution = distribution_utils.get_distribution_strategy(
num_gpus=num_gpus)
elif num_gpus > 1:
# Run with multi_gpu_model
# If eager execution is enabled, only one GPU is utilized even if multiple
# GPUs are provided.
if FLAGS.eager:
tf.logging.warning(
"{} GPUs are provided, but only one GPU is utilized as "
"eager execution is enabled.".format(num_gpus))
model = tf.keras.utils.multi_gpu_model(model, gpus=num_gpus)
# Adam optimizer and some other optimizers doesn't work well with
# distribution strategy (b/113076709)
# Use GradientDescentOptimizer here
optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.001)
model.compile(loss="categorical_crossentropy",
optimizer=optimizer,
metrics=["accuracy"],
distribute=distribution)
# Create benchmark logger for benchmark logging
run_params = {
"batch_size": FLAGS.batch_size,
"synthetic_data": FLAGS.use_synthetic_data,
"train_epochs": FLAGS.train_epochs,
"num_train_images": FLAGS.num_train_images,
"num_eval_images": FLAGS.num_eval_images,
}
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info(model_name=FLAGS.model,
dataset_name=dataset_name,
run_params=run_params,
test_id=FLAGS.benchmark_test_id)
class LossHistory(tf.keras.callbacks.Callback):
def __init__(self):
self.start = time.time()
def on_train_begin(self, logs={}):
return
def on_epoch_end(self, epoch, logs={}):
global training_flags, have_trained
if job_status == 'g':
training_flags = 1
have_trained = epoch + 1
self.model.stop_training = True
if job_status == 's':
training_flags = 1
have_trained = epoch + 1
self.model.stop_training = True
def on_batch_end(self, batch, logs={}):
global lock
if batch == 49 and lock is True:
hundred = time.time() - self.start
# calculate the speed and unlock job
msg = {}
msg['id'] = FLAGS.id
msg['status'] = 'un'
msg['ep_tm'] = FLAGS.num_train_images * hundred / (
FLAGS.batch_size * 50)
send_msg(FLAGS.server_address, msg)
lock = False
# Create callbacks that log metric values about the training and evaluation
callbacks = model_callbacks.get_model_callbacks(
FLAGS.callbacks,
batch_size=FLAGS.batch_size,
metric_logger=benchmark_logger)
callbacks.append(LossHistory())
# Train and evaluate the model
history = model.fit(
train_dataset,
epochs=FLAGS.train_epochs,
callbacks=callbacks,
validation_data=val_dataset,
steps_per_epoch=int(np.ceil(FLAGS.num_train_images /
FLAGS.batch_size)),
)
''' No need for evaluation part
tf.logging.info("Logging the evaluation results...")
for epoch in range(FLAGS.train_epochs):
eval_results = {
"accuracy": history.history["val_acc"][epoch],
"loss": history.history["val_loss"][epoch],
tf.GraphKeys.GLOBAL_STEP: (epoch + 1) * np.ceil(
FLAGS.num_eval_images/FLAGS.batch_size)
}
benchmark_logger.log_evaluation_result(eval_results)
'''
# Clear the session explicitly to avoid session delete error
tf.keras.backend.clear_session()
# Now end the training send back message
msg = {}
remain_ep = FLAGS.train_epochs - have_trained
if training_flags == 0 or remain_ep == 0:
msg['status'] = 'e'
msg['id'] = FLAGS.id
# send_msg(FLAGS.server_address, msg)
else:
# ask the scheduler to re-run
# growing is needed
gpus_loc = {}
flags_gpu_list = [int(i) for i in FLAGS.gpus_list]
if job_status == 'g':
new_gpus_list = gpus + flags_gpu_list
msg['status'] = 'g'
else:
new_gpus_list = list(set(flags_gpu_list).difference(set(gpus)))
msg['status'] = 's'
# TODO hardcoded here
gpus_loc['localhost'] = new_gpus_list
msg['gpus_loc'] = gpus_loc
msg['id'] = FLAGS.id
msg['ep'] = FLAGS.train_epochs - have_trained
# send_msg(FLAGS.server_address, msg)
global exit_code
exit_code = True
time.sleep(1)
send_msg(FLAGS.server_address, msg)
print('exit')
exit()
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:
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,
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.
with tf.Session() as sess:
run_metadata = tf.RunMetadata()
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:
sess.run(classifier.train(
input_fn=lambda: input_fn_train(num_train_epochs),
hooks=train_hooks,
max_steps=flags_obj.max_train_steps),
run_metadata=run_metadata)
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 = sess.run(classifier.evaluate(
input_fn=input_fn_eval, steps=flags_obj.max_train_steps),
run_metadata=run_metadata)
benchmark_logger.log_evaluation_result(eval_results)
if model_helpers.past_stop_threshold(flags_obj.stop_threshold,
eval_results['accuracy']):
break
fetched_timeline = timeline.Timeline(run_metadata.step_stats)
chrome_trace = fetched_timeline.generate_chrome_trace_format()
with open('timeline_run_loop.json', 'w') as f:
f.write(chrome_trace)
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 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.
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 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)
'''
#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 = 30
#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)
tpu_cluster_resolver = tf.contrib.cluster_resolver.TPUClusterResolver(
flags_obj.tpu,
zone=flags_obj.tpu_zone,
project=flags_obj.gcp_project
)
run_config = tf.contrib.tpu.RunConfig(
cluster=tpu_cluster_resolver,
model_dir=flags_obj.model_dir,
session_config=tf.ConfigProto(
allow_soft_placement=True, log_device_placement=True),
tpu_config=tf.contrib.tpu.TPUConfig(flags_obj.iterations,flags_obj.num_shards),
)
#run_config = tf.estimator.RunConfig(train_distribute=distribution_strategy)
estimator = tf.contrib.tpu.TPUEstimator(
model_fn=model_fn,
model_dir=flags_obj.model_dir,
use_tpu=flags_obj.use_tpu,
train_batch_size=flags_obj.batch_size,
eval_batch_size=flags_obj.batch_size,
params={"num_classes": num_classes,
},
config=run_config)
# 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 = "Tuda Data"
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_device_batch_size = distribution_utils.per_device_batch_size(
flags_obj.batch_size, num_gpus
)
#TODO
def input_fn_train(params):
#ds = dataset.input_fn(per_device_batch_size, train_speech_dataset)
ds = test.input_fn(per_device_batch_size,'/content/records_test.csv')
return ds
def input_fn_eval(params):
return test.input_fn(params['batch_size'], eval_speech_dataset)
#def input_fn_predict(features, batch_size):
#dataset = tf.data.Dataset.from_tensor_slices(features)
#dataset = dataset.batch(batch_size)
#return dataset
# return dataset.input_fn(per_device_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
'''
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)
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_ncf(_):
"""Run NCF training and eval loop."""
# Data preprocessing
# The file name of training and test dataset
train_fname = os.path.join(
FLAGS.data_dir, FLAGS.dataset + "-" + constants.TRAIN_RATINGS_FILENAME)
test_fname = os.path.join(
FLAGS.data_dir, FLAGS.dataset + "-" + constants.TEST_RATINGS_FILENAME)
neg_fname = os.path.join(
FLAGS.data_dir, FLAGS.dataset + "-" + constants.TEST_NEG_FILENAME)
assert os.path.exists(train_fname), (
"Run data_download.py first to download and extract {} dataset".format(
FLAGS.dataset))
tf.logging.info("Data preprocessing...")
ncf_dataset = dataset.data_preprocessing(
train_fname, test_fname, neg_fname, FLAGS.num_neg)
# Create NeuMF model and convert it to Estimator
tf.logging.info("Creating Estimator from Keras model...")
layers = [int(layer) for layer in FLAGS.layers]
mlp_regularization = [float(reg) for reg in FLAGS.mlp_regularization]
keras_model = neumf_model.NeuMF(
ncf_dataset.num_users, ncf_dataset.num_items, FLAGS.num_factors,
layers, FLAGS.batch_size, FLAGS.mf_regularization,
mlp_regularization)
num_gpus = flags_core.get_num_gpus(FLAGS)
estimator = convert_keras_to_estimator(keras_model, num_gpus, FLAGS.model_dir)
# Create hooks that log information about the training and metric values
train_hooks = hooks_helper.get_train_hooks(
FLAGS.hooks,
batch_size=FLAGS.batch_size # for ExamplesPerSecondHook
)
run_params = {
"batch_size": FLAGS.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)
# Training and evaluation cycle
def train_input_fn():
return dataset.input_fn(
True,
distribution_utils.per_device_batch_size(FLAGS.batch_size, num_gpus),
ncf_dataset, FLAGS.epochs_between_evals)
total_training_cycle = FLAGS.train_epochs // FLAGS.epochs_between_evals
for cycle_index in range(total_training_cycle):
tf.logging.info("Starting a training cycle: {}/{}".format(
cycle_index + 1, total_training_cycle))
# Train the model
estimator.train(input_fn=train_input_fn, hooks=train_hooks)
# Evaluate the model
eval_results = evaluate_model(
estimator, FLAGS.batch_size, num_gpus, ncf_dataset)
# Benchmark the evaluation results
benchmark_logger.log_evaluation_result(eval_results)
# Log the HR and NDCG results.
hr = eval_results[_HR_KEY]
ndcg = eval_results[_NDCG_KEY]
tf.logging.info(
"Iteration {}: HR = {:.4f}, NDCG = {:.4f}".format(
cycle_index + 1, hr, ndcg))
# If some evaluation threshold is met
if model_helpers.past_stop_threshold(FLAGS.hr_threshold, hr):
break
# Clear the session explicitly to avoid session delete error
tf.keras.backend.clear_session()
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.
"""
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:
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,
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.
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)
return eval_results
def input_fn_eval():
return input_function(
is_training=False, data_dir=flags_obj.data_dir,
batch_size=per_device_batch_size(
flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
num_epochs=1)
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)
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_device_batch_size = distribution_utils.per_device_batch_size(
flags_obj.batch_size, num_gpus)
def input_fn_train():
return dataset.input_fn(per_device_batch_size, train_speech_dataset)
def input_fn_eval():
return dataset.input_fn(per_device_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_ncf(_):
"""Run NCF training and eval loop."""
if FLAGS.download_if_missing:
movielens.download(FLAGS.dataset, FLAGS.data_dir)
movielens_dataset.construct_train_eval_csv(
data_dir=FLAGS.data_dir, dataset=FLAGS.dataset)
tf.logging.info("Data preprocessing...")
ncf_dataset = movielens_dataset.data_preprocessing(
FLAGS.data_dir, FLAGS.dataset, FLAGS.num_neg)
model_helpers.apply_clean(flags.FLAGS)
# Create NeuMF model and convert it to Estimator
tf.logging.info("Creating Estimator from Keras model...")
layers = [int(layer) for layer in FLAGS.layers]
mlp_regularization = [float(reg) for reg in FLAGS.mlp_regularization]
keras_model = neumf_model.NeuMF(
ncf_dataset.num_users, ncf_dataset.num_items, FLAGS.num_factors,
layers, FLAGS.batch_size, FLAGS.mf_regularization,
mlp_regularization)
num_gpus = flags_core.get_num_gpus(FLAGS)
estimator = convert_keras_to_estimator(keras_model, num_gpus, FLAGS.model_dir)
# 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
)
run_params = {
"batch_size": FLAGS.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)
# Training and evaluation cycle
def get_train_input_fn():
return movielens_dataset.get_input_fn(
True,
distribution_utils.per_device_batch_size(FLAGS.batch_size, num_gpus),
ncf_dataset, FLAGS.data_dir, FLAGS.dataset, FLAGS.epochs_between_evals)
def get_pred_input_fn():
return movielens_dataset.get_input_fn(
False,
distribution_utils.per_device_batch_size(FLAGS.batch_size, num_gpus),
ncf_dataset, FLAGS.data_dir, FLAGS.dataset, 1)
total_training_cycle = FLAGS.train_epochs // FLAGS.epochs_between_evals
for cycle_index in range(total_training_cycle):
tf.logging.info("Starting a training cycle: {}/{}".format(
cycle_index + 1, total_training_cycle))
# Train the model
estimator.train(input_fn=get_train_input_fn(), hooks=train_hooks)
# Evaluate the model
eval_results = evaluate_model(
estimator, FLAGS.batch_size, num_gpus, ncf_dataset, get_pred_input_fn())
# Benchmark the evaluation results
benchmark_logger.log_evaluation_result(eval_results)
# Log the HR and NDCG results.
hr = eval_results[_HR_KEY]
ndcg = eval_results[_NDCG_KEY]
tf.logging.info(
"Iteration {}: HR = {:.4f}, NDCG = {:.4f}".format(
cycle_index + 1, hr, ndcg))
# If some evaluation threshold is met
if model_helpers.past_stop_threshold(FLAGS.hr_threshold, hr):
break
# Clear the session explicitly to avoid session delete error
tf.keras.backend.clear_session()
