def define_resnet_flags(resnet_size_choices=None,
dynamic_loss_scale=False,
fp16_implementation=False):
"""Add flags and validators for ResNet."""
flags_core.define_base(clean=True,
train_epochs=True,
epochs_between_evals=True,
stop_threshold=True,
num_gpu=True,
hooks=True,
export_dir=True,
distribution_strategy=True)
flags_core.define_performance(num_parallel_calls=False,
inter_op=True,
intra_op=True,
synthetic_data=True,
dtype=True,
all_reduce_alg=True,
num_packs=True,
tf_gpu_thread_mode=True,
datasets_num_private_threads=True,
dynamic_loss_scale=dynamic_loss_scale,
fp16_implementation=fp16_implementation,
loss_scale=True,
tf_data_experimental_slack=True,
max_train_steps=True)
flags_core.define_image()
flags_core.define_benchmark()
flags_core.define_distribution()
flags.adopt_module_key_flags(flags_core)
flags.DEFINE_enum(
name='resnet_version',
short_name='rv',
default='1',
enum_values=['1', '2'],
help=flags_core.help_wrap(
'Version of ResNet. (1 or 2) See README.md for details.'))
flags.DEFINE_bool(
name='fine_tune',
short_name='ft',
default=False,
help=flags_core.help_wrap(
'If True do not train any parameters except for the final layer.'))
flags.DEFINE_string(
name='pretrained_model_checkpoint_path',
short_name='pmcp',
default=None,
help=flags_core.help_wrap(
'If not None initialize all the network except the final layer with '
'these values'))
flags.DEFINE_boolean(name='eval_only',
default=False,
help=flags_core.help_wrap(
'Skip training and only perform evaluation on '
'the latest checkpoint.'))
flags.DEFINE_boolean(
name='image_bytes_as_serving_input',
default=False,
help=flags_core.help_wrap(
'If True exports savedmodel with serving signature that accepts '
'JPEG image bytes instead of a fixed size [HxWxC] tensor that '
'represents the image. The former is easier to use for serving at '
'the expense of image resize/cropping being done as part of model '
'inference. Note, this flag only applies to ImageNet and cannot '
'be used for CIFAR.'))
flags.DEFINE_boolean(
name='use_train_and_evaluate',
default=False,
help=flags_core.help_wrap(
'If True, uses `tf.estimator.train_and_evaluate` for the training '
'and evaluation loop, instead of separate calls to `classifier.train '
'and `classifier.evaluate`, which is the default behavior.'))
flags.DEFINE_bool(name='enable_lars',
default=False,
help=flags_core.help_wrap(
'Enable LARS optimizer for large batch training.'))
flags.DEFINE_float(
name='label_smoothing',
default=0.0,
help=flags_core.help_wrap(
'Label smoothing parameter used in the softmax_cross_entropy'))
flags.DEFINE_float(name='weight_decay',
default=1e-4,
help=flags_core.help_wrap(
'Weight decay coefficiant for l2 regularization.'))
flags.DEFINE_float(name='percent',
default=0,
help=flags_core.help_wrap('percent of data to poison'))
flags.DEFINE_bool(
name='adv_train',
default=False,
help=flags_core.help_wrap('whether adversarial training'))
choice_kwargs = dict(
name='resnet_size',
short_name='rs',
default='50',
help=flags_core.help_wrap('The size of the ResNet model to use.'))
if resnet_size_choices is None:
flags.DEFINE_string(**choice_kwargs)
else:
flags.DEFINE_enum(enum_values=resnet_size_choices, **choice_kwargs)
def define_keras_flags(dynamic_loss_scale=True):
"""Define flags for Keras models."""
flags_core.define_base(run_eagerly=True)
flags_core.define_performance(num_parallel_calls=False,
synthetic_data=True,
dtype=True,
all_reduce_alg=True,
num_packs=True,
tf_gpu_thread_mode=True,
datasets_num_private_threads=True,
dynamic_loss_scale=dynamic_loss_scale,
loss_scale=True,
tf_data_experimental_slack=True,
enable_xla=True,
force_v2_in_keras_compile=True)
flags_core.define_image()
flags_core.define_benchmark()
flags_core.define_distribution()
flags.adopt_module_key_flags(flags_core)
flags.DEFINE_boolean(name='enable_eager',
default=False,
help='Enable eager?')
flags.DEFINE_boolean(name='skip_eval',
default=False,
help='Skip evaluation?')
# TODO(b/135607288): Remove this flag once we understand the root cause of
# slowdown when setting the learning phase in Keras backend.
flags.DEFINE_boolean(
name='set_learning_phase_to_train',
default=True,
help='If skip eval, also set Keras learning phase to 1 (training).')
flags.DEFINE_boolean(
name='explicit_gpu_placement',
default=False,
help='If not using distribution strategy, explicitly set device scope '
'for the Keras training loop.')
flags.DEFINE_boolean(name='use_trivial_model',
default=False,
help='Whether to use a trivial Keras model.')
flags.DEFINE_boolean(name='report_accuracy_metrics',
default=True,
help='Report metrics during training and evaluation.')
flags.DEFINE_boolean(
name='use_tensor_lr',
default=False,
help='Use learning rate tensor instead of a callback.')
flags.DEFINE_boolean(name='enable_tensorboard',
default=False,
help='Whether to enable Tensorboard callback.')
flags.DEFINE_integer(
name='train_steps',
default=None,
help='The number of steps to run for training. If it is larger than '
'# batches per epoch, then use # batches per epoch. When this flag is '
'set, only one epoch is going to run for training.')
flags.DEFINE_string(
name='profile_steps',
default=None,
help='Save profiling data to model dir at given range of steps. The '
'value must be a comma separated pair of positive integers, specifying '
'the first and last step to profile. For example, "--profile_steps=2,4" '
'triggers the profiler to process 3 steps, starting from the 2nd step. '
'Note that profiler has a non-trivial performance overhead, and the '
'output file can be gigantic if profiling many steps.')
flags.DEFINE_boolean(
name='data_delay_prefetch',
default=False,
help=
'Add a small delay in tf.data prefetch to prioritize memory copy of '
'other tensors over the data minibatch for the (T+1)th step. It should '
'help improve performance using EagerIterator and function. The codepath '
'when enabling this feature is experimental and will be removed once the '
'corresponding performance features are fully supported in TensorFlow.'
)
flags.DEFINE_boolean(
name='batchnorm_spatial_persistent',
default=True,
help='Enable the spacial persistent mode for CuDNN batch norm kernel.')
flags.DEFINE_boolean(
name='enable_get_next_as_optional',
default=False,
help='Enable get_next_as_optional behavior in DistributedIterator.')
def define_keras_flags(dynamic_loss_scale=True,
model=False,
optimizer=False,
pretrained_filepath=False):
"""Define flags for Keras models."""
flags_core.define_base(clean=True,
num_gpu=True,
run_eagerly=True,
train_epochs=True,
epochs_between_evals=True,
distribution_strategy=True)
flags_core.define_performance(num_parallel_calls=False,
synthetic_data=True,
dtype=True,
all_reduce_alg=True,
num_packs=True,
tf_gpu_thread_mode=True,
datasets_num_private_threads=True,
dynamic_loss_scale=dynamic_loss_scale,
loss_scale=True,
fp16_implementation=True,
tf_data_experimental_slack=True,
enable_xla=True,
force_v2_in_keras_compile=True,
training_dataset_cache=True)
flags_core.define_image()
flags_core.define_benchmark()
flags_core.define_distribution()
flags.adopt_module_key_flags(flags_core)
flags.DEFINE_boolean(name='enable_eager',
default=False,
help='Enable eager?')
flags.DEFINE_boolean(name='skip_eval',
default=False,
help='Skip evaluation?')
# TODO(b/135607288): Remove this flag once we understand the root cause of
# slowdown when setting the learning phase in Keras backend.
flags.DEFINE_boolean(
name='set_learning_phase_to_train',
default=True,
help='If skip eval, also set Keras learning phase to 1 (training).')
flags.DEFINE_boolean(
name='explicit_gpu_placement',
default=False,
help='If not using distribution strategy, explicitly set device scope '
'for the Keras training loop.')
flags.DEFINE_boolean(name='use_trivial_model',
default=False,
help='Whether to use a trivial Keras model.')
flags.DEFINE_boolean(name='report_accuracy_metrics',
default=True,
help='Report metrics during training and evaluation.')
flags.DEFINE_boolean(
name='use_tensor_lr',
default=False,
help='Use learning rate tensor instead of a callback.')
flags.DEFINE_boolean(name='enable_tensorboard',
default=False,
help='Whether to enable Tensorboard callback.')
flags.DEFINE_integer(
name='train_steps',
default=None,
help='The number of steps to run for training. If it is larger than '
'# batches per epoch, then use # batches per epoch. This flag will be '
'ignored if train_epochs is set to be larger than 1. ')
flags.DEFINE_string(
name='profile_steps',
default=None,
help=
'Save profiling data to model dir at given range of global steps. The '
'value must be a comma separated pair of positive integers, specifying '
'the first and last step to profile. For example, "--profile_steps=2,4" '
'triggers the profiler to process 3 steps, starting from the 2nd step. '
'Note that profiler has a non-trivial performance overhead, and the '
'output file can be gigantic if profiling many steps.')
flags.DEFINE_boolean(
name='batchnorm_spatial_persistent',
default=True,
help='Enable the spacial persistent mode for CuDNN batch norm kernel.')
flags.DEFINE_boolean(
name='enable_get_next_as_optional',
default=False,
help='Enable get_next_as_optional behavior in DistributedIterator.')
flags.DEFINE_boolean(
name='enable_checkpoint_and_export',
default=False,
help=
'Whether to enable a checkpoint callback and export the savedmodel.')
flags.DEFINE_string(name='tpu',
default='',
help='TPU address to connect to.')
flags.DEFINE_integer(
name='steps_per_loop',
default=1,
help='Number of steps per graph-mode loop. Only training step happens '
'inside the loop. Callbacks will not be called inside. Will be capped at '
'steps per epoch.')
flags.DEFINE_boolean(
name='use_tf_keras_layers',
default=False,
help='Whether to use tf.keras.layers instead of tf.python.keras.layers.'
'It only changes imagenet resnet model layers for now. This flag is '
'a temporal flag during transition to tf.keras.layers. Do not use this '
'flag for external usage. this will be removed shortly.')
if model:
flags.DEFINE_string(
'model', 'resnet50_v1.5',
'Name of model preset. (mobilenet, resnet50_v1.5)')
if optimizer:
flags.DEFINE_string(
'optimizer', 'resnet50_default', 'Name of optimizer preset. '
'(mobilenet_default, resnet50_default)')
if pretrained_filepath:
flags.DEFINE_string('pretrained_filepath', '', 'Pretrained file path.')
def define_common_bert_flags():
"""Define common flags for BERT tasks."""
flags_core.define_base(data_dir=False,
model_dir=True,
clean=False,
train_epochs=False,
epochs_between_evals=False,
stop_threshold=False,
batch_size=False,
num_gpu=True,
export_dir=False,
distribution_strategy=True,
run_eagerly=True)
flags_core.define_distribution()
flags.DEFINE_string('bert_config_file', None,
'Bert configuration file to define core bert layers.')
flags.DEFINE_string(
'model_export_path', None,
'Path to the directory, where trainined model will be '
'exported.')
flags.DEFINE_string('tpu', '', 'TPU address to connect to.')
flags.DEFINE_string(
'init_checkpoint', None,
'Initial checkpoint (usually from a pre-trained BERT model).')
flags.DEFINE_integer('num_train_epochs', 3,
'Total number of training epochs to perform.')
flags.DEFINE_integer(
'steps_per_loop', None,
'Number of steps per graph-mode loop. Only training step '
'happens inside the loop. Callbacks will not be called '
'inside. If not set the value will be configured depending on the '
'devices available.')
flags.DEFINE_float('learning_rate', 5e-5,
'The initial learning rate for Adam.')
flags.DEFINE_float('end_lr', 0.0,
'The end learning rate for learning rate decay.')
flags.DEFINE_string(
'optimizer_type', 'adamw',
'The type of optimizer to use for training (adamw|lamb)')
flags.DEFINE_boolean(
'scale_loss', False,
'Whether to divide the loss by number of replica inside the per-replica '
'loss function.')
flags.DEFINE_boolean(
'use_keras_compile_fit', False,
'If True, uses Keras compile/fit() API for training logic. Otherwise '
'use custom training loop.')
flags.DEFINE_string(
'hub_module_url', None, 'TF-Hub path/url to Bert module. '
'If specified, init_checkpoint flag should not be used.')
flags.DEFINE_bool(
'hub_module_trainable', True,
'True to make keras layers in the hub module trainable.')
flags_core.define_log_steps()
# Adds flags for mixed precision and multi-worker training.
flags_core.define_performance(
num_parallel_calls=False,
inter_op=False,
intra_op=False,
synthetic_data=False,
max_train_steps=False,
dtype=True,
dynamic_loss_scale=True,
loss_scale=True,
all_reduce_alg=True,
num_packs=False,
tf_gpu_thread_mode=True,
datasets_num_private_threads=True,
enable_xla=True,
fp16_implementation=True,
)
# Adds gin configuration flags.
hyperparams_flags.define_gin_flags()
def define_transformer_flags():
"""Add flags and flag validators for running transformer_main."""
# Add common flags (data_dir, model_dir, etc.).
flags_core.define_base(num_gpu=True, distribution_strategy=True)
flags_core.define_performance(
num_parallel_calls=True,
inter_op=False,
intra_op=False,
synthetic_data=True,
max_train_steps=False,
dtype=True,
loss_scale=True,
all_reduce_alg=True,
num_packs=True,
tf_gpu_thread_mode=True,
datasets_num_private_threads=True,
enable_xla=True,
force_v2_in_keras_compile=True,
fp16_implementation=True
)
# Additional performance flags
# TODO(b/76028325): Remove when generic layout optimizer is ready.
flags.DEFINE_boolean(
name='enable_grappler_layout_optimizer',
default=True,
help='Enable Grappler layout optimizer. Currently Grappler can '
'de-optimize fp16 graphs by forcing NCHW layout for all '
'convolutions and batch normalizations, and this flag allows to '
'disable it.'
)
flags_core.define_benchmark()
flags_core.define_device(tpu=True)
flags_core.define_distribution()
flags.DEFINE_integer(
name='train_steps', short_name='ts', default=300000,
help=flags_core.help_wrap('The number of steps used to train.'))
flags.DEFINE_integer(
name='steps_between_evals', short_name='sbe', default=1000,
help=flags_core.help_wrap(
'The Number of training steps to run between evaluations. This is '
'used if --train_steps is defined.'))
flags.DEFINE_boolean(
name='enable_time_history', default=True,
help='Whether to enable TimeHistory callback.')
flags.DEFINE_boolean(
name='enable_tensorboard', default=False,
help='Whether to enable Tensorboard callback.')
flags.DEFINE_boolean(
name='enable_metrics_in_training', default=False,
help='Whether to enable metrics during training.')
flags.DEFINE_string(
name='profile_steps', default=None,
help='Save profiling data to model dir at given range of steps. The '
'value must be a comma separated pair of positive integers, specifying '
'the first and last step to profile. For example, "--profile_steps=2,4" '
'triggers the profiler to process 3 steps, starting from the 2nd step. '
'Note that profiler has a non-trivial performance overhead, and the '
'output file can be gigantic if profiling many steps.')
# Set flags from the flags_core module as 'key flags' so they're listed when
# the '-h' flag is used. Without this line, the flags defined above are
# only shown in the full `--helpful` help text.
flags.adopt_module_key_flags(flags_core)
# Add transformer-specific flags
flags.DEFINE_enum(
name='param_set', short_name='mp', default='big',
enum_values=PARAMS_MAP.keys(),
help=flags_core.help_wrap(
'Parameter set to use when creating and training the model. The '
'parameters define the input shape (batch size and max length), '
'model configuration (size of embedding, # of hidden layers, etc.), '
'and various other settings. The big parameter set increases the '
'default batch size, embedding/hidden size, and filter size. For a '
'complete list of parameters, please see model/model_params.py.'))
flags.DEFINE_bool(
name='static_batch', short_name='sb', default=False,
help=flags_core.help_wrap(
'Whether the batches in the dataset should have static shapes. In '
'general, this setting should be False. Dynamic shapes allow the '
'inputs to be grouped so that the number of padding tokens is '
'minimized, and helps model training. In cases where the input shape '
'must be static (e.g. running on TPU), this setting will be ignored '
'and static batching will always be used.'))
flags.DEFINE_integer(
name='max_length', short_name='ml', default=256,
help=flags_core.help_wrap(
'Max sentence length for Transformer. Default is 256. Note: Usually '
'it is more effective to use a smaller max length if static_batch is '
'enabled, e.g. 64.'))
# Flags for training with steps (may be used for debugging)
flags.DEFINE_integer(
name='validation_steps', short_name='vs', default=64,
help=flags_core.help_wrap('The number of steps used in validation.'))
# BLEU score computation
flags.DEFINE_string(
name='bleu_source', short_name='bls', default=None,
help=flags_core.help_wrap(
'Path to source file containing text translate when calculating the '
'official BLEU score. Both --bleu_source and --bleu_ref must be set. '
))
flags.DEFINE_string(
name='bleu_ref', short_name='blr', default=None,
help=flags_core.help_wrap(
'Path to source file containing text translate when calculating the '
'official BLEU score. Both --bleu_source and --bleu_ref must be set. '
))
flags.DEFINE_string(
name='vocab_file', short_name='vf', default=None,
help=flags_core.help_wrap(
'Path to subtoken vocabulary file. If data_download.py was used to '
'download and encode the training data, look in the data_dir to find '
'the vocab file.'))
flags.DEFINE_string(
name='mode', default='train',
help=flags_core.help_wrap('mode: train, eval, or predict'))
flags.DEFINE_bool(
name='use_ctl',
default=False,
help=flags_core.help_wrap(
'Whether the model runs with custom training loop.'))
flags.DEFINE_integer(
name='decode_batch_size',
default=32,
help=flags_core.help_wrap(
'Global batch size used for Transformer autoregressive decoding on '
'TPU.'))
flags.DEFINE_integer(
name='decode_max_length',
default=97,
help=flags_core.help_wrap(
'Max sequence length of the decode/eval data. This is used by '
'Transformer autoregressive decoding on TPU to have minimum '
'paddings.'))
flags.DEFINE_bool(
name='padded_decode',
default=False,
help=flags_core.help_wrap(
'Whether the autoregressive decoding runs with input data padded to '
'the decode_max_length. For TPU/XLA-GPU runs, this flag has to be '
'set due the static shape requirement. Although CPU/GPU could also '
'use padded_decode, it has not been tested. In addition, this method '
'will introduce unnecessary overheads which grow quadratically with '
'the max sequence length.'))
flags_core.set_defaults(data_dir='/tmp/translate_ende',
model_dir='/tmp/transformer_model',
batch_size=None)
# pylint: disable=unused-variable
@flags.multi_flags_validator(
['bleu_source', 'bleu_ref'],
message='Both or neither --bleu_source and --bleu_ref must be defined.')
def _check_bleu_files(flags_dict):
return (flags_dict['bleu_source'] is None) == (
flags_dict['bleu_ref'] is None)
@flags.multi_flags_validator(
['bleu_source', 'bleu_ref', 'vocab_file'],
message='--vocab_file must be defined if --bleu_source and --bleu_ref '
'are defined.')
def _check_bleu_vocab_file(flags_dict):
if flags_dict['bleu_source'] and flags_dict['bleu_ref']:
return flags_dict['vocab_file'] is not None
return True
def define_keras_flags(dynamic_loss_scale=True):
"""Define flags for Keras models."""
flags_core.define_base(
clean=True,
num_gpu=True,
run_eagerly=True,
train_epochs=True,
epochs_between_evals=True,
distribution_strategy=True,
)
flags_core.define_performance(
num_parallel_calls=False,
synthetic_data=True,
dtype=True,
all_reduce_alg=True,
num_packs=True,
tf_gpu_thread_mode=True,
datasets_num_private_threads=True,
dynamic_loss_scale=dynamic_loss_scale,
loss_scale=True,
fp16_implementation=True,
tf_data_experimental_slack=True,
enable_xla=True,
force_v2_in_keras_compile=True,
training_dataset_cache=True,
)
flags_core.define_image()
flags_core.define_benchmark()
flags_core.define_distribution()
flags.adopt_module_key_flags(flags_core)
flags.DEFINE_boolean(name="enable_eager",
default=False,
help="Enable eager?")
flags.DEFINE_boolean(name="skip_eval",
default=False,
help="Skip evaluation?")
# TODO(b/135607288): Remove this flag once we understand the root cause of
# slowdown when setting the learning phase in Keras backend.
flags.DEFINE_boolean(
name="set_learning_phase_to_train",
default=True,
help="If skip eval, also set Keras learning phase to 1 (training).",
)
flags.DEFINE_boolean(
name="explicit_gpu_placement",
default=False,
help="If not using distribution strategy, explicitly set device scope "
"for the Keras training loop.",
)
flags.DEFINE_boolean(name="use_trivial_model",
default=False,
help="Whether to use a trivial Keras model.")
flags.DEFINE_boolean(
name="report_accuracy_metrics",
default=True,
help="Report metrics during training and evaluation.",
)
flags.DEFINE_boolean(
name="use_tensor_lr",
default=False,
help="Use learning rate tensor instead of a callback.")
flags.DEFINE_boolean(name="enable_tensorboard",
default=False,
help="Whether to enable Tensorboard callback.")
flags.DEFINE_integer(
name="train_steps",
default=None,
help="The number of steps to run for training. If it is larger than "
"# batches per epoch, then use # batches per epoch. This flag will be "
"ignored if train_epochs is set to be larger than 1. ",
)
flags.DEFINE_string(
name="profile_steps",
default=None,
help="Save profiling data to model dir at given range of steps. The "
"value must be a comma separated pair of positive integers, specifying "
'the first and last step to profile. For example, "--profile_steps=2,4" '
"triggers the profiler to process 3 steps, starting from the 2nd step. "
"Note that profiler has a non-trivial performance overhead, and the "
"output file can be gigantic if profiling many steps.",
)
flags.DEFINE_boolean(
name="data_delay_prefetch",
default=False,
help=
"Add a small delay in tf.data prefetch to prioritize memory copy of "
"other tensors over the data minibatch for the (T+1)th step. It should "
"help improve performance using EagerIterator and function. The codepath "
"when enabling this feature is experimental and will be removed once the "
"corresponding performance features are fully supported in TensorFlow.",
)
flags.DEFINE_boolean(
name="batchnorm_spatial_persistent",
default=True,
help="Enable the spacial persistent mode for CuDNN batch norm kernel.",
)
flags.DEFINE_boolean(
name="enable_get_next_as_optional",
default=False,
help="Enable get_next_as_optional behavior in DistributedIterator.",
)
flags.DEFINE_boolean(
name="enable_checkpoint_and_export",
default=False,
help=
"Whether to enable a checkpoint callback and export the savedmodel.",
)
flags.DEFINE_string(name="tpu",
default="",
help="TPU address to connect to.")
flags.DEFINE_integer(
name="steps_per_loop",
default=1,
help="Number of steps per graph-mode loop. Only training step happens "
"inside the loop. Callbacks will not be called inside. Will be capped at "
"steps per epoch.",
)
def define_common_bert_flags():
"""Define common flags for BERT tasks."""
flags_core.define_base(
data_dir=False,
model_dir=True,
clean=False,
train_epochs=False,
epochs_between_evals=False,
stop_threshold=False,
batch_size=False,
num_gpu=True,
hooks=False,
export_dir=False,
distribution_strategy=True,
run_eagerly=True)
flags_core.define_distribution()
flags.DEFINE_string('bert_config_file', None,
'Bert configuration file to define core bert layers.')
flags.DEFINE_string(
'model_export_path', None,
'Path to the directory, where trainined model will be '
'exported.')
flags.DEFINE_string('tpu', '', 'TPU address to connect to.')
flags.DEFINE_string(
'init_checkpoint', None,
'Initial checkpoint (usually from a pre-trained BERT model).')
flags.DEFINE_integer('num_train_epochs', 3,
'Total number of training epochs to perform.')
flags.DEFINE_integer(
'steps_per_loop', 200,
'Number of steps per graph-mode loop. Only training step '
'happens inside the loop. Callbacks will not be called '
'inside.')
flags.DEFINE_float('learning_rate', 5e-5,
'The initial learning rate for Adam.')
flags.DEFINE_boolean(
'scale_loss', False,
'Whether to divide the loss by number of replica inside the per-replica '
'loss function.')
flags.DEFINE_boolean(
'use_keras_compile_fit', False,
'If True, uses Keras compile/fit() API for training logic. Otherwise '
'use custom training loop.')
flags.DEFINE_string(
'hub_module_url', None, 'TF-Hub path/url to Bert module. '
'If specified, init_checkpoint flag should not be used.')
flags.DEFINE_enum(
'model_type', 'bert', ['bert', 'albert'],
'Specifies the type of the model. '
'If "bert", will use canonical BERT; if "albert", will use ALBERT model.')
# Adds flags for mixed precision training.
flags_core.define_performance(
num_parallel_calls=False,
inter_op=False,
intra_op=False,
synthetic_data=False,
max_train_steps=False,
dtype=True,
dynamic_loss_scale=True,
loss_scale=True,
all_reduce_alg=False,
num_packs=False,
enable_xla=True,
fp16_implementation=True,
)
def define_keras_flags(dynamic_loss_scale=True,
model=False,
optimizer=False,
pretrained_filepath=False):
"""Define flags for Keras models."""
flags_core.define_base(clean=True,
num_gpu=True,
run_eagerly=True,
train_epochs=True,
epochs_between_evals=True,
distribution_strategy=True)
flags_core.define_performance(num_parallel_calls=False,
synthetic_data=True,
dtype=True,
all_reduce_alg=True,
num_packs=True,
tf_gpu_thread_mode=True,
datasets_num_private_threads=True,
dynamic_loss_scale=dynamic_loss_scale,
loss_scale=True,
fp16_implementation=True,
tf_data_experimental_slack=True,
enable_xla=True,
training_dataset_cache=True)
flags_core.define_image()
flags_core.define_benchmark()
flags_core.define_distribution()
flags.adopt_module_key_flags(flags_core)
flags.DEFINE_boolean(name='enable_eager',
default=False,
help='Enable eager?')
flags.DEFINE_boolean(name='skip_eval',
default=False,
help='Skip evaluation?')
# TODO(b/135607288): Remove this flag once we understand the root cause of
# slowdown when setting the learning phase in Keras backend.
flags.DEFINE_boolean(
name='set_learning_phase_to_train',
default=True,
help='If skip eval, also set Keras learning phase to 1 (training).')
flags.DEFINE_boolean(
name='explicit_gpu_placement',
default=False,
help='If not using distribution strategy, explicitly set device scope '
'for the Keras training loop.')
flags.DEFINE_boolean(name='use_trivial_model',
default=False,
help='Whether to use a trivial Keras model.')
flags.DEFINE_boolean(name='report_accuracy_metrics',
default=True,
help='Report metrics during training and evaluation.')
flags.DEFINE_boolean(
name='use_tensor_lr',
default=True,
help='Use learning rate tensor instead of a callback.')
flags.DEFINE_boolean(name='enable_tensorboard',
default=False,
help='Whether to enable Tensorboard callback.')
flags.DEFINE_integer(
name='train_steps',
default=None,
help='The number of steps to run for training. If it is larger than '
'# batches per epoch, then use # batches per epoch. This flag will be '
'ignored if train_epochs is set to be larger than 1. ')
flags.DEFINE_boolean(
name='batchnorm_spatial_persistent',
default=True,
help='Enable the spacial persistent mode for CuDNN batch norm kernel.')
flags.DEFINE_boolean(
name='enable_get_next_as_optional',
default=False,
help='Enable get_next_as_optional behavior in DistributedIterator.')
flags.DEFINE_boolean(
name='enable_checkpoint_and_export',
default=False,
help=
'Whether to enable a checkpoint callback and export the savedmodel.')
flags.DEFINE_string(name='tpu',
default='',
help='TPU address to connect to.')
flags.DEFINE_integer(
name='steps_per_loop',
default=500,
help='Number of steps per training loop. Only training step happens '
'inside the loop. Callbacks will not be called inside. Will be capped at '
'steps per epoch.')
flags.DEFINE_boolean(
name='use_tf_while_loop',
default=True,
help='Whether to build a tf.while_loop inside the training loop on the '
'host. Setting it to True is critical to have peak performance on '
'TPU.')
if model:
flags.DEFINE_string(
'model', 'resnet50_v1.5',
'Name of model preset. (mobilenet, resnet50_v1.5)')
if optimizer:
flags.DEFINE_string(
'optimizer', 'resnet50_default', 'Name of optimizer preset. '
'(mobilenet_default, resnet50_default)')
# TODO(kimjaehong): Replace as general hyper-params not only for mobilenet.
flags.DEFINE_float(
'initial_learning_rate_per_sample', 0.00007,
'Initial value of learning rate per sample for '
'mobilenet_default.')
flags.DEFINE_float(
'lr_decay_factor', 0.94,
'Learning rate decay factor for mobilenet_default.')
flags.DEFINE_float(
'num_epochs_per_decay', 2.5,
'Number of epochs per decay for mobilenet_default.')
if pretrained_filepath:
flags.DEFINE_string('pretrained_filepath', '', 'Pretrained file path.')