def run_wide_deep(flags_obj):
"""Run Wide-Deep training and eval loop.
Args:
flags_obj: An object containing parsed flag values.
"""
# Clean up the model directory if present
shutil.rmtree(flags_obj.model_dir, ignore_errors=True)
model = build_estimator(flags_obj.model_dir, flags_obj.model_type)
train_file = os.path.join(flags_obj.data_dir, 'adult.data')
test_file = os.path.join(flags_obj.data_dir, 'adult.test')
# Train and evaluate the model every `flags.epochs_between_evals` epochs.
def train_input_fn():
return input_fn(
train_file, flags_obj.epochs_between_evals, True, flags_obj.batch_size)
def eval_input_fn():
return input_fn(test_file, 1, False, flags_obj.batch_size)
run_params = {
'batch_size': flags_obj.batch_size,
'train_epochs': flags_obj.train_epochs,
'model_type': flags_obj.model_type,
}
benchmark_logger = logger.config_benchmark_logger(flags_obj)
benchmark_logger.log_run_info('wide_deep', 'Census Income', run_params)
loss_prefix = LOSS_PREFIX.get(flags_obj.model_type, '')
train_hooks = hooks_helper.get_train_hooks(
flags_obj.hooks, batch_size=flags_obj.batch_size,
tensors_to_log={'average_loss': loss_prefix + 'head/truediv',
'loss': loss_prefix + 'head/weighted_loss/Sum'})
# Train and evaluate the model every `flags.epochs_between_evals` epochs.
for n in range(flags_obj.train_epochs // flags_obj.epochs_between_evals):
model.train(input_fn=train_input_fn, hooks=train_hooks)
results = model.evaluate(input_fn=eval_input_fn)
# Display evaluation metrics
tf.logging.info('Results at epoch %d / %d',
(n + 1) * flags_obj.epochs_between_evals,
flags_obj.train_epochs)
tf.logging.info('-' * 60)
for key in sorted(results):
tf.logging.info('%s: %s' % (key, results[key]))
benchmark_logger.log_evaluation_result(results)
if model_helpers.past_stop_threshold(
flags_obj.stop_threshold, results['accuracy']):
break
# Export the model
if flags_obj.export_dir is not None:
export_model(model, flags_obj.model_type, flags_obj.export_dir)
def validate_train_hook_name(self,
test_hook_name,
expected_hook_name,
**kwargs):
returned_hook = hooks_helper.get_train_hooks([test_hook_name], **kwargs)
self.assertEqual(len(returned_hook), 1)
self.assertIsInstance(returned_hook[0], tf.train.SessionRunHook)
self.assertEqual(returned_hook[0].__class__.__name__.lower(),
expected_hook_name)
def run_loop(name, train_input_fn, eval_input_fn, model_column_fn,
build_estimator_fn, flags_obj, tensors_to_log, early_stop=False):
"""Define training loop."""
model_helpers.apply_clean(flags.FLAGS)
model = build_estimator_fn(
model_dir=flags_obj.model_dir, model_type=flags_obj.model_type,
model_column_fn=model_column_fn,
inter_op=flags_obj.inter_op_parallelism_threads,
intra_op=flags_obj.intra_op_parallelism_threads)
run_params = {
'batch_size': flags_obj.batch_size,
'train_epochs': flags_obj.train_epochs,
'model_type': flags_obj.model_type,
}
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info('wide_deep', name, run_params,
test_id=flags_obj.benchmark_test_id)
loss_prefix = LOSS_PREFIX.get(flags_obj.model_type, '')
tensors_to_log = {k: v.format(loss_prefix=loss_prefix)
for k, v in tensors_to_log.items()}
train_hooks = hooks_helper.get_train_hooks(
flags_obj.hooks, model_dir=flags_obj.model_dir,
batch_size=flags_obj.batch_size, tensors_to_log=tensors_to_log)
# Train and evaluate the model every `flags.epochs_between_evals` epochs.
for n in range(flags_obj.train_epochs // flags_obj.epochs_between_evals):
model.train(input_fn=train_input_fn, hooks=train_hooks)
results = model.evaluate(input_fn=eval_input_fn)
# Display evaluation metrics
tf.logging.info('Results at epoch %d / %d',
(n + 1) * flags_obj.epochs_between_evals,
flags_obj.train_epochs)
tf.logging.info('-' * 60)
for key in sorted(results):
tf.logging.info('%s: %s' % (key, results[key]))
benchmark_logger.log_evaluation_result(results)
if early_stop and model_helpers.past_stop_threshold(
flags_obj.stop_threshold, results['accuracy']):
break
# Export the model
if flags_obj.export_dir is not None:
export_model(model, flags_obj.model_type, flags_obj.export_dir,
model_column_fn)
def main(argv):
parser = WideDeepArgParser()
flags = parser.parse_args(args=argv[1:])
# Clean up the model directory if present
shutil.rmtree(flags.model_dir, ignore_errors=True)
model = build_estimator(flags.model_dir, flags.model_type)
train_file = os.path.join(flags.data_dir, 'adult.data')
test_file = os.path.join(flags.data_dir, 'adult.test')
# Train and evaluate the model every `flags.epochs_between_evals` epochs.
def train_input_fn():
return input_fn(
train_file, flags.epochs_between_evals, True, flags.batch_size)
def eval_input_fn():
return input_fn(test_file, 1, False, flags.batch_size)
loss_prefix = LOSS_PREFIX.get(flags.model_type, '')
train_hooks = hooks_helper.get_train_hooks(
flags.hooks, batch_size=flags.batch_size,
tensors_to_log={'average_loss': loss_prefix + 'head/truediv',
'loss': loss_prefix + 'head/weighted_loss/Sum'})
# Train and evaluate the model every `flags.epochs_between_evals` epochs.
for n in range(flags.train_epochs // flags.epochs_between_evals):
model.train(input_fn=train_input_fn, hooks=train_hooks)
results = model.evaluate(input_fn=eval_input_fn)
# Display evaluation metrics
print('Results at epoch', (n + 1) * flags.epochs_between_evals)
print('-' * 60)
for key in sorted(results):
print('%s: %s' % (key, results[key]))
if model_helpers.past_stop_threshold(
flags.stop_threshold, results['accuracy']):
break
def log_and_get_hooks(eval_batch_size):
"""Convenience function for hook and logger creation."""
# Create hooks that log information about the training and metric values
train_hooks = hooks_helper.get_train_hooks(
FLAGS.hooks,
model_dir=FLAGS.model_dir,
batch_size=FLAGS.batch_size, # for ExamplesPerSecondHook
tensors_to_log={"cross_entropy": "cross_entropy"}
)
run_params = {
"batch_size": FLAGS.batch_size,
"eval_batch_size": eval_batch_size,
"number_factors": FLAGS.num_factors,
"hr_threshold": FLAGS.hr_threshold,
"train_epochs": FLAGS.train_epochs,
}
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info(
model_name="recommendation",
dataset_name=FLAGS.dataset,
run_params=run_params,
test_id=FLAGS.benchmark_test_id)
return benchmark_logger, train_hooks
def 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 test_raise_in_non_list_names(self):
with self.assertRaises(ValueError):
hooks_helper.get_train_hooks(
'LoggingTensorHook, ProfilerHook', model_dir="", batch_size=256)
def run(_):
tf.compat.v1.set_random_seed(flags_obj.seed)
print("Data pre-processing...")
train_speech_dataset = generate_dataset(flags_obj.train_data_dir)
eval_speech_dataset = generate_dataset(flags_obj.eval_data_dir)
num_classes = len(train_speech_dataset.speech_labels)
num_gpus = flags_core.get_num_gpus(flags_obj)
distribution_strategy = distribution_utils.get_distribution_strategy(
num_gpus=num_gpus)
run_config = tf.estimator.RunConfig(train_distribute=distribution_strategy)
estimator = tf.estimator.Estimator(model_fn=model_fn,
model_dir=flags_obj.model_dir,
config=run_config,
params={
"num_classes": num_classes,
})
run_params = {
"batch_size": flags_obj.batch_size,
"train_epochs": flags_obj.train_epochs,
"rnn_hidden_size": flags_obj.rnn_hidden_size,
"rnn_hidden_layers": flags_obj.rnn_hidden_layers,
"rnn_type": flags_obj.rnn_type,
"is_bidirectional": flags_obj.is_bidirectional,
"use_bias": flags_obj.use_bias
}
dataset_name = "LibriSpeech"
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info("deep_speech",
dataset_name,
run_params,
test_id=flags_obj.benchmark_test_id)
train_hooks = hooks_helper.get_train_hooks(flags_obj.hooks,
model_dir=flags_obj.model_dir,
batch_size=flags_obj.batch_size)
per_replica_batch_size = distribution_utils.per_replica_batch_size(
flags_obj.batch_size, num_gpus)
def input_fn_train():
return input_fn(per_replica_batch_size, train_speech_dataset)
def input_fn_eval():
return input_fn(per_replica_batch_size, eval_speech_dataset)
total_training_cycle = (flags_obj.train_epochs //
flags_obj.epochs_between_evals)
for cycle_index in range(total_training_cycle):
print('Starting a training cycle: %d/%d' %
(cycle_index + 1, total_training_cycle))
train_speech_dataset.entries = fn.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)
eval_results = evaluate_model(estimator,
eval_speech_dataset.speech_labels,
eval_speech_dataset.entries,
input_fn_eval)
print("Iteration {}: WER = {:.2f}, CER = {:.2f}".format(
cycle_index + 1, eval_results[_WER_KEY], eval_results[_CER_KEY]))
if model_helpers.past_stop_threshold(flags_obj.wer_threshold,
eval_results[_WER_KEY]):
break
def run_mnist(flags_obj):
"""Run MNIST training and eval loop.
Args:
flags_obj: An object containing parsed flag values.
"""
model_function = model_fn
if flags_obj.multi_gpu:
validate_batch_size_for_multi_gpu(flags_obj.batch_size)
# 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)
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': flags_obj.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, 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_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_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)
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
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
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_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_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]
approx_train_steps = None
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)
num_users = ncf_dataset.num_users
num_items = ncf_dataset.num_items
approx_train_steps = int(ncf_dataset.num_train_positives
* (1 + FLAGS.num_neg) // FLAGS.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,
}, 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 = 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 approx_train_steps and 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)
if train_record_dir:
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 run_deep_speech(_):
"""Run deep speech training and eval loop."""
tf.set_random_seed(flags_obj.seed)
# Data preprocessing
tf.logging.info("Data preprocessing...")
train_speech_dataset = generate_dataset(flags_obj.train_data_dir)
eval_speech_dataset = generate_dataset(flags_obj.eval_data_dir)
# Number of label classes. Label string is "[a-z]' -"
num_classes = len(train_speech_dataset.speech_labels)
# Use distribution strategy for multi-gpu training
num_gpus = flags_core.get_num_gpus(flags_obj)
distribution_strategy = distribution_utils.get_distribution_strategy(num_gpus=num_gpus)
run_config = tf.estimator.RunConfig(
train_distribute=distribution_strategy)
estimator = tf.estimator.Estimator(
model_fn=model_fn,
model_dir=flags_obj.model_dir,
config=run_config,
params={
"num_classes": num_classes,
}
)
# Benchmark logging
run_params = {
"batch_size": flags_obj.batch_size,
"train_epochs": flags_obj.train_epochs,
"rnn_hidden_size": flags_obj.rnn_hidden_size,
"rnn_hidden_layers": flags_obj.rnn_hidden_layers,
"rnn_type": flags_obj.rnn_type,
"is_bidirectional": flags_obj.is_bidirectional,
"use_bias": flags_obj.use_bias
}
dataset_name = "LibriSpeech"
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info("deep_speech", dataset_name, run_params,
test_id=flags_obj.benchmark_test_id)
train_hooks = hooks_helper.get_train_hooks(
flags_obj.hooks,
model_dir=flags_obj.model_dir,
batch_size=flags_obj.batch_size)
per_replica_batch_size = per_device_batch_size(flags_obj.batch_size, num_gpus)
def input_fn_train():
return dataset.input_fn(
per_replica_batch_size, train_speech_dataset)
def input_fn_eval():
return dataset.input_fn(
per_replica_batch_size, eval_speech_dataset)
total_training_cycle = (flags_obj.train_epochs //
flags_obj.epochs_between_evals)
for cycle_index in range(total_training_cycle):
tf.logging.info("Starting a training cycle: %d/%d",
cycle_index + 1, total_training_cycle)
# Perform batch_wise dataset shuffling
train_speech_dataset.entries = dataset.batch_wise_dataset_shuffle(
train_speech_dataset.entries, cycle_index, flags_obj.sortagrad,
flags_obj.batch_size)
estimator.train(input_fn=input_fn_train, hooks=train_hooks)
# Evaluation
tf.logging.info("Starting to evaluate...")
eval_results = evaluate_model(
estimator, eval_speech_dataset.speech_labels,
eval_speech_dataset.entries, input_fn_eval)
# Log the WER and CER results.
benchmark_logger.log_evaluation_result(eval_results)
tf.logging.info(
"Iteration {}: WER = {:.2f}, CER = {:.2f}".format(
cycle_index + 1, eval_results[_WER_KEY], eval_results[_CER_KEY]))
# If some evaluation threshold is met
if model_helpers.past_stop_threshold(
flags_obj.wer_threshold, eval_results[_WER_KEY]):
break
def resnet_main(seed, flags, model_function, input_function, shape=None):
"""Shared main loop for ResNet Models.
Args:
flags: FLAGS object that contains the params for running. See
ResnetArgParser for created flags.
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.
shape: list of ints representing the shape of the images used for training.
This is only used if flags.export_dir is passed.
"""
mlperf_log.resnet_print(key=mlperf_log.RUN_START)
# 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.inter_op_parallelism_threads,
intra_op_parallelism_threads=flags.intra_op_parallelism_threads,
allow_soft_placement=True)
if flags.num_gpus == 0:
distribution = tf.contrib.distribute.OneDeviceStrategy('device:CPU:0')
elif flags.num_gpus == 1:
distribution = tf.contrib.distribute.OneDeviceStrategy('device:GPU:0')
else:
distribution = tf.contrib.distribute.MirroredStrategy(
num_gpus=flags.num_gpus)
mlperf_log.resnet_print(key=mlperf_log.RUN_SET_RANDOM_SEED, value=seed)
run_config = tf.estimator.RunConfig(train_distribute=distribution,
session_config=session_config,
tf_random_seed=seed)
mlperf_log.resnet_print(key=mlperf_log.INPUT_BATCH_SIZE,
value=flags.batch_size)
classifier = tf.estimator.Estimator(model_fn=model_function,
model_dir=flags.model_dir,
config=run_config,
params={
'resnet_size': flags.resnet_size,
'data_format': flags.data_format,
'batch_size': flags.batch_size,
'version': flags.version,
'loss_scale': flags.loss_scale,
'dtype': flags.dtype
})
if flags.benchmark_log_dir is not None:
benchmark_logger = logger.BenchmarkLogger(flags.benchmark_log_dir)
benchmark_logger.log_run_info('resnet')
else:
benchmark_logger = None
mlperf_log.resnet_print(key=mlperf_log.TRAIN_LOOP)
# The reference performs the first evaluation on the fourth epoch. (offset
# eval by 3 epochs)
mlperf_log.resnet_print(key=mlperf_log.EVAL_EPOCH_OFFSET, value=3)
success = False
for i in range(flags.train_epochs // flags.epochs_between_evals):
# Data for epochs_between_evals (i.e. 4 epochs between evals) worth of
# epochs is concatenated and run as a single block inside a session. For
# this reason we declare all of the epochs that will be run at the start.
# Submitters may report in a way which is reasonable for their control flow.
for j in range(flags.epochs_between_evals):
mlperf_log.resnet_print(key=mlperf_log.TRAIN_EPOCH,
value=i * flags.epochs_between_evals + j)
train_hooks = hooks_helper.get_train_hooks(
flags.hooks,
batch_size=flags.batch_size,
benchmark_log_dir=flags.benchmark_log_dir)
_log_cache = []
def formatter(x):
"""Abuse side effects to get tensors out of the model_fn."""
if _log_cache:
_log_cache.pop()
_log_cache.append(x.copy())
return str(x)
compliance_hook = tf.train.LoggingTensorHook(
tensors={_NUM_EXAMPLES_NAME: _NUM_EXAMPLES_NAME},
every_n_iter=int(1e10),
at_end=True,
formatter=formatter)
print('Starting a training cycle.')
def input_fn_train():
return input_function(is_training=True,
data_dir=flags.data_dir,
batch_size=per_device_batch_size(
flags.batch_size, flags.num_gpus),
num_epochs=flags.epochs_between_evals,
num_gpus=flags.num_gpus,
dtype=flags.dtype)
classifier.train(input_fn=input_fn_train,
hooks=train_hooks + [compliance_hook],
max_steps=flags.max_train_steps)
train_examples = int(_log_cache.pop()[_NUM_EXAMPLES_NAME])
mlperf_log.resnet_print(key=mlperf_log.INPUT_SIZE,
value=train_examples)
print('Starting to evaluate.')
# Evaluate the model and print results
def input_fn_eval():
return input_function(is_training=False,
data_dir=flags.data_dir,
batch_size=per_device_batch_size(
flags.batch_size, flags.num_gpus),
num_epochs=1,
dtype=flags.dtype)
mlperf_log.resnet_print(key=mlperf_log.EVAL_START)
# flags.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.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.max_train_steps)
mlperf_log.resnet_print(key=mlperf_log.EVAL_STOP)
mlperf_log.resnet_print(key=mlperf_log.EVAL_SIZE,
value=int(eval_results[_NUM_EXAMPLES_NAME]))
mlperf_log.resnet_print(key=mlperf_log.EVAL_ACCURACY,
value=float(eval_results['accuracy']))
mlperf_log.resnet_print(key=mlperf_log.EVAL_TARGET,
value=flags.stop_threshold)
print(eval_results)
if benchmark_logger:
benchmark_logger.log_estimator_evaluation_result(eval_results)
if model_helpers.past_stop_threshold(flags.stop_threshold,
eval_results['accuracy']):
success = True
break
mlperf_log.resnet_print(key=mlperf_log.RUN_STOP,
value={"success": success})
mlperf_log.resnet_print(key=mlperf_log.RUN_FINAL)
def convinh_main(
flags_obj, model_function, input_function, dataset_name, shape=None):
"""Shared main loop for convinh Models.
Args:
flags_obj: An object containing parsed flags. See define_convinh_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.
_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(
tf_random_seed=flags_obj.seed,
train_distribute=distribution_strategy,
session_config=session_config,
keep_checkpoint_max = flags_obj.num_ckpt
)
classifier = tf.estimator.Estimator(
model_fn=model_function, model_dir=flags_obj.model_dir, config=run_config,
params={
'model_params':{
'data_format':flags_obj.data_format,
'filters':list(map(int,flags_obj.filters)),
'ratio_PV': flags_obj.ratio_PV,
'ratio_SST': flags_obj.ratio_SST,
'conv_kernel_size':list(map(int,flags_obj.conv_kernel_size)),
'conv_kernel_size_inh':list(map(int,flags_obj.conv_kernel_size_inh)),
'conv_strides':list(map(int,flags_obj.conv_strides)),
'pool_size':list(map(int,flags_obj.pool_size)),
'pool_strides':list(map(int,flags_obj.pool_strides)),
'num_ff_layers':flags_obj.num_ff_layers,
'num_rnn_layers':flags_obj.num_rnn_layers,
'connection':flags_obj.connection,
'n_time':flags_obj.n_time,
'cell_fn':flags_obj.cell_fn,
'act_fn':flags_obj.act_fn,
'pvsst_circuit':flags_obj.pvsst_circuit,
'gating':flags_obj.gating,
'normalize':flags_obj.normalize,
'num_classes':flags_obj.num_classes
},
'batch_size' : flags_obj.batch_size,
'weight_decay': flags_obj.weight_decay,
'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),
'convinh_size': flags_obj.convinh_size, # deprecated
'convinh_version': flags_obj.convinh_version, # deprecated
'synthetic_data': flags_obj.use_synthetic_data, # deprecated
'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('convinh', 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)
class input_fn_train(object):
def __init__(self,num_epochs):
self._num_epochs = num_epochs
def __call__(self):
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=self._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)
tf.logging.info('Evaluate the intial model.')
eval_results = classifier.evaluate(input_fn=input_fn_eval,
steps=flags_obj.max_train_steps)
benchmark_logger.log_evaluation_result(eval_results)
# training
total_training_cycle = (flags_obj.train_epochs //
flags_obj.epochs_between_evals) + 1
for cycle_index in range(total_training_cycle):
cur_train_epochs = flags_obj.epochs_between_evals if cycle_index else 1
tf.logging.info('Starting a training cycle: %d/%d, with %d epochs',
cycle_index, total_training_cycle, cur_train_epochs)
classifier.train(input_fn=input_fn_train(cur_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=1)
if cycle_index==0:
classifier.export_savedmodel(flags_obj.export_dir, input_receiver_fn,
checkpoint_path='{}/model.ckpt-0'.format(flags_obj.model_dir))
classifier.export_savedmodel(flags_obj.export_dir, input_receiver_fn)
def resnet_main(flags_obj,
model_function,
input_function,
dataset_name,
shape=None,
num_images=None,
zeroshot_eval=False):
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 = config_utils.get_session_config(flags_obj)
run_config = config_utils.get_run_config(flags_obj, flags_core,
session_config,
num_images['train'])
def gen_estimator(period=None):
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_tf = flags_core.get_tf_dtype(flags_obj)
num_epochs_per_decay = flags_obj.num_epochs_per_decay
learning_rate_decay_factor = flags_obj.learning_rate_decay_factor
end_learning_rate = flags_obj.end_learning_rate
learning_rate_decay_type = flags_obj.learning_rate_decay_type
weight_decay = flags_obj.weight_decay
zero_gamma = flags_obj.zero_gamma
lr_warmup_epochs = flags_obj.lr_warmup_epochs
base_learning_rate = flags_obj.base_learning_rate
use_resnet_d = flags_obj.use_resnet_d
use_dropblock = flags_obj.use_dropblock
dropblock_kp = [float(be) for be in flags_obj.dropblock_kp]
label_smoothing = flags_obj.label_smoothing
momentum = flags_obj.momentum
bn_momentum = flags_obj.bn_momentum
train_epochs = flags_obj.train_epochs
piecewise_lr_boundary_epochs = [
int(be) for be in flags_obj.piecewise_lr_boundary_epochs
]
piecewise_lr_decay_rates = [
float(dr) for dr in flags_obj.piecewise_lr_decay_rates
]
use_ranking_loss = flags_obj.use_ranking_loss
use_se_block = flags_obj.use_se_block
use_sk_block = flags_obj.use_sk_block
mixup_type = flags_obj.mixup_type
dataset_name = flags_obj.dataset_name
kd_temp = flags_obj.kd_temp
no_downsample = flags_obj.no_downsample
anti_alias_filter_size = flags_obj.anti_alias_filter_size
anti_alias_type = flags_obj.anti_alias_type
cls_loss_type = flags_obj.cls_loss_type
logit_type = flags_obj.logit_type
embedding_size = flags_obj.embedding_size
pool_type = flags_obj.pool_type
arc_s = flags_obj.arc_s
arc_m = flags_obj.arc_m
bl_alpha = flags_obj.bl_alpha
bl_beta = flags_obj.bl_beta
exp = None
if install_hyperdash and flags_obj.use_hyperdash:
exp = Experiment(flags_obj.model_dir.split("/")[-1])
resnet_size = exp.param("resnet_size", int(flags_obj.resnet_size))
batch_size = exp.param("batch_size", flags_obj.batch_size)
exp.param("dtype", flags_obj.dtype)
learning_rate_decay_type = exp.param(
"learning_rate_decay_type", flags_obj.learning_rate_decay_type)
weight_decay = exp.param("weight_decay", flags_obj.weight_decay)
zero_gamma = exp.param("zero_gamma", flags_obj.zero_gamma)
lr_warmup_epochs = exp.param("lr_warmup_epochs",
flags_obj.lr_warmup_epochs)
base_learning_rate = exp.param("base_learning_rate",
flags_obj.base_learning_rate)
use_dropblock = exp.param("use_dropblock", flags_obj.use_dropblock)
dropblock_kp = exp.param(
"dropblock_kp", [float(be) for be in flags_obj.dropblock_kp])
piecewise_lr_boundary_epochs = exp.param(
"piecewise_lr_boundary_epochs",
[int(be) for be in flags_obj.piecewise_lr_boundary_epochs])
piecewise_lr_decay_rates = exp.param(
"piecewise_lr_decay_rates",
[float(dr) for dr in flags_obj.piecewise_lr_decay_rates])
mixup_type = exp.param("mixup_type", flags_obj.mixup_type)
dataset_name = exp.param("dataset_name", flags_obj.dataset_name)
exp.param("autoaugment_type", flags_obj.autoaugment_type)
classifier = tf.estimator.Estimator(
model_fn=model_function,
model_dir=flags_obj.model_dir,
config=run_config,
params={
'resnet_size': resnet_size,
'data_format': data_format,
'batch_size': batch_size,
'resnet_version': resnet_version,
'loss_scale': loss_scale,
'dtype': dtype_tf,
'num_epochs_per_decay': num_epochs_per_decay,
'learning_rate_decay_factor': learning_rate_decay_factor,
'end_learning_rate': end_learning_rate,
'learning_rate_decay_type': learning_rate_decay_type,
'weight_decay': weight_decay,
'zero_gamma': zero_gamma,
'lr_warmup_epochs': lr_warmup_epochs,
'base_learning_rate': base_learning_rate,
'use_resnet_d': use_resnet_d,
'use_dropblock': use_dropblock,
'dropblock_kp': dropblock_kp,
'label_smoothing': label_smoothing,
'momentum': momentum,
'bn_momentum': bn_momentum,
'embedding_size': embedding_size,
'train_epochs': train_epochs,
'piecewise_lr_boundary_epochs': piecewise_lr_boundary_epochs,
'piecewise_lr_decay_rates': piecewise_lr_decay_rates,
'with_drawing_bbox': flags_obj.with_drawing_bbox,
'use_ranking_loss': use_ranking_loss,
'use_se_block': use_se_block,
'use_sk_block': use_sk_block,
'mixup_type': mixup_type,
'kd_temp': kd_temp,
'no_downsample': no_downsample,
'dataset_name': dataset_name,
'anti_alias_filter_size': anti_alias_filter_size,
'anti_alias_type': anti_alias_type,
'cls_loss_type': cls_loss_type,
'logit_type': logit_type,
'arc_s': arc_s,
'arc_m': arc_m,
'pool_type': pool_type,
'bl_alpha': bl_alpha,
'bl_beta': bl_beta,
'train_steps': total_train_steps,
})
return classifier, exp
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,
use_random_crop=flags_obj.training_random_crop,
num_epochs=num_epochs,
flags_obj=flags_obj)
def input_fn_eval():
return input_function(is_training=False,
use_random_crop=False,
num_epochs=1,
flags_obj=flags_obj)
ckpt_keeper = checkpoint_utils.CheckpointKeeper(
save_dir=flags_obj.model_dir,
num_to_keep=flags_obj.num_best_ckpt_to_keep,
keep_epoch=flags_obj.keep_ckpt_every_eval,
maximize=True)
if zeroshot_eval:
dataset = data_config.get_config(dataset_name)
model = model_fns.Model(
int(flags_obj.resnet_size),
flags_obj.data_format,
resnet_version=int(flags_obj.resnet_version),
num_classes=dataset.num_classes,
zero_gamma=flags_obj.zero_gamma,
use_se_block=flags_obj.use_se_block,
use_sk_block=flags_obj.use_sk_block,
no_downsample=flags_obj.no_downsample,
anti_alias_filter_size=flags_obj.anti_alias_filter_size,
anti_alias_type=flags_obj.anti_alias_type,
bn_momentum=flags_obj.bn_momentum,
embedding_size=flags_obj.embedding_size,
pool_type=flags_obj.pool_type,
bl_alpha=flags_obj.bl_alpha,
bl_beta=flags_obj.bl_beta,
dtype=flags_core.get_tf_dtype(flags_obj),
loss_type=flags_obj.cls_loss_type)
def train_and_evaluate(hooks):
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=hooks,
steps=flags_obj.max_train_steps)
tf.logging.info('Starting to evaluate.')
if zeroshot_eval:
tf.reset_default_graph()
eval_results = recall_metric.recall_at_k(
flags_obj,
flags_core,
input_fns.input_fn_ir_eval,
model,
num_images['validation'],
eval_similarity=flags_obj.eval_similarity,
return_embedding=True)
else:
eval_results = classifier.evaluate(input_fn=input_fn_eval,
steps=flags_obj.max_train_steps)
return eval_results
total_train_steps = flags_obj.train_epochs * int(
num_images['train'] / flags_obj.batch_size)
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
elif flags_obj.export_only:
schedule, n_loops = [], 0
else:
n_loops = math.ceil(flags_obj.train_epochs /
flags_obj.epochs_between_evals)
schedule = [
flags_obj.epochs_between_evals for _ in range(int(n_loops))
]
schedule[-1] = flags_obj.train_epochs - sum(
schedule[:-1]) # over counting.
schedule = config_utils.get_epoch_schedule(flags_obj, schedule,
num_images)
tf.logging.info('epoch schedule:')
tf.logging.info(schedule)
classifier, exp = gen_estimator()
if flags_obj.pretrained_model_checkpoint_path:
warm_start_hook = WarmStartHook(
flags_obj.pretrained_model_checkpoint_path)
train_hooks.append(warm_start_hook)
for cycle_index, num_train_epochs in enumerate(schedule):
eval_results = train_and_evaluate(train_hooks)
if zeroshot_eval:
metric = eval_results['recall_at_1']
else:
metric = eval_results['accuracy']
ckpt_keeper.save(metric, flags_obj.model_dir)
if exp:
exp.metric("accuracy", metric)
benchmark_logger.log_evaluation_result(eval_results)
if model_helpers.past_stop_threshold(flags_obj.stop_threshold, metric):
break
if model_helpers.past_stop_threshold(total_train_steps,
eval_results['global_step']):
break
if exp:
exp.end()
if flags_obj.export_dir is not None:
export_utils.export_pb(flags_core, flags_obj, shape, classifier)
def main(_):
"""Train NCF model and evaluate its hit rate (HR) metric."""
params = create_params()
if FLAGS.seed is not None:
np.random.seed(FLAGS.seed)
if FLAGS.use_synthetic_data:
producer = data_pipeline.DummyConstructor()
num_users, num_items = data_preprocessing.DATASET_TO_NUM_USERS_AND_ITEMS[
FLAGS.dataset]
num_train_steps = rconst.SYNTHETIC_BATCHES_PER_EPOCH
num_eval_steps = rconst.SYNTHETIC_BATCHES_PER_EPOCH
else:
num_users, num_items, producer = data_preprocessing.instantiate_pipeline(
dataset=FLAGS.dataset, data_dir=FLAGS.data_dir,
epoch_dir=os.path.join(params["model_dir"], "epoch"),
params=get_params_for_dataset(params),
constructor_type=FLAGS.constructor_type,
deterministic=FLAGS.seed is not None)
num_train_steps = (producer.train_batches_per_epoch //
params["batches_per_step"])
num_eval_steps = (producer.eval_batches_per_epoch //
params["batches_per_step"])
assert not producer.train_batches_per_epoch % params["batches_per_step"]
assert not producer.eval_batches_per_epoch % params["batches_per_step"]
producer.start()
params["num_users"] = num_users
params["num_items"] = num_items
feature_columns = create_feature_columns(params)
model_fn = create_model_fn(feature_columns)
estimator = create_tpu_estimator(model_fn, feature_columns, params)
train_hooks = hooks_helper.get_train_hooks(
["ProfilerHook"],
model_dir=FLAGS.model_dir,
batch_size=FLAGS.batch_size, # for ExamplesPerSecondHook
tensors_to_log={"cross_entropy": "cross_entropy"}
)
for cycle_index in range(FLAGS.train_epochs):
tf.logging.info("Starting a training cycle: {}/{}".format(
cycle_index + 1, FLAGS.train_epochs))
train_input_fn = producer.make_input_fn(is_training=True)
estimator.train(input_fn=train_input_fn, hooks=train_hooks,
steps=num_train_steps)
tf.logging.info("Beginning evaluation.")
eval_input_fn = producer.make_input_fn(is_training=False)
eval_results = estimator.evaluate(eval_input_fn, steps=num_eval_steps)
tf.logging.info("Evaluation complete.")
hr = float(eval_results[rconst.HR_KEY])
ndcg = float(eval_results[rconst.NDCG_KEY])
loss = float(eval_results["loss"])
tf.logging.info(
"Iteration {}: HR = {:.4f}, NDCG = {:.4f}, Loss = {:.4f}".format(
cycle_index + 1, hr, ndcg, loss))
producer.stop_loop()
producer.join()
def resnet_main(flags_obj,
model_function,
input_function,
dataset_name,
shape=None):
"""Shared main loop for ResNet Models.
Args:
flags_obj: An object containing parsed flags. See define_resnet_flags()
for details.
model_function: the function that instantiates the Model and builds the
ops for train/eval. This will be passed directly into the estimator.
input_function: the function that processes the dataset and returns a
dataset that the estimator can train on. This will be wrapped with
all the relevant flags for running and passed to estimator.
dataset_name: the name of the dataset for training and evaluation. This is
used for logging purpose.
shape: list of ints representing the shape of the images used for training.
This is only used if flags_obj.export_dir is passed.
"""
model_helpers.apply_clean(flags.FLAGS)
# Using the Winograd non-fused algorithms provides a small performance boost.
os.environ['TF_ENABLE_WINOGRAD_NONFUSED'] = '1'
# Create session config based on values of inter_op_parallelism_threads and
# intra_op_parallelism_threads. Note that we default to having
# allow_soft_placement = True, which is required for multi-GPU and not
# harmful for other modes.
session_config = tf.ConfigProto(
inter_op_parallelism_threads=flags_obj.inter_op_parallelism_threads,
intra_op_parallelism_threads=flags_obj.intra_op_parallelism_threads,
allow_soft_placement=True)
distribution_strategy = distribution_utils.get_distribution_strategy(
flags_core.get_num_gpus(flags_obj), flags_obj.all_reduce_alg)
run_config = tf.estimator.RunConfig(train_distribute=distribution_strategy,
session_config=session_config)
classifier = tf.estimator.Estimator(
model_fn=model_function,
model_dir=flags_obj.model_dir,
config=run_config,
params={
'resnet_size': int(flags_obj.resnet_size),
'data_format': flags_obj.data_format,
'batch_size': flags_obj.batch_size,
'resnet_version': int(flags_obj.resnet_version),
'loss_scale': flags_core.get_loss_scale(flags_obj),
'dtype': flags_core.get_tf_dtype(flags_obj)
})
run_params = {
'batch_size': flags_obj.batch_size,
'dtype': flags_core.get_tf_dtype(flags_obj),
'resnet_size': flags_obj.resnet_size,
'resnet_version': flags_obj.resnet_version,
'synthetic_data': flags_obj.use_synthetic_data,
'train_epochs': flags_obj.train_epochs,
}
if flags_obj.use_synthetic_data:
dataset_name = dataset_name + '-synthetic'
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info('resnet',
dataset_name,
run_params,
test_id=flags_obj.benchmark_test_id)
train_hooks = hooks_helper.get_train_hooks(flags_obj.hooks,
model_dir=flags_obj.model_dir,
batch_size=flags_obj.batch_size)
def input_fn_train():
return input_function(
is_training=True,
data_dir=flags_obj.data_dir,
batch_size=distribution_utils.per_device_batch_size(
flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
num_epochs=flags_obj.epochs_between_evals,
num_gpus=flags_core.get_num_gpus(flags_obj))
def input_fn_eval():
return input_function(
is_training=False,
data_dir=flags_obj.data_dir,
batch_size=distribution_utils.per_device_batch_size(
flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
num_epochs=1)
total_training_cycle = (flags_obj.train_epochs //
flags_obj.epochs_between_evals)
for cycle_index in range(total_training_cycle):
tf.logging.info('Starting a training cycle: %d/%d', cycle_index,
total_training_cycle)
classifier.train(input_fn=input_fn_train,
hooks=train_hooks,
max_steps=flags_obj.max_train_steps)
tf.logging.info('Starting to evaluate.')
# flags_obj.max_train_steps is generally associated with testing and
# profiling. As a result it is frequently called with synthetic data, which
# will iterate forever. Passing steps=flags_obj.max_train_steps allows the
# eval (which is generally unimportant in those circumstances) to terminate.
# Note that eval will run for max_train_steps each loop, regardless of the
# global_step count.
eval_results = classifier.evaluate(input_fn=input_fn_eval,
steps=flags_obj.max_train_steps)
tf.logging.info(
"Convert types of eval_results (np.xx types) to std python (needed for json.dump():"
)
for key, value in eval_results.items():
eval_results[key] = value.item()
#print(key, eval_results[key], type(eval_results[key]))
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)
savedmodel_raw_path = classifier.export_savedmodel(
flags_obj.export_dir, input_receiver_fn)
savedmodel_raw_path = savedmodel_raw_path.decode(
) # convert a byte string to a normal string
if (flags_obj.benchmark_logger_type == "BenchmarkFileLogger"
and hasattr(flags_obj, "benchmark_log_dir")):
log_dir = flags_obj.benchmark_log_dir
benchmark_log_path = os.path.join(
log_dir, logger.BENCHMARK_RUN_LOG_FILE_NAME)
# attempt to identify a number of GPUs or was it on CPU
with open(benchmark_log_path, "r") as json_file:
log = json.load(json_file)
num_gpus = int(log["machine_config"]["gpu_info"]["count"])
gpu_info = ""
if num_gpus == 0:
gpu_info = "_cpu"
elif num_gpus > 0:
gpu_info = "_" + str(num_gpus) + "gpu"
# add "_cpu" or "_gpu" to just saved graph directory
savedmodel_path = savedmodel_raw_path + gpu_info
shutil.move(savedmodel_raw_path, savedmodel_path)
# move benchmark_run.log and metric.log to the just saved graph
shutil.move(benchmark_log_path, savedmodel_path)
shutil.move(os.path.join(log_dir, logger.METRIC_LOG_FILE_NAME),
savedmodel_path)
tf.logging.info("[INFO]: SavedModel path: %s", savedmodel_path)
# zip the trained graph, aka savedmodel:
# adapted from https://stackoverflow.com/questions/1855095/how-to-create-a-zip-archive-of-a-directory-in-python
# full path to the zip file
graph_zip_path = savedmodel_path + '.zip'
# cd to the directory with the trained graph
os.chdir(os.path.dirname(savedmodel_path.rstrip('/')))
dir_to_zip = savedmodel_path.rstrip('/').split('/')[-1]
graph_zip = zipfile.ZipFile(graph_zip_path, 'w', zipfile.ZIP_DEFLATED)
for root, dirs, files in os.walk(dir_to_zip):
for file in files:
graph_zip.write(os.path.join(root, file))
graph_zip.close()
return graph_zip_path
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 resnet_main(seed, flags, model_function, input_function, shape=None):
"""Shared main loop for ResNet Models.
Args:
flags: FLAGS object that contains the params for running. See
ResnetArgParser for created flags.
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.
shape: list of ints representing the shape of the images used for training.
This is only used if flags.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.inter_op_parallelism_threads,
intra_op_parallelism_threads=flags.intra_op_parallelism_threads,
allow_soft_placement=True)
if flags.num_gpus == 0:
distribution = tf.contrib.distribute.OneDeviceStrategy('device:CPU:0')
elif flags.num_gpus == 1:
distribution = tf.contrib.distribute.OneDeviceStrategy('device:GPU:0')
else:
distribution = tf.contrib.distribute.MirroredStrategy(
num_gpus=flags.num_gpus
)
run_config = tf.estimator.RunConfig(train_distribute=distribution,
session_config=session_config, tf_random_seed=seed)
classifier = tf.estimator.Estimator(
model_fn=model_function, model_dir=flags.model_dir, config=run_config,
params={
'resnet_size': flags.resnet_size,
'data_format': flags.data_format,
'batch_size': flags.batch_size,
'version': flags.version,
'loss_scale': flags.loss_scale,
'dtype': flags.dtype
})
if flags.benchmark_log_dir is not None:
benchmark_logger = logger.BenchmarkLogger(flags.benchmark_log_dir)
benchmark_logger.log_run_info('resnet')
else:
benchmark_logger = None
for _ in range(flags.train_epochs // flags.epochs_between_evals):
train_hooks = hooks_helper.get_train_hooks(
flags.hooks,
batch_size=flags.batch_size,
benchmark_log_dir=flags.benchmark_log_dir)
print('Starting a training cycle.')
def input_fn_train():
return input_function(
is_training=True,
data_dir=flags.data_dir,
batch_size=per_device_batch_size(flags.batch_size, flags.num_gpus),
num_epochs=flags.epochs_between_evals,
)
classifier.train(input_fn=input_fn_train, hooks=train_hooks,
max_steps=flags.max_train_steps)
print('Starting to evaluate.')
# Evaluate the model and print results
def input_fn_eval():
return input_function(
is_training=False,
data_dir=flags.data_dir,
batch_size=per_device_batch_size(flags.batch_size, flags.num_gpus),
num_epochs=1,
)
# flags.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.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.max_train_steps)
print(eval_results)
if benchmark_logger:
benchmark_logger.log_estimator_evaluation_result(eval_results)
if model_helpers.past_stop_threshold(
flags.stop_threshold, eval_results['accuracy']):
break
if flags.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.batch_size)
classifier.export_savedmodel(flags.export_dir, input_receiver_fn)
def test_raise_in_invalid_names(self):
invalid_names = ['StepCounterHook', 'StopAtStepHook']
with self.assertRaises(ValueError):
hooks_helper.get_train_hooks(invalid_names, model_dir="", batch_size=256)
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_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_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=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 test_raise_in_non_list_names(self):
with self.assertRaises(ValueError):
hooks_helper.get_train_hooks(
'LoggingTensorHook, ProfilerHook', batch_size=256)
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.
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
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,
'num_workers': num_workers,
})
run_params = {
'batch_size': flags_obj.batch_size,
'dtype': flags_core.get_tf_dtype(flags_obj),
'resnet_size': flags_obj.resnet_size,
'resnet_version': flags_obj.resnet_version,
'synthetic_data': flags_obj.use_synthetic_data,
'train_epochs': flags_obj.train_epochs,
'num_workers': num_workers,
}
if flags_obj.use_synthetic_data:
dataset_name = dataset_name + '-synthetic'
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info('resnet', dataset_name, run_params,
test_id=flags_obj.benchmark_test_id)
train_hooks = hooks_helper.get_train_hooks(
flags_obj.hooks,
model_dir=flags_obj.model_dir,
batch_size=flags_obj.batch_size)
def input_fn_train(num_epochs, input_context=None):
return input_function(
is_training=True,
data_dir=flags_obj.data_dir,
batch_size=distribution_utils.per_replica_batch_size(
flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
num_epochs=num_epochs,
dtype=flags_core.get_tf_dtype(flags_obj),
datasets_num_private_threads=flags_obj.datasets_num_private_threads,
num_parallel_batches=flags_obj.datasets_num_parallel_batches,
input_context=input_context)
def input_fn_eval():
return input_function(
is_training=False,
data_dir=flags_obj.data_dir,
batch_size=distribution_utils.per_replica_batch_size(
flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
num_epochs=1,
dtype=flags_core.get_tf_dtype(flags_obj))
train_epochs = (0 if flags_obj.eval_only or not flags_obj.train_epochs else
flags_obj.train_epochs)
use_train_and_evaluate = flags_obj.use_train_and_evaluate or num_workers > 1
if use_train_and_evaluate:
train_spec = tf.estimator.TrainSpec(
input_fn=lambda input_context=None: input_fn_train(
train_epochs, input_context=input_context),
hooks=train_hooks,
max_steps=flags_obj.max_train_steps)
eval_spec = tf.estimator.EvalSpec(input_fn=input_fn_eval)
tf.compat.v1.logging.info('Starting to train and evaluate.')
tf.estimator.train_and_evaluate(classifier, train_spec, eval_spec)
# tf.estimator.train_and_evalute doesn't return anything in multi-worker
# case.
eval_results = {}
else:
if train_epochs == 0:
# If --eval_only is set, perform a single loop with zero train epochs.
schedule, n_loops = [0], 1
else:
# Compute the number of times to loop while training. All but the last
# pass will train for `epochs_between_evals` epochs, while the last will
# train for the number needed to reach `training_epochs`. For instance if
# train_epochs = 25 and epochs_between_evals = 10
# schedule will be set to [10, 10, 5]. That is to say, the loop will:
# Train for 10 epochs and then evaluate.
# Train for another 10 epochs and then evaluate.
# Train for a final 5 epochs (to reach 25 epochs) and then evaluate.
n_loops = math.ceil(train_epochs / flags_obj.epochs_between_evals)
schedule = [flags_obj.epochs_between_evals for _ in range(int(n_loops))]
schedule[-1] = train_epochs - sum(schedule[:-1]) # over counting.
for cycle_index, num_train_epochs in enumerate(schedule):
tf.compat.v1.logging.info('Starting cycle: %d/%d', cycle_index,
int(n_loops))
if num_train_epochs:
# Since we are calling classifier.train immediately in each loop, the
# value of num_train_epochs in the lambda function will not be changed
# before it is used. So it is safe to ignore the pylint error here
# pylint: disable=cell-var-from-loop
classifier.train(
input_fn=lambda input_context=None: input_fn_train(
num_train_epochs, input_context=input_context),
hooks=train_hooks,
max_steps=flags_obj.max_train_steps)
# flags_obj.max_train_steps is generally associated with testing and
# profiling. As a result it is frequently called with synthetic data,
# which will iterate forever. Passing steps=flags_obj.max_train_steps
# allows the eval (which is generally unimportant in those circumstances)
# to terminate. Note that eval will run for max_train_steps each loop,
# regardless of the global_step count.
tf.compat.v1.logging.info('Starting to evaluate.')
eval_results = classifier.evaluate(input_fn=input_fn_eval,
steps=flags_obj.max_train_steps)
benchmark_logger.log_evaluation_result(eval_results)
if model_helpers.past_stop_threshold(
flags_obj.stop_threshold, eval_results['accuracy']):
break
if flags_obj.export_dir is not None:
# Exports a saved model for the given classifier.
export_dtype = flags_core.get_tf_dtype(flags_obj)
if flags_obj.image_bytes_as_serving_input:
input_receiver_fn = functools.partial(
image_bytes_serving_input_fn, shape, dtype=export_dtype)
else:
input_receiver_fn = export.build_tensor_serving_input_receiver_fn(
shape, batch_size=flags_obj.batch_size, dtype=export_dtype)
classifier.export_savedmodel(flags_obj.export_dir, input_receiver_fn,
strip_default_attrs=True)
stats = {}
stats['eval_results'] = eval_results
stats['train_hooks'] = train_hooks
return stats
def run_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 run_loop(name,
train_input_fn,
eval_input_fn,
model_column_fn,
build_estimator_fn,
flags_obj,
tensors_to_log,
early_stop=False):
"""Define training loop."""
model_helpers.apply_clean(flags.FLAGS)
model = build_estimator_fn(model_dir=flags_obj.model_dir,
model_type=flags_obj.model_type,
model_column_fn=model_column_fn,
inter_op=flags_obj.inter_op_parallelism_threads,
intra_op=flags_obj.intra_op_parallelism_threads)
run_params = {
'batch_size': flags_obj.batch_size,
'train_epochs': flags_obj.train_epochs,
'model_type': flags_obj.model_type,
}
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info('wide_deep',
name,
run_params,
test_id=flags_obj.benchmark_test_id)
loss_prefix = LOSS_PREFIX.get(flags_obj.model_type, '')
tensors_to_log = {
k: v.format(loss_prefix=loss_prefix)
for k, v in tensors_to_log.items()
}
train_hooks = hooks_helper.get_train_hooks(flags_obj.hooks,
model_dir=flags_obj.model_dir,
batch_size=flags_obj.batch_size,
tensors_to_log=tensors_to_log)
# Train and evaluate the model every `flags.epochs_between_evals` epochs.
for n in range(flags_obj.train_epochs // flags_obj.epochs_between_evals):
model.train(input_fn=train_input_fn, hooks=train_hooks)
results = model.evaluate(input_fn=eval_input_fn)
# Display evaluation metrics
tf.logging.info('Results at epoch %d / %d',
(n + 1) * flags_obj.epochs_between_evals,
flags_obj.train_epochs)
tf.logging.info('-' * 60)
for key in sorted(results):
tf.logging.info('%s: %s' % (key, results[key]))
benchmark_logger.log_evaluation_result(results)
if early_stop and model_helpers.past_stop_threshold(
flags_obj.stop_threshold, results['accuracy']):
break
# Export the model
if flags_obj.export_dir is not None:
export_model(model, flags_obj.model_type, flags_obj.export_dir,
model_column_fn)
def resnet_main(flags, model_function, input_function, shape=None):
"""Shared main loop for ResNet Models.
Args:
flags: FLAGS object that contains the params for running. See
ResnetArgParser for created flags.
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.
shape: list of ints representing the shape of the images used for training.
This is only used if flags.export_dir is passed.
"""
# Using the Winograd non-fused algorithms provides a small performance boost.
os.environ['TF_ENABLE_WINOGRAD_NONFUSED'] = '1'
if flags.multi_gpu:
validate_batch_size_for_multi_gpu(flags.batch_size)
# 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_function,
loss_reduction=tf.losses.Reduction.MEAN)
# 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.inter_op_parallelism_threads,
intra_op_parallelism_threads=flags.intra_op_parallelism_threads,
allow_soft_placement=True)
# Set up a RunConfig to save checkpoint and set session config.
run_config = tf.estimator.RunConfig().replace(save_checkpoints_secs=1e9,
session_config=session_config)
classifier = tf.estimator.Estimator(
model_fn=model_function, model_dir=flags.model_dir, config=run_config,
params={
'resnet_size': flags.resnet_size,
'data_format': flags.data_format,
'batch_size': flags.batch_size,
'multi_gpu': flags.multi_gpu,
'version': flags.version,
})
if flags.benchmark_log_dir is not None:
benchmark_logger = logger.BenchmarkLogger(flags.benchmark_log_dir)
benchmark_logger.log_run_info("resnet")
else:
benchmark_logger = None
for _ in range(flags.train_epochs // flags.epochs_between_evals):
train_hooks = hooks_helper.get_train_hooks(
flags.hooks,
batch_size=flags.batch_size,
benchmark_log_dir=flags.benchmark_log_dir)
print('Starting a training cycle.')
def input_fn_train():
return input_function(True, flags.data_dir, flags.batch_size,
flags.epochs_between_evals,
flags.num_parallel_calls, flags.multi_gpu)
classifier.train(input_fn=input_fn_train, hooks=train_hooks,
max_steps=flags.max_train_steps)
print('Starting to evaluate.')
# Evaluate the model and print results
def input_fn_eval():
return input_function(False, flags.data_dir, flags.batch_size,
1, flags.num_parallel_calls, flags.multi_gpu)
# flags.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.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.max_train_steps)
print(eval_results)
if benchmark_logger:
benchmark_logger.log_estimator_evaluation_result(eval_results)
if flags.export_dir is not None:
warn_on_multi_gpu_export(flags.multi_gpu)
# Exports a saved model for the given classifier.
input_receiver_fn = export.build_tensor_serving_input_receiver_fn(
shape, batch_size=flags.batch_size)
classifier.export_savedmodel(flags.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)
# 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_mnist(flags_obj):
"""Run MNIST training and eval loop.
Args:
flags_obj: An object containing parsed flag values.
"""
model_function = model_fn
if flags_obj.multi_gpu:
validate_batch_size_for_multi_gpu(flags_obj.batch_size)
# 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)
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': flags_obj.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, 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_Simple_model(flags_obj):
"""Run Simple_model 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 = 'channels_first'
classifier = tf.estimator.Estimator(
model_fn=model_function,
model_dir=flags_obj.model_dir,
config=run_config,
params={
'data_format': data_format,
})
# Use `tf.parse_single_example()` to extract data from a `tf.Example`
# protocol buffer, and perform any additional per-record preprocessing.
def parser_train(record):
keys_to_features = {
"image_raw": tf.FixedLenFeature((), tf.string, default_value=""),
"label": tf.FixedLenFeature((), tf.int64,
default_value=tf.zeros([], dtype=tf.int64)),
}
parsed = tf.parse_single_example(record, keys_to_features)
# Perform additional preprocessing on the parsed data.
image = tf.image.decode_jpeg(parsed["image_raw"])
image = tf.reshape(image, [64, 64, 3])
# subtract mean
mean = tf.convert_to_tensor(mean_value, dtype=tf.uint8)
image = image - mean
image = tf.image.random_crop(image, [48, 48, 3])
image = tf.image.random_flip_left_right(image) # flip image randomly with a 50% chance
image = tf.image.random_brightness(image, max_delta=0.3) # random brightness
l = [1, 2, 3]
deg = random.choice(l)
image = tf.image.rot90(image, k=deg)
image = tf.transpose(image, perm=[2, 0, 1]) # channels first
image = tf.cast(image, tf.int32)
image = tf.truediv(image, 255)
image = tf.cast(image, tf.float32)
label = tf.cast(parsed["label"], tf.int32)
return {"image": image}, label
# Use `tf.parse_single_example()` to extract data from a `tf.Example`
# protocol buffer, and perform any additional per-record preprocessing.
def parser_eval(record):
keys_to_features = {
"image_raw": tf.FixedLenFeature((), tf.string, default_value=""),
"label": tf.FixedLenFeature((), tf.int64,
default_value=tf.zeros([], dtype=tf.int64)),
}
parsed = tf.parse_single_example(record, keys_to_features)
# Perform additional preprocessing on the parsed data.
image = tf.image.decode_jpeg(parsed["image_raw"])
image = tf.reshape(image, [64, 64, 3])
# subtract mean
mean = tf.convert_to_tensor(mean_value, dtype=tf.uint8)
image = image - mean
image = tf.image.random_crop(image, [48, 48, 3])
image = tf.transpose(image, perm=[2, 0, 1]) # channels first
image = tf.cast(image, tf.int32)
image = tf.truediv(image, 255)
image = tf.cast(image, tf.float32)
label = tf.cast(parsed["label"], tf.int32)
return {"image": image}, label
# Set up training and evaluation input functions.
def train_input_fn():
"""Prepare data for training."""
train_tfrecord = 'dataset/train_images.tfrecords'
dataset = tf.data.TFRecordDataset(train_tfrecord)
# Use `Dataset.map()` to build a pair of a feature dictionary and a label
# tensor for each example.
dataset = dataset.map(parser_train)
dataset = dataset.shuffle(buffer_size=5000)
dataset = dataset.batch(flags_obj.batch_size)
# Iterate through the dataset a set number (`epochs_between_evals`) of times
# during each training session.
dataset = dataset.repeat(flags_obj.epochs_between_evals)
return dataset
def eval_input_fn():
filenames = ["dataset/val_images.tfrecords"]
dataset = tf.data.TFRecordDataset(filenames)
# Use `Dataset.map()` to build a pair of a feature dictionary and a label
# tensor for each example.
dataset = dataset.map(parser_eval)
dataset = dataset.batch(flags_obj.batch_size)
# repeat:restarts dataset when reaches the end so here is not needed
# dataset = dataset.repeat()
return dataset
# 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):
classifier.train(input_fn=train_input_fn, hooks=train_hooks)
eval_results = 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, 3, 48, 48])
input_fn = tf.estimator.export.build_raw_serving_input_receiver_fn({
'image': image,
})
classifier.export_savedmodel(flags_obj.export_dir, input_fn,
strip_default_attrs=True)
def run_wide_deep(flags_obj):
"""Run Wide-Deep training and eval loop.
Args:
flags_obj: An object containing parsed flag values.
"""
# Clean up the model directory if present
shutil.rmtree(flags_obj.model_dir, ignore_errors=True)
model = build_estimator(flags_obj.model_dir, flags_obj.model_type)
train_file = os.path.join(flags_obj.data_dir, 'adult.data')
test_file = os.path.join(flags_obj.data_dir, 'adult.test')
# Train and evaluate the model every `flags.epochs_between_evals` epochs.
def train_input_fn():
return input_fn(train_file, flags_obj.epochs_between_evals, True,
flags_obj.batch_size)
def eval_input_fn():
return input_fn(test_file, 1, False, flags_obj.batch_size)
run_params = {
'batch_size': flags_obj.batch_size,
'train_epochs': flags_obj.train_epochs,
'model_type': flags_obj.model_type,
}
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info('wide_deep',
'Census Income',
run_params,
test_id=flags_obj.benchmark_test_id)
loss_prefix = LOSS_PREFIX.get(flags_obj.model_type, '')
train_hooks = hooks_helper.get_train_hooks(
flags_obj.hooks,
batch_size=flags_obj.batch_size,
tensors_to_log={
'average_loss': loss_prefix + 'head/truediv',
'loss': loss_prefix + 'head/weighted_loss/Sum'
})
# Train and evaluate the model every `flags.epochs_between_evals` epochs.
for n in range(flags_obj.train_epochs // flags_obj.epochs_between_evals):
model.train(input_fn=train_input_fn, hooks=train_hooks)
results = model.evaluate(input_fn=eval_input_fn)
# Display evaluation metrics
tf.logging.info('Results at epoch %d / %d',
(n + 1) * flags_obj.epochs_between_evals,
flags_obj.train_epochs)
tf.logging.info('-' * 60)
for key in sorted(results):
tf.logging.info('%s: %s' % (key, results[key]))
benchmark_logger.log_evaluation_result(results)
if model_helpers.past_stop_threshold(flags_obj.stop_threshold,
results['accuracy']):
break
# Export the model
if flags_obj.export_dir is not None:
export_model(model, flags_obj.model_type, flags_obj.export_dir)
def test_raise_in_invalid_names(self):
invalid_names = ['StepCounterHook', 'StopAtStepHook']
with self.assertRaises(ValueError):
hooks_helper.get_train_hooks(invalid_names, batch_size=256)
def net_main(flags_obj,
model_function,
input_function,
net_data_configs,
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)
is_metriclog = True
if is_metriclog:
metric_logfn = os.path.join(flags_obj.model_dir, 'log_metric.txt')
metric_logf = open(metric_logfn, 'a')
from tensorflow.contrib.memory_stats.ops import gen_memory_stats_ops
max_memory_usage = gen_memory_stats_ops.max_bytes_in_use()
# 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={
'data_format': flags_obj.data_format,
'batch_size': flags_obj.batch_size,
'loss_scale': flags_core.get_loss_scale(flags_obj),
'weight_decay': flags_obj.weight_decay,
'dtype': flags_core.get_tf_dtype(flags_obj),
'fine_tune': flags_obj.fine_tune,
'examples_per_epoch': flags_obj.examples_per_epoch,
'net_data_configs': net_data_configs
})
run_params = {
'batch_size': flags_obj.batch_size,
'dtype': flags_core.get_tf_dtype(flags_obj),
'synthetic_data': flags_obj.use_synthetic_data,
'train_epochs': flags_obj.train_epochs,
}
dataset_name = net_data_configs['dataset_name']
if flags_obj.use_synthetic_data:
dataset_name = dataset_name + '-synthetic'
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info('meshnet',
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),
examples_per_epoch=flags_obj.examples_per_epoch,
sg_settings=net_data_configs['sg_settings'])
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,
sg_settings=net_data_configs['sg_settings'])
if flags_obj.eval_only or flags_obj.pred_ply 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.
classifier.train(input_fn=lambda: input_fn_train(1),
hooks=train_hooks,
max_steps=10)
with tf.Session() as sess:
max_memory_usage_v = sess.run(max_memory_usage)
tf.logging.info('\n\nmemory usage: %0.3f G\n\n' %
(max_memory_usage_v * 1.0 / 1e9))
best_acc, best_acc_checkpoint = load_saved_best(flags_obj.model_dir)
for cycle_index, num_train_epochs in enumerate(schedule):
tf.logging.info('Starting cycle: %d/%d', cycle_index, int(n_loops))
t0 = time.time()
train_t = 0
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)
train_t = (time.time() - t0) / num_train_epochs
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.
only_train = False and (not flags_obj.eval_only) and (
not flags_obj.pred_ply)
if not only_train:
t0 = time.time()
eval_results = classifier.evaluate(
input_fn=input_fn_eval,
steps=flags_obj.max_train_steps,
)
#checkpoint_path=best_acc_checkpoint)
eval_t = time.time() - t0
if flags_obj.pred_ply:
pred_generator = classifier.predict(input_fn=input_fn_eval)
num_classes = net_data_configs['dset_metas'].num_classes
gen_pred_ply(eval_results, pred_generator, flags_obj.model_dir,
num_classes)
benchmark_logger.log_evaluation_result(eval_results)
if model_helpers.past_stop_threshold(flags_obj.stop_threshold,
eval_results['accuracy']):
break
cur_is_best = ''
if num_train_epochs and eval_results['accuracy'] > best_acc:
best_acc = eval_results['accuracy']
save_cur_model_as_best_acc(flags_obj.model_dir, best_acc)
cur_is_best = 'best'
global_step = cur_global_step(flags_obj.model_dir)
epoch = int(global_step / flags_obj.examples_per_epoch *
flags_obj.num_gpus)
ious_str = get_ious_str(eval_results['cm'],
net_data_configs['dset_metas'],
eval_results['mean_iou'])
metric_logf.write(
'\n{} train t:{:.1f} eval t:{:.1f} \teval acc:{:.3f} \tmean_iou:{:.3f} {} {}\n'
.format(epoch, train_t, eval_t, eval_results['accuracy'],
eval_results['mean_iou'], cur_is_best, ious_str))
metric_logf.flush()
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."""
# 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.
"""
# 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,
}
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=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 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 = 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_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 resnet_main(flags_obj, model_function, input_function, 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.
shape: list of ints representing the shape of the images used for training.
This is only used if flags.export_dir is passed.
"""
# Using the Winograd non-fused algorithms provides a small performance boost.
os.environ['TF_ENABLE_WINOGRAD_NONFUSED'] = '1'
if flags_obj.multi_gpu:
validate_batch_size_for_multi_gpu(flags_obj.batch_size)
# 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_function, loss_reduction=tf.losses.Reduction.MEAN)
# 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)
# Set up a RunConfig to save checkpoint and set session config.
run_config = tf.estimator.RunConfig().replace(
save_checkpoints_secs=1e9, 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,
'multi_gpu': flags_obj.multi_gpu,
'version': int(flags_obj.version),
'loss_scale': flags_core.get_loss_scale(flags_obj),
'dtype': flags_core.get_tf_dtype(flags_obj)
})
benchmark_logger = logger.config_benchmark_logger(
flags_obj.benchmark_log_dir)
benchmark_logger.log_run_info('resnet')
train_hooks = hooks_helper.get_train_hooks(
flags_obj.hooks,
batch_size=flags_obj.batch_size,
benchmark_log_dir=flags_obj.benchmark_log_dir)
def input_fn_train():
return input_function(True, flags_obj.data_dir, flags_obj.batch_size,
flags_obj.epochs_between_evals,
flags_obj.num_parallel_calls,
flags_obj.multi_gpu)
def input_fn_eval():
return input_function(False, flags_obj.data_dir, flags_obj.batch_size,
1, flags_obj.num_parallel_calls,
flags_obj.multi_gpu)
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.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.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:
warn_on_multi_gpu_export(flags_obj.multi_gpu)
# 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 model_fn(features, mode, params):
# 기존 dict 의 key,value 에 사용자 입력 값을 추가함
# extend dict values to defaultdict
_params = Transformer_params.copy()
for k in params:
v = params[k]
_params[k] = v
params = _params
if mode == tf.estimator.ModeKeys.PREDICT: features['answer'] = None
# define transformer
transformer = Transformer(params, (mode == tf.estimator.ModeKeys.TRAIN))
logits = transformer(features['question'], features['answer'])
if mode == tf.estimator.ModeKeys.TRAIN or mode == tf.estimator.ModeKeys.EVAL:
# 네트워크 출력 logits 와 실제 answer 간의 loss 를 계산
xentropy, weights = metrics.padded_cross_entropy_loss(
logits, features['answer'], params["label_smoothing"],
params["vocab_size"])
loss = tf.reduce_sum(xentropy) / tf.reduce_sum(weights)
# loss 를 minimize
learning_rate = get_learning_rate(params['learning_rate'],
params['hidden_size'],
params['learning_rate_warmup_steps'])
optimizer = tf.train.AdamOptimizer(
learning_rate=learning_rate,
beta1=params['optimizer_adam_beta1'],
beta2=params['optimizer_adam_beta2'],
epsilon=params['optimizer_adam_epsilon'])
train_op = optimizer.minimize(loss,
global_step=tf.train.get_global_step())
# 매 100번 마다 logitmax 과 answer 값을 보여줌
logging_hook = tf.train.LoggingTensorHook(
{
"logitmax": tf.argmax(logits[0], -1),
"answer": features['answer'][0]
},
every_n_iter=100)
# 여러가지 metric 을 계산하여 보여줌 (accuracy, BLEU score, ..)
eval_metric_ops = metrics.get_eval_metrics(logits, features['answer'],
params)
tensors_to_log = {}
for k in eval_metric_ops:
tensors_to_log[k.split('/')[-1]] = eval_metric_ops[k][1].name
tf.summary.scalar(k.split('/')[-1], eval_metric_ops[k][1])
tensors_to_log = {'learning_rate': learning_rate}
tf.summary.scalar('learning_rate', learning_rate)
train_hooks = hooks_helper.get_train_hooks(
['LoggingTensorHook'],
model_dir=params['model_dir'],
tensors_to_log=tensors_to_log,
batch_size=params['batch_size'],
use_tpu=params["use_tpu"])
# train
if mode == tf.estimator.ModeKeys.TRAIN:
return tf.estimator.EstimatorSpec(mode,
loss=loss,
train_op=train_op,
predictions=logits,
training_hooks=[logging_hook] +
train_hooks,
eval_metric_ops=eval_metric_ops)
# evaluate
elif mode == tf.estimator.ModeKeys.EVAL:
return tf.estimator.EstimatorSpec(mode,
loss=loss,
predictions=logits,
eval_metric_ops=eval_metric_ops)
# predict
else:
# predict 시에도 summary 저장
summary_hook = tf.train.SummarySaverHook(
save_secs=1000,
output_dir='./output/ckpt/pred',
scaffold=tf.train.Scaffold(summary_op=tf.summary.merge_all()))
return tf.estimator.EstimatorSpec(mode,
predictions=logits,
prediction_hooks=[summary_hook])
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_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_mnist(flags_obj):
model_helpers.apply_clean(flags_obj)
model_function = model_fn
# 通过命令 "with tf.device('/cpu:0'):",允许手动设置操作运行的设备。如果手动设置的设备不存在或者不可用，就会导致tf程序等待或异常
# allow_soft_placement, 自动选择可用设备。
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')
data_format = '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})
def train_input_fn():
ds = dataset.train(flags_obj.data_dir)
ds = ds.cache().shuffle(buffer_size=50000).batch(flags_obj.batch_size)
# repeat的功能就是将整个序列重复多次，主要用来处理机器学习中的epoch
ds = ds.repeat(flags_obj.epochs_between_evals)
return ds
def eval_input_fn():
# make_one_shot_iterator
# Creates an `Iterator` for enumerating the elements of this dataset
return dataset.test(flags_obj.data_dir).batch(
flags_obj.batch_size).make_one_shot_iterator().get_next()
train_hooks = hooks_helper.get_train_hooks(flags_obj.hooks,
model_dir=flags_obj.model_dir,
batch_size=flags_obj.batch_size)
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
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 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 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, 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()