def export_tfhub(model_path, hub_destination): """Restores a tf.keras.Model and saves for TF-Hub.""" model = resnet_model.resnet50(num_classes=imagenet_preprocessing.NUM_CLASSES) model.load_weights(model_path) # Extracts a sub-model to use pooling feature vector as model output. image_input = model.get_layer(index=0).get_output_at(0) feature_vector_output = model.get_layer(name='reduce_mean').get_output_at(0) hub_model = tf.keras.Model(image_input, feature_vector_output) # Exports a SavedModel. hub_model.save(hub_destination, include_optimizer=False)
def dump_resnet50_v1():
root = "/tmp/tfmodels"
model = resm.resnet50(1000)
# model.compile(optimizer=tf.keras.optimizers.Adam(),
# loss=tf.keras.losses.sparse_categorical_crossentropy,
# metrics=['accuracy'])
os.makedirs(root, exist_ok=True)
model.save(os.path.join(root, "keras.h5"))
# tf.contrib.saved_model.save_keras_model(model, root,
# serving_only=False)
K.set_learning_phase(0)
model = load_model(os.path.join(root, "keras.h5"))
print("DDDDDDD", model.outputs)
frozen_graph = freeze_session(K.get_session(),
output_names=[out.op.name for out in model.outputs])
# with K.get_session() as sess:
# frozen_graph = convert_variables_to_constants(sess, sess.graph_def,
# [x.op.name for x in model.outputs])
tf.train.write_graph(frozen_graph, root, "resnet50_v1.pb", as_text=False)
def run(flags_obj):
"""Run ResNet ImageNet training and eval loop using custom training loops.
Args:
flags_obj: An object containing parsed flag values.
Raises:
ValueError: If fp16 is passed as it is not currently supported.
Returns:
Dictionary of training and eval stats.
"""
keras_utils.set_session_config(enable_eager=flags_obj.enable_eager,
enable_xla=flags_obj.enable_xla)
# TODO(anj-s): Set data_format without using Keras.
data_format = flags_obj.data_format
if data_format is None:
data_format = ('channels_first'
if tf.test.is_built_with_cuda() else 'channels_last')
tf.keras.backend.set_image_data_format(data_format)
strategy = distribution_utils.get_distribution_strategy(
distribution_strategy=flags_obj.distribution_strategy,
num_gpus=flags_obj.num_gpus,
num_workers=distribution_utils.configure_cluster(),
all_reduce_alg=flags_obj.all_reduce_alg,
num_packs=flags_obj.num_packs)
train_ds, test_ds = get_input_dataset(flags_obj, strategy)
train_steps, train_epochs, eval_steps = get_num_train_iterations(flags_obj)
time_callback = keras_utils.TimeHistory(flags_obj.batch_size,
flags_obj.log_steps)
strategy_scope = distribution_utils.get_strategy_scope(strategy)
with strategy_scope:
model = resnet_model.resnet50(
num_classes=imagenet_preprocessing.NUM_CLASSES,
batch_size=flags_obj.batch_size,
use_l2_regularizer=not flags_obj.single_l2_loss_op)
optimizer = tf.keras.optimizers.SGD(
learning_rate=common.BASE_LEARNING_RATE,
momentum=0.9,
nesterov=True)
if flags_obj.fp16_implementation == "graph_rewrite":
if not flags_obj.use_tf_function:
raise ValueError(
"--fp16_implementation=graph_rewrite requires "
"--use_tf_function to be true")
loss_scale = flags_core.get_loss_scale(flags_obj,
default_for_fp16=128)
optimizer = tf.train.experimental.enable_mixed_precision_graph_rewrite(
optimizer, loss_scale)
training_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(
'training_accuracy', dtype=tf.float32)
test_loss = tf.keras.metrics.Mean('test_loss', dtype=tf.float32)
test_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(
'test_accuracy', dtype=tf.float32)
trainable_variables = model.trainable_variables
def train_step(train_ds_inputs):
"""Training StepFn."""
def step_fn(inputs):
"""Per-Replica StepFn."""
images, labels = inputs
with tf.GradientTape() as tape:
logits = model(images, training=True)
prediction_loss = tf.keras.losses.sparse_categorical_crossentropy(
labels, logits)
loss = tf.reduce_sum(prediction_loss) * (
1.0 / flags_obj.batch_size)
num_replicas = tf.distribute.get_strategy(
).num_replicas_in_sync
if flags_obj.single_l2_loss_op:
filtered_variables = [
tf.reshape(v, (-1, )) for v in trainable_variables
if 'bn' not in v.name
]
l2_loss = resnet_model.L2_WEIGHT_DECAY * 2 * tf.nn.l2_loss(
tf.concat(filtered_variables, axis=0))
loss += (l2_loss / num_replicas)
else:
loss += (tf.reduce_sum(model.losses) / num_replicas)
# Scale the loss
if flags_obj.dtype == "fp16":
loss = optimizer.get_scaled_loss(loss)
grads = tape.gradient(loss, trainable_variables)
# Unscale the grads
if flags_obj.dtype == "fp16":
grads = optimizer.get_unscaled_gradients(grads)
optimizer.apply_gradients(zip(grads, trainable_variables))
training_accuracy.update_state(labels, logits)
return loss
if strategy:
per_replica_losses = strategy.experimental_run_v2(
step_fn, args=(train_ds_inputs, ))
return strategy.reduce(tf.distribute.ReduceOp.SUM,
per_replica_losses,
axis=None)
else:
return step_fn(train_ds_inputs)
def test_step(test_ds_inputs):
"""Evaluation StepFn."""
def step_fn(inputs):
images, labels = inputs
logits = model(images, training=False)
loss = tf.keras.losses.sparse_categorical_crossentropy(
labels, logits)
loss = tf.reduce_sum(loss) * (1.0 / flags_obj.batch_size)
test_loss.update_state(loss)
test_accuracy.update_state(labels, logits)
if strategy:
strategy.experimental_run_v2(step_fn, args=(test_ds_inputs, ))
else:
step_fn(test_ds_inputs)
if flags_obj.use_tf_function:
train_step = tf.function(train_step)
test_step = tf.function(test_step)
time_callback.on_train_begin()
for epoch in range(train_epochs):
train_iter = iter(train_ds)
total_loss = 0.0
training_accuracy.reset_states()
for step in range(train_steps):
optimizer.lr = common.learning_rate_schedule(
epoch, step, train_steps, flags_obj.batch_size)
time_callback.on_batch_begin(step + epoch * train_steps)
total_loss += train_step(next(train_iter))
time_callback.on_batch_end(step + epoch * train_steps)
train_loss = total_loss / train_steps
logging.info('Training loss: %s, accuracy: %s%% at epoch: %d',
train_loss.numpy(),
training_accuracy.result().numpy(), epoch)
if (not flags_obj.skip_eval
and (epoch + 1) % flags_obj.epochs_between_evals == 0):
test_loss.reset_states()
test_accuracy.reset_states()
test_iter = iter(test_ds)
for _ in range(eval_steps):
test_step(next(test_iter))
logging.info('Test loss: %s, accuracy: %s%% at epoch: %d',
test_loss.result().numpy(),
test_accuracy.result().numpy(), epoch)
time_callback.on_train_end()
eval_result = None
train_result = None
if not flags_obj.skip_eval:
eval_result = [
test_loss.result().numpy(),
test_accuracy.result().numpy()
]
train_result = [
train_loss.numpy(),
training_accuracy.result().numpy()
]
stats = build_stats(train_result, eval_result, time_callback)
return stats
def run(flags_obj):
"""Run ResNet ImageNet training and eval loop using native Keras APIs.
Args:
flags_obj: An object containing parsed flag values.
Raises:
ValueError: If fp16 is passed as it is not currently supported.
Returns:
Dictionary of training and eval stats.
"""
keras_utils.set_session_config(enable_eager=flags_obj.enable_eager,
enable_xla=flags_obj.enable_xla)
# Execute flag override logic for better model performance
if flags_obj.tf_gpu_thread_mode:
common.set_gpu_thread_mode_and_count(flags_obj)
if flags_obj.data_delay_prefetch:
common.data_delay_prefetch()
common.set_cudnn_batchnorm_mode()
dtype = flags_core.get_tf_dtype(flags_obj)
if dtype == 'float16':
policy = tf.keras.mixed_precision.experimental.Policy(
'infer_float32_vars')
tf.keras.mixed_precision.experimental.set_policy(policy)
data_format = flags_obj.data_format
if data_format is None:
data_format = ('channels_first'
if tf.test.is_built_with_cuda() else 'channels_last')
tf.keras.backend.set_image_data_format(data_format)
# Configures cluster spec for distribution strategy.
num_workers = distribution_utils.configure_cluster(flags_obj.worker_hosts,
flags_obj.task_index)
strategy = distribution_utils.get_distribution_strategy(
distribution_strategy=flags_obj.distribution_strategy,
num_gpus=flags_obj.num_gpus,
num_workers=num_workers,
all_reduce_alg=flags_obj.all_reduce_alg,
num_packs=flags_obj.num_packs)
if strategy:
# flags_obj.enable_get_next_as_optional controls whether enabling
# get_next_as_optional behavior in DistributedIterator. If true, last
# partial batch can be supported.
strategy.extended.experimental_enable_get_next_as_optional = (
flags_obj.enable_get_next_as_optional)
strategy_scope = distribution_utils.get_strategy_scope(strategy)
# pylint: disable=protected-access
if flags_obj.use_synthetic_data:
distribution_utils.set_up_synthetic_data()
input_fn = common.get_synth_input_fn(
height=imagenet_preprocessing.DEFAULT_IMAGE_SIZE,
width=imagenet_preprocessing.DEFAULT_IMAGE_SIZE,
num_channels=imagenet_preprocessing.NUM_CHANNELS,
num_classes=imagenet_preprocessing.NUM_CLASSES,
dtype=dtype,
drop_remainder=True)
else:
distribution_utils.undo_set_up_synthetic_data()
input_fn = imagenet_preprocessing.input_fn
# When `enable_xla` is True, we always drop the remainder of the batches
# in the dataset, as XLA-GPU doesn't support dynamic shapes.
drop_remainder = flags_obj.enable_xla
train_input_dataset = input_fn(
is_training=True,
data_dir=flags_obj.data_dir,
batch_size=flags_obj.batch_size,
num_epochs=flags_obj.train_epochs,
parse_record_fn=imagenet_preprocessing.parse_record,
datasets_num_private_threads=flags_obj.datasets_num_private_threads,
dtype=dtype,
drop_remainder=drop_remainder,
tf_data_experimental_slack=flags_obj.tf_data_experimental_slack,
)
eval_input_dataset = None
if not flags_obj.skip_eval:
eval_input_dataset = input_fn(
is_training=False,
data_dir=flags_obj.data_dir,
batch_size=flags_obj.batch_size,
num_epochs=flags_obj.train_epochs,
parse_record_fn=imagenet_preprocessing.parse_record,
dtype=dtype,
drop_remainder=drop_remainder)
lr_schedule = 0.1
if flags_obj.use_tensor_lr:
lr_schedule = common.PiecewiseConstantDecayWithWarmup(
batch_size=flags_obj.batch_size,
epoch_size=imagenet_preprocessing.NUM_IMAGES['train'],
warmup_epochs=LR_SCHEDULE[0][1],
boundaries=list(p[1] for p in LR_SCHEDULE[1:]),
multipliers=list(p[0] for p in LR_SCHEDULE),
compute_lr_on_cpu=True)
with strategy_scope:
optimizer = common.get_optimizer(lr_schedule)
if dtype == 'float16':
# TODO(reedwm): Remove manually wrapping optimizer once mixed precision
# can be enabled with a single line of code.
optimizer = tf.keras.mixed_precision.experimental.LossScaleOptimizer(
optimizer,
loss_scale=flags_core.get_loss_scale(flags_obj,
default_for_fp16=128))
if flags_obj.use_trivial_model:
model = trivial_model.trivial_model(
imagenet_preprocessing.NUM_CLASSES, dtype)
else:
model = resnet_model.resnet50(
num_classes=imagenet_preprocessing.NUM_CLASSES, dtype=dtype)
# TODO(b/138957587): Remove when force_v2_in_keras_compile is on longer
# a valid arg for this model. Also remove as a valid flag.
if flags_obj.force_v2_in_keras_compile is not None:
model.compile(
loss='sparse_categorical_crossentropy',
optimizer=optimizer,
metrics=(['sparse_categorical_accuracy']
if flags_obj.report_accuracy_metrics else None),
run_eagerly=flags_obj.run_eagerly,
experimental_run_tf_function=flags_obj.
force_v2_in_keras_compile)
else:
model.compile(
loss='sparse_categorical_crossentropy',
optimizer=optimizer,
metrics=(['sparse_categorical_accuracy']
if flags_obj.report_accuracy_metrics else None),
run_eagerly=flags_obj.run_eagerly)
callbacks = common.get_callbacks(
learning_rate_schedule, imagenet_preprocessing.NUM_IMAGES['train'])
train_steps = (imagenet_preprocessing.NUM_IMAGES['train'] //
flags_obj.batch_size)
train_epochs = flags_obj.train_epochs
if flags_obj.train_steps:
train_steps = min(flags_obj.train_steps, train_steps)
train_epochs = 1
num_eval_steps = (imagenet_preprocessing.NUM_IMAGES['validation'] //
flags_obj.batch_size)
validation_data = eval_input_dataset
if flags_obj.skip_eval:
# Only build the training graph. This reduces memory usage introduced by
# control flow ops in layers that have different implementations for
# training and inference (e.g., batch norm).
if flags_obj.set_learning_phase_to_train:
# TODO(haoyuzhang): Understand slowdown of setting learning phase when
# not using distribution strategy.
tf.keras.backend.set_learning_phase(1)
num_eval_steps = None
validation_data = None
if not strategy and flags_obj.explicit_gpu_placement:
# TODO(b/135607227): Add device scope automatically in Keras training loop
# when not using distribition strategy.
no_dist_strat_device = tf.device('/device:GPU:0')
no_dist_strat_device.__enter__()
history = model.fit(train_input_dataset,
epochs=train_epochs,
steps_per_epoch=train_steps // 15,
callbacks=callbacks,
validation_steps=num_eval_steps,
validation_data=validation_data,
validation_freq=flags_obj.epochs_between_evals,
verbose=1)
eval_output = None
if not flags_obj.skip_eval:
eval_output = model.evaluate(eval_input_dataset,
steps=num_eval_steps,
verbose=1)
if not strategy and flags_obj.explicit_gpu_placement:
no_dist_strat_device.__exit__()
stats = common.build_stats(history, eval_output, callbacks)
return stats
def run(flags_obj):
"""Run ResNet ImageNet training and eval loop using custom training loops.
Args:
flags_obj: An object containing parsed flag values.
Raises:
ValueError: If fp16 is passed as it is not currently supported.
Returns:
Dictionary of training and eval stats.
"""
keras_utils.set_session_config(enable_eager=flags_obj.enable_eager,
enable_xla=flags_obj.enable_xla)
dtype = flags_core.get_tf_dtype(flags_obj)
if dtype == tf.bfloat16:
policy = tf.compat.v2.keras.mixed_precision.experimental.Policy(
'mixed_bfloat16')
tf.compat.v2.keras.mixed_precision.experimental.set_policy(policy)
# TODO(anj-s): Set data_format without using Keras.
data_format = flags_obj.data_format
if data_format is None:
data_format = ('channels_first'
if tf.test.is_built_with_cuda() else 'channels_last')
tf.keras.backend.set_image_data_format(data_format)
strategy = distribution_utils.get_distribution_strategy(
distribution_strategy=flags_obj.distribution_strategy,
num_gpus=flags_obj.num_gpus,
num_workers=distribution_utils.configure_cluster(),
all_reduce_alg=flags_obj.all_reduce_alg,
num_packs=flags_obj.num_packs,
tpu_address=flags_obj.tpu)
train_ds, test_ds = get_input_dataset(flags_obj, strategy)
per_epoch_steps, train_epochs, eval_steps = get_num_train_iterations(
flags_obj)
steps_per_loop = min(flags_obj.steps_per_loop, per_epoch_steps)
logging.info(
"Training %d epochs, each epoch has %d steps, "
"total steps: %d; Eval %d steps", train_epochs, per_epoch_steps,
train_epochs * per_epoch_steps, eval_steps)
time_callback = keras_utils.TimeHistory(flags_obj.batch_size,
flags_obj.log_steps)
with distribution_utils.get_strategy_scope(strategy):
model = resnet_model.resnet50(
num_classes=imagenet_preprocessing.NUM_CLASSES,
batch_size=flags_obj.batch_size,
use_l2_regularizer=not flags_obj.single_l2_loss_op)
lr_schedule = common.PiecewiseConstantDecayWithWarmup(
batch_size=flags_obj.batch_size,
epoch_size=imagenet_preprocessing.NUM_IMAGES['train'],
warmup_epochs=common.LR_SCHEDULE[0][1],
boundaries=list(p[1] for p in common.LR_SCHEDULE[1:]),
multipliers=list(p[0] for p in common.LR_SCHEDULE),
compute_lr_on_cpu=True)
optimizer = common.get_optimizer(lr_schedule)
if flags_obj.fp16_implementation == 'graph_rewrite':
if not flags_obj.use_tf_function:
raise ValueError(
'--fp16_implementation=graph_rewrite requires '
'--use_tf_function to be true')
loss_scale = flags_core.get_loss_scale(flags_obj,
default_for_fp16=128)
optimizer = tf.train.experimental.enable_mixed_precision_graph_rewrite(
optimizer, loss_scale)
train_loss = tf.keras.metrics.Mean('train_loss', dtype=tf.float32)
training_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(
'training_accuracy', dtype=tf.float32)
test_loss = tf.keras.metrics.Mean('test_loss', dtype=tf.float32)
test_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(
'test_accuracy', dtype=tf.float32)
trainable_variables = model.trainable_variables
def step_fn(inputs):
"""Per-Replica StepFn."""
images, labels = inputs
with tf.GradientTape() as tape:
logits = model(images, training=True)
prediction_loss = tf.keras.losses.sparse_categorical_crossentropy(
labels, logits)
loss = tf.reduce_sum(prediction_loss) * (1.0 /
flags_obj.batch_size)
num_replicas = tf.distribute.get_strategy(
).num_replicas_in_sync
if flags_obj.single_l2_loss_op:
filtered_variables = [
tf.reshape(v, (-1, )) for v in trainable_variables
if 'bn' not in v.name
]
l2_loss = resnet_model.L2_WEIGHT_DECAY * 2 * tf.nn.l2_loss(
tf.concat(filtered_variables, axis=0))
loss += (l2_loss / num_replicas)
else:
loss += (tf.reduce_sum(model.losses) / num_replicas)
# Scale the loss
if flags_obj.dtype == "fp16":
loss = optimizer.get_scaled_loss(loss)
grads = tape.gradient(loss, trainable_variables)
# Unscale the grads
if flags_obj.dtype == "fp16":
grads = optimizer.get_unscaled_gradients(grads)
optimizer.apply_gradients(zip(grads, trainable_variables))
train_loss.update_state(loss)
training_accuracy.update_state(labels, logits)
@tf.function
def train_steps(iterator, steps):
"""Performs distributed training steps in a loop."""
for _ in tf.range(steps):
strategy.experimental_run_v2(step_fn, args=(next(iterator), ))
def train_single_step(iterator):
if strategy:
strategy.experimental_run_v2(step_fn, args=(next(iterator), ))
else:
return step_fn(next(iterator))
def test_step(iterator):
"""Evaluation StepFn."""
def step_fn(inputs):
images, labels = inputs
logits = model(images, training=False)
loss = tf.keras.losses.sparse_categorical_crossentropy(
labels, logits)
loss = tf.reduce_sum(loss) * (1.0 / flags_obj.batch_size)
test_loss.update_state(loss)
test_accuracy.update_state(labels, logits)
if strategy:
strategy.experimental_run_v2(step_fn, args=(next(iterator), ))
else:
step_fn(next(iterator))
if flags_obj.use_tf_function:
train_single_step = tf.function(train_single_step)
test_step = tf.function(test_step)
train_iter = iter(train_ds)
time_callback.on_train_begin()
for epoch in range(train_epochs):
train_loss.reset_states()
training_accuracy.reset_states()
steps_in_current_epoch = 0
while steps_in_current_epoch < per_epoch_steps:
time_callback.on_batch_begin(steps_in_current_epoch +
epoch * per_epoch_steps)
steps = _steps_to_run(steps_in_current_epoch, per_epoch_steps,
steps_per_loop)
if steps == 1:
train_single_step(train_iter)
else:
# Converts steps to a Tensor to avoid tf.function retracing.
train_steps(train_iter,
tf.convert_to_tensor(steps, dtype=tf.int32))
time_callback.on_batch_end(steps_in_current_epoch +
epoch * per_epoch_steps)
steps_in_current_epoch += steps
logging.info('Training loss: %s, accuracy: %s at epoch %d',
train_loss.result().numpy(),
training_accuracy.result().numpy(), epoch + 1)
if (not flags_obj.skip_eval
and (epoch + 1) % flags_obj.epochs_between_evals == 0):
test_loss.reset_states()
test_accuracy.reset_states()
test_iter = iter(test_ds)
for _ in range(eval_steps):
test_step(test_iter)
logging.info('Test loss: %s, accuracy: %s%% at epoch: %d',
test_loss.result().numpy(),
test_accuracy.result().numpy(), epoch + 1)
time_callback.on_train_end()
eval_result = None
train_result = None
if not flags_obj.skip_eval:
eval_result = [
test_loss.result().numpy(),
test_accuracy.result().numpy()
]
train_result = [
train_loss.result().numpy(),
training_accuracy.result().numpy()
]
stats = build_stats(train_result, eval_result, time_callback)
return stats
def __init__(self, flags_obj, time_callback, epoch_steps):
standard_runnable.StandardTrainable.__init__(
self, flags_obj.use_tf_while_loop, flags_obj.use_tf_function)
standard_runnable.StandardEvaluable.__init__(self,
flags_obj.use_tf_function)
self.strategy = tf.distribute.get_strategy()
self.flags_obj = flags_obj
self.dtype = flags_core.get_tf_dtype(flags_obj)
self.time_callback = time_callback
# Input pipeline related
batch_size = flags_obj.batch_size
if batch_size % self.strategy.num_replicas_in_sync != 0:
raise ValueError(
'Batch size must be divisible by number of replicas : {}'.
format(self.strategy.num_replicas_in_sync))
# As auto rebatching is not supported in
# `experimental_distribute_datasets_from_function()` API, which is
# required when cloning dataset to multiple workers in eager mode,
# we use per-replica batch size.
self.batch_size = int(batch_size / self.strategy.num_replicas_in_sync)
if self.flags_obj.use_synthetic_data:
self.input_fn = common.get_synth_input_fn(
height=imagenet_preprocessing.DEFAULT_IMAGE_SIZE,
width=imagenet_preprocessing.DEFAULT_IMAGE_SIZE,
num_channels=imagenet_preprocessing.NUM_CHANNELS,
num_classes=imagenet_preprocessing.NUM_CLASSES,
dtype=self.dtype,
drop_remainder=True)
else:
self.input_fn = imagenet_preprocessing.input_fn
resnet_model.change_keras_layer(flags_obj.use_tf_keras_layers)
self.model = resnet_model.resnet50(
num_classes=imagenet_preprocessing.NUM_CLASSES,
batch_size=flags_obj.batch_size,
use_l2_regularizer=not flags_obj.single_l2_loss_op)
lr_schedule = common.PiecewiseConstantDecayWithWarmup(
batch_size=flags_obj.batch_size,
epoch_size=imagenet_preprocessing.NUM_IMAGES['train'],
warmup_epochs=common.LR_SCHEDULE[0][1],
boundaries=list(p[1] for p in common.LR_SCHEDULE[1:]),
multipliers=list(p[0] for p in common.LR_SCHEDULE),
compute_lr_on_cpu=True)
self.optimizer = common.get_optimizer(lr_schedule)
# Make sure iterations variable is created inside scope.
self.global_step = self.optimizer.iterations
if self.dtype == tf.float16:
loss_scale = flags_core.get_loss_scale(flags_obj,
default_for_fp16=128)
self.optimizer = (
tf.keras.mixed_precision.experimental.LossScaleOptimizer(
self.optimizer, loss_scale))
elif flags_obj.fp16_implementation == 'graph_rewrite':
# `dtype` is still float32 in this case. We built the graph in float32
# and let the graph rewrite change parts of it float16.
if not flags_obj.use_tf_function:
raise ValueError(
'--fp16_implementation=graph_rewrite requires '
'--use_tf_function to be true')
loss_scale = flags_core.get_loss_scale(flags_obj,
default_for_fp16=128)
self.optimizer = (
tf.train.experimental.enable_mixed_precision_graph_rewrite(
self.optimizer, loss_scale))
self.train_loss = tf.keras.metrics.Mean('train_loss', dtype=tf.float32)
self.train_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(
'train_accuracy', dtype=tf.float32)
self.test_loss = tf.keras.metrics.Mean('test_loss', dtype=tf.float32)
self.test_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(
'test_accuracy', dtype=tf.float32)
self.checkpoint = tf.train.Checkpoint(model=self.model,
optimizer=self.optimizer)
# Handling epochs.
self.epoch_steps = epoch_steps
self.epoch_helper = utils.EpochHelper(epoch_steps, self.global_step)
def run(flags_obj):
"""Run ResNet ImageNet training and eval loop using native Keras APIs.
Args:
flags_obj: An object containing parsed flag values.
Raises:
ValueError: If fp16 is passed as it is not currently supported.
Returns:
Dictionary of training and eval stats.
"""
keras_utils.set_session_config(
enable_eager=flags_obj.enable_eager,
enable_xla=flags_obj.enable_xla)
# Execute flag override logic for better model performance
if flags_obj.tf_gpu_thread_mode:
common.set_gpu_thread_mode_and_count(flags_obj)
common.set_cudnn_batchnorm_mode()
dtype = flags_core.get_tf_dtype(flags_obj)
if dtype == tf.float16:
loss_scale = flags_core.get_loss_scale(flags_obj, default_for_fp16=128)
policy = tf.compat.v2.keras.mixed_precision.experimental.Policy(
'mixed_float16', loss_scale=loss_scale)
tf.compat.v2.keras.mixed_precision.experimental.set_policy(policy)
if not keras_utils.is_v2_0():
raise ValueError('--dtype=fp16 is not supported in TensorFlow 1.')
elif dtype == tf.bfloat16:
policy = tf.compat.v2.keras.mixed_precision.experimental.Policy(
'mixed_bfloat16')
tf.compat.v2.keras.mixed_precision.experimental.set_policy(policy)
data_format = flags_obj.data_format
if data_format is None:
data_format = ('channels_first'
if tf.test.is_built_with_cuda() else 'channels_last')
tf.keras.backend.set_image_data_format(data_format)
# Configures cluster spec for distribution strategy.
num_workers = distribution_utils.configure_cluster(flags_obj.worker_hosts,
flags_obj.task_index)
strategy = distribution_utils.get_distribution_strategy(
distribution_strategy=flags_obj.distribution_strategy,
num_gpus=flags_obj.num_gpus,
num_workers=num_workers,
all_reduce_alg=flags_obj.all_reduce_alg,
num_packs=flags_obj.num_packs,
tpu_address=flags_obj.tpu)
if strategy:
# flags_obj.enable_get_next_as_optional controls whether enabling
# get_next_as_optional behavior in DistributedIterator. If true, last
# partial batch can be supported.
strategy.extended.experimental_enable_get_next_as_optional = (
flags_obj.enable_get_next_as_optional
)
strategy_scope = distribution_utils.get_strategy_scope(strategy)
# pylint: disable=protected-access
if flags_obj.use_synthetic_data:
distribution_utils.set_up_synthetic_data()
input_fn = common.get_synth_input_fn(
height=imagenet_preprocessing.DEFAULT_IMAGE_SIZE,
width=imagenet_preprocessing.DEFAULT_IMAGE_SIZE,
num_channels=imagenet_preprocessing.NUM_CHANNELS,
num_classes=imagenet_preprocessing.NUM_CLASSES,
dtype=dtype,
drop_remainder=True)
else:
distribution_utils.undo_set_up_synthetic_data()
input_fn = imagenet_preprocessing.input_fn
# When `enable_xla` is True, we always drop the remainder of the batches
# in the dataset, as XLA-GPU doesn't support dynamic shapes.
drop_remainder = flags_obj.enable_xla
train_input_dataset = input_fn(
is_training=True,
data_dir=flags_obj.data_dir,
batch_size=flags_obj.batch_size,
num_epochs=flags_obj.train_epochs,
parse_record_fn=imagenet_preprocessing.parse_record,
datasets_num_private_threads=flags_obj.datasets_num_private_threads,
dtype=dtype,
drop_remainder=drop_remainder,
tf_data_experimental_slack=flags_obj.tf_data_experimental_slack,
training_dataset_cache=flags_obj.training_dataset_cache,
)
eval_input_dataset = None
if not flags_obj.skip_eval:
eval_input_dataset = input_fn(
is_training=False,
data_dir=flags_obj.data_dir,
batch_size=flags_obj.batch_size,
num_epochs=flags_obj.train_epochs,
parse_record_fn=imagenet_preprocessing.parse_record,
dtype=dtype,
drop_remainder=drop_remainder)
lr_schedule = 0.1
if flags_obj.use_tensor_lr:
lr_schedule = common.PiecewiseConstantDecayWithWarmup(
batch_size=flags_obj.batch_size,
epoch_size=imagenet_preprocessing.NUM_IMAGES['train'],
warmup_epochs=common.LR_SCHEDULE[0][1],
boundaries=list(p[1] for p in common.LR_SCHEDULE[1:]),
multipliers=list(p[0] for p in common.LR_SCHEDULE),
compute_lr_on_cpu=True)
with strategy_scope:
optimizer = common.get_optimizer(lr_schedule)
if flags_obj.fp16_implementation == 'graph_rewrite':
# Note: when flags_obj.fp16_implementation == "graph_rewrite", dtype as
# determined by flags_core.get_tf_dtype(flags_obj) would be 'float32'
# which will ensure tf.compat.v2.keras.mixed_precision and
# tf.train.experimental.enable_mixed_precision_graph_rewrite do not double
# up.
optimizer = tf.train.experimental.enable_mixed_precision_graph_rewrite(
optimizer)
# TODO(hongkuny): Remove trivial model usage and move it to benchmark.
if flags_obj.use_trivial_model:
model = trivial_model.trivial_model(
imagenet_preprocessing.NUM_CLASSES)
else:
model = resnet_model.resnet50(
num_classes=imagenet_preprocessing.NUM_CLASSES)
# TODO(b/138957587): Remove when force_v2_in_keras_compile is on longer
# a valid arg for this model. Also remove as a valid flag.
if flags_obj.force_v2_in_keras_compile is not None:
model.compile(
loss='sparse_categorical_crossentropy',
optimizer=optimizer,
metrics=(['sparse_categorical_accuracy']
if flags_obj.report_accuracy_metrics else None),
run_eagerly=flags_obj.run_eagerly,
experimental_run_tf_function=flags_obj.force_v2_in_keras_compile)
else:
model.compile(
loss='sparse_categorical_crossentropy',
optimizer=optimizer,
metrics=(['sparse_categorical_accuracy']
if flags_obj.report_accuracy_metrics else None),
run_eagerly=flags_obj.run_eagerly)
steps_per_epoch = (
imagenet_preprocessing.NUM_IMAGES['train'] // flags_obj.batch_size)
train_epochs = flags_obj.train_epochs
callbacks = common.get_callbacks(steps_per_epoch,
common.learning_rate_schedule)
if flags_obj.enable_checkpoint_and_export:
ckpt_full_path = os.path.join(flags_obj.model_dir, 'model.ckpt-{epoch:04d}')
callbacks.append(tf.keras.callbacks.ModelCheckpoint(ckpt_full_path,
save_weights_only=True))
# if mutliple epochs, ignore the train_steps flag.
if train_epochs <= 1 and flags_obj.train_steps:
steps_per_epoch = min(flags_obj.train_steps, steps_per_epoch)
train_epochs = 1
num_eval_steps = (
imagenet_preprocessing.NUM_IMAGES['validation'] // flags_obj.batch_size)
validation_data = eval_input_dataset
if flags_obj.skip_eval:
# Only build the training graph. This reduces memory usage introduced by
# control flow ops in layers that have different implementations for
# training and inference (e.g., batch norm).
if flags_obj.set_learning_phase_to_train:
# TODO(haoyuzhang): Understand slowdown of setting learning phase when
# not using distribution strategy.
tf.keras.backend.set_learning_phase(1)
num_eval_steps = None
validation_data = None
if not strategy and flags_obj.explicit_gpu_placement:
# TODO(b/135607227): Add device scope automatically in Keras training loop
# when not using distribition strategy.
no_dist_strat_device = tf.device('/device:GPU:0')
no_dist_strat_device.__enter__()
history = model.fit(train_input_dataset,
epochs=train_epochs,
steps_per_epoch=steps_per_epoch,
callbacks=callbacks,
validation_steps=num_eval_steps,
validation_data=validation_data,
validation_freq=flags_obj.epochs_between_evals,
verbose=2)
if flags_obj.enable_checkpoint_and_export:
if dtype == tf.bfloat16:
logging.warning("Keras model.save does not support bfloat16 dtype.")
else:
# Keras model.save assumes a float32 input designature.
export_path = os.path.join(flags_obj.model_dir, 'saved_model')
model.save(export_path, include_optimizer=False)
eval_output = None
if not flags_obj.skip_eval:
eval_output = model.evaluate(eval_input_dataset,
steps=num_eval_steps,
verbose=2)
if not strategy and flags_obj.explicit_gpu_placement:
no_dist_strat_device.__exit__()
stats = common.build_stats(history, eval_output, callbacks)
return stats
def run(flags_obj):
"""Run ResNet ImageNet training and eval loop using native Keras APIs.
Args:
flags_obj: An object containing parsed flag values.
Raises:
ValueError: If fp16 is passed as it is not currently supported.
NotImplementedError: If some features are not currently supported.
Returns:
Dictionary of training and eval stats.
"""
keras_utils.set_session_config(enable_eager=flags_obj.enable_eager,
enable_xla=flags_obj.enable_xla)
# Execute flag override logic for better model performance
if flags_obj.tf_gpu_thread_mode:
keras_utils.set_gpu_thread_mode_and_count(
per_gpu_thread_count=flags_obj.per_gpu_thread_count,
gpu_thread_mode=flags_obj.tf_gpu_thread_mode,
num_gpus=flags_obj.num_gpus,
datasets_num_private_threads=flags_obj.datasets_num_private_threads
)
common.set_cudnn_batchnorm_mode()
dtype = flags_core.get_tf_dtype(flags_obj)
performance.set_mixed_precision_policy(
flags_core.get_tf_dtype(flags_obj),
flags_core.get_loss_scale(flags_obj, default_for_fp16=128))
data_format = flags_obj.data_format
if data_format is None:
data_format = ('channels_first'
if tf.test.is_built_with_cuda() else 'channels_last')
tf.keras.backend.set_image_data_format(data_format)
# Configures cluster spec for distribution strategy.
_ = distribution_utils.configure_cluster(flags_obj.worker_hosts,
flags_obj.task_index)
strategy = distribution_utils.get_distribution_strategy(
distribution_strategy=flags_obj.distribution_strategy,
num_gpus=flags_obj.num_gpus,
all_reduce_alg=flags_obj.all_reduce_alg,
num_packs=flags_obj.num_packs,
tpu_address=flags_obj.tpu)
if strategy:
# flags_obj.enable_get_next_as_optional controls whether enabling
# get_next_as_optional behavior in DistributedIterator. If true, last
# partial batch can be supported.
strategy.extended.experimental_enable_get_next_as_optional = (
flags_obj.enable_get_next_as_optional)
strategy_scope = distribution_utils.get_strategy_scope(strategy)
# pylint: disable=protected-access
if flags_obj.use_synthetic_data:
distribution_utils.set_up_synthetic_data()
input_fn = common.get_synth_input_fn(
height=imagenet_preprocessing.DEFAULT_IMAGE_SIZE,
width=imagenet_preprocessing.DEFAULT_IMAGE_SIZE,
num_channels=imagenet_preprocessing.NUM_CHANNELS,
num_classes=imagenet_preprocessing.NUM_CLASSES,
dtype=dtype,
drop_remainder=True)
else:
distribution_utils.undo_set_up_synthetic_data()
input_fn = imagenet_preprocessing.input_fn
# When `enable_xla` is True, we always drop the remainder of the batches
# in the dataset, as XLA-GPU doesn't support dynamic shapes.
drop_remainder = flags_obj.enable_xla
# Current resnet_model.resnet50 input format is always channel-last.
# We use keras_application mobilenet model which input format is depends on
# the keras beckend image data format.
# This use_keras_image_data_format flags indicates whether image preprocessor
# output format should be same as the keras backend image data format or just
# channel-last format.
use_keras_image_data_format = (flags_obj.model == 'mobilenet')
train_input_dataset = input_fn(
is_training=True,
data_dir=flags_obj.data_dir,
batch_size=flags_obj.batch_size,
parse_record_fn=imagenet_preprocessing.get_parse_record_fn(
use_keras_image_data_format=use_keras_image_data_format),
datasets_num_private_threads=flags_obj.datasets_num_private_threads,
dtype=dtype,
drop_remainder=drop_remainder,
tf_data_experimental_slack=flags_obj.tf_data_experimental_slack,
training_dataset_cache=flags_obj.training_dataset_cache,
)
eval_input_dataset = None
if not flags_obj.skip_eval:
eval_input_dataset = input_fn(
is_training=False,
data_dir=flags_obj.data_dir,
batch_size=flags_obj.batch_size,
parse_record_fn=imagenet_preprocessing.get_parse_record_fn(
use_keras_image_data_format=use_keras_image_data_format),
dtype=dtype,
drop_remainder=drop_remainder)
lr_schedule = common.PiecewiseConstantDecayWithWarmup(
batch_size=flags_obj.batch_size,
epoch_size=imagenet_preprocessing.NUM_IMAGES['train'],
warmup_epochs=common.LR_SCHEDULE[0][1],
boundaries=list(p[1] for p in common.LR_SCHEDULE[1:]),
multipliers=list(p[0] for p in common.LR_SCHEDULE),
compute_lr_on_cpu=True)
steps_per_epoch = (imagenet_preprocessing.NUM_IMAGES['train'] //
flags_obj.batch_size)
with strategy_scope:
if flags_obj.optimizer == 'resnet50_default':
optimizer = common.get_optimizer(lr_schedule)
elif flags_obj.optimizer == 'mobilenet_default':
initial_learning_rate = \
flags_obj.initial_learning_rate_per_sample * flags_obj.batch_size
optimizer = tf.keras.optimizers.SGD(
learning_rate=tf.keras.optimizers.schedules.ExponentialDecay(
initial_learning_rate,
decay_steps=steps_per_epoch *
flags_obj.num_epochs_per_decay,
decay_rate=flags_obj.lr_decay_factor,
staircase=True),
momentum=0.9)
if flags_obj.fp16_implementation == 'graph_rewrite':
# Note: when flags_obj.fp16_implementation == "graph_rewrite", dtype as
# determined by flags_core.get_tf_dtype(flags_obj) would be 'float32'
# which will ensure tf.compat.v2.keras.mixed_precision and
# tf.train.experimental.enable_mixed_precision_graph_rewrite do not double
# up.
optimizer = tf.train.experimental.enable_mixed_precision_graph_rewrite(
optimizer)
# TODO(hongkuny): Remove trivial model usage and move it to benchmark.
if flags_obj.use_trivial_model:
model = trivial_model.trivial_model(
imagenet_preprocessing.NUM_CLASSES)
elif flags_obj.model == 'resnet50_v1.5':
resnet_model.change_keras_layer(flags_obj.use_tf_keras_layers)
model = resnet_model.resnet50(
num_classes=imagenet_preprocessing.NUM_CLASSES)
elif flags_obj.model == 'mobilenet':
# TODO(kimjaehong): Remove layers attribute when minimum TF version
# support 2.0 layers by default.
model = tf.keras.applications.mobilenet.MobileNet(
weights=None,
classes=imagenet_preprocessing.NUM_CLASSES,
layers=tf.keras.layers)
if flags_obj.pretrained_filepath:
model.load_weights(flags_obj.pretrained_filepath)
if flags_obj.pruning_method == 'polynomial_decay':
if dtype != tf.float32:
raise NotImplementedError(
'Pruning is currently only supported on dtype=tf.float32.')
pruning_params = {
'pruning_schedule':
tfmot.sparsity.keras.PolynomialDecay(
initial_sparsity=flags_obj.pruning_initial_sparsity,
final_sparsity=flags_obj.pruning_final_sparsity,
begin_step=flags_obj.pruning_begin_step,
end_step=flags_obj.pruning_end_step,
frequency=flags_obj.pruning_frequency),
}
model = tfmot.sparsity.keras.prune_low_magnitude(
model, **pruning_params)
elif flags_obj.pruning_method:
raise NotImplementedError(
'Only polynomial_decay is currently supported.')
model.compile(loss='sparse_categorical_crossentropy',
optimizer=optimizer,
metrics=(['sparse_categorical_accuracy']
if flags_obj.report_accuracy_metrics else None),
run_eagerly=flags_obj.run_eagerly)
train_epochs = flags_obj.train_epochs
callbacks = common.get_callbacks(
steps_per_epoch=steps_per_epoch,
pruning_method=flags_obj.pruning_method,
enable_checkpoint_and_export=flags_obj.enable_checkpoint_and_export,
model_dir=flags_obj.model_dir)
# if mutliple epochs, ignore the train_steps flag.
if train_epochs <= 1 and flags_obj.train_steps:
steps_per_epoch = min(flags_obj.train_steps, steps_per_epoch)
train_epochs = 1
num_eval_steps = (imagenet_preprocessing.NUM_IMAGES['validation'] //
flags_obj.batch_size)
validation_data = eval_input_dataset
if flags_obj.skip_eval:
# Only build the training graph. This reduces memory usage introduced by
# control flow ops in layers that have different implementations for
# training and inference (e.g., batch norm).
if flags_obj.set_learning_phase_to_train:
# TODO(haoyuzhang): Understand slowdown of setting learning phase when
# not using distribution strategy.
tf.keras.backend.set_learning_phase(1)
num_eval_steps = None
validation_data = None
if not strategy and flags_obj.explicit_gpu_placement:
# TODO(b/135607227): Add device scope automatically in Keras training loop
# when not using distribition strategy.
no_dist_strat_device = tf.device('/device:GPU:0')
no_dist_strat_device.__enter__()
history = model.fit(train_input_dataset,
epochs=train_epochs,
steps_per_epoch=steps_per_epoch,
callbacks=callbacks,
validation_steps=num_eval_steps,
validation_data=validation_data,
validation_freq=flags_obj.epochs_between_evals,
verbose=2)
eval_output = None
if not flags_obj.skip_eval:
eval_output = model.evaluate(eval_input_dataset,
steps=num_eval_steps,
verbose=2)
if flags_obj.pruning_method:
model = tfmot.sparsity.keras.strip_pruning(model)
if flags_obj.enable_checkpoint_and_export:
if dtype == tf.bfloat16:
logging.warning(
'Keras model.save does not support bfloat16 dtype.')
else:
# Keras model.save assumes a float32 input designature.
export_path = os.path.join(flags_obj.model_dir, 'saved_model')
model.save(export_path, include_optimizer=False)
if not strategy and flags_obj.explicit_gpu_placement:
no_dist_strat_device.__exit__()
stats = common.build_stats(history, eval_output, callbacks)
return stats
def run(flags_obj):
"""Run ResNet ImageNet training and eval loop using native Keras APIs.
Args:
flags_obj: An object containing parsed flag values.
Raises:
ValueError: If fp16 is passed as it is not currently supported.
Returns:
Dictionary of training and eval stats.
"""
keras_utils.set_session_config(enable_eager=flags_obj.enable_eager,
enable_xla=flags_obj.enable_xla)
# Execute flag override logic for better model performance
if flags_obj.tf_gpu_thread_mode:
keras_utils.set_gpu_thread_mode_and_count(
per_gpu_thread_count=flags_obj.per_gpu_thread_count,
gpu_thread_mode=flags_obj.tf_gpu_thread_mode,
num_gpus=flags_obj.num_gpus,
datasets_num_private_threads=flags_obj.datasets_num_private_threads
)
common.set_cudnn_batchnorm_mode()
dtype = flags_core.get_tf_dtype(flags_obj)
if dtype == tf.float16:
loss_scale = flags_core.get_loss_scale(flags_obj, default_for_fp16=128)
policy = tf.compat.v2.keras.mixed_precision.experimental.Policy(
'mixed_float16', loss_scale=loss_scale)
tf.compat.v2.keras.mixed_precision.experimental.set_policy(policy)
if not keras_utils.is_v2_0():
raise ValueError('--dtype=fp16 is not supported in TensorFlow 1.')
elif dtype == tf.bfloat16:
policy = tf.compat.v2.keras.mixed_precision.experimental.Policy(
'mixed_bfloat16')
tf.compat.v2.keras.mixed_precision.experimental.set_policy(policy)
data_format = flags_obj.data_format
if data_format is None:
data_format = ('channels_first'
if tf.test.is_built_with_cuda() else 'channels_last')
tf.keras.backend.set_image_data_format(data_format)
preprocessing_seed = 12345
# pylint: disable=protected-access
if flags_obj.use_synthetic_data:
distribution_utils.set_up_synthetic_data()
input_fn = common.get_synth_input_fn(
height=imagenet_preprocessing.DEFAULT_IMAGE_SIZE,
width=imagenet_preprocessing.DEFAULT_IMAGE_SIZE,
num_channels=imagenet_preprocessing.NUM_CHANNELS,
num_classes=imagenet_preprocessing.NUM_CLASSES,
dtype=dtype,
drop_remainder=True)
else:
distribution_utils.undo_set_up_synthetic_data()
input_fn = imagenet_preprocessing.input_fn
# When `enable_xla` is True, we always drop the remainder of the batches
# in the dataset, as XLA-GPU doesn't support dynamic shapes.
drop_remainder = flags_obj.enable_xla
train_input_dataset = input_fn(
is_training=True,
data_dir=flags_obj.data_dir,
batch_size=flags_obj.batch_size,
num_epochs=flags_obj.train_epochs,
parse_record_fn=imagenet_preprocessing.parse_record,
datasets_num_private_threads=flags_obj.datasets_num_private_threads,
dtype=dtype,
drop_remainder=drop_remainder,
random_seed=preprocessing_seed, #addition
num_workers=current_cluster_size(), #addition
worker_ID=current_rank(), #addition
tf_data_experimental_slack=flags_obj.tf_data_experimental_slack,
training_dataset_cache=flags_obj.training_dataset_cache,
)
eval_input_dataset = None
if not flags_obj.skip_eval:
eval_input_dataset = input_fn(
is_training=False,
data_dir=flags_obj.data_dir,
batch_size=flags_obj.batch_size,
num_epochs=flags_obj.train_epochs,
parse_record_fn=imagenet_preprocessing.parse_record,
dtype=dtype,
drop_remainder=drop_remainder)
lr_schedule = 0.1
if flags_obj.use_tensor_lr:
lr_schedule = common.PiecewiseConstantDecayWithWarmup(
batch_size=flags_obj.batch_size,
epoch_size=imagenet_preprocessing.NUM_IMAGES['train'],
warmup_epochs=common.LR_SCHEDULE[0][1],
boundaries=list(p[1] for p in common.LR_SCHEDULE[1:]),
multipliers=list(p[0] for p in common.LR_SCHEDULE),
compute_lr_on_cpu=True)
# Build KungFu optimizer
opt = common.get_optimizer(lr_schedule)
# logging.info(opt.__dict__)
optimizer = SynchronousSGDOptimizer(opt, reshape=False, use_locking=True)
optimizer._hyper = opt._hyper
# logging.info(optimizer.__dict__)
if flags_obj.fp16_implementation == 'graph_rewrite':
# Note: when flags_obj.fp16_implementation == "graph_rewrite", dtype as
# determined by flags_core.get_tf_dtype(flags_obj) would be 'float32'
# which will ensure tf.compat.v2.keras.mixed_precision and
# tf.train.experimental.enable_mixed_precision_graph_rewrite do not double
# up.
optimizer = tf.train.experimental.enable_mixed_precision_graph_rewrite(
optimizer)
# TODO(hongkuny): Remove trivial model usage and move it to benchmark.
if flags_obj.use_trivial_model:
model = trivial_model.trivial_model(imagenet_preprocessing.NUM_CLASSES)
else:
model = resnet_model.resnet50(
num_classes=imagenet_preprocessing.NUM_CLASSES)
# TODO(b/138957587): Remove when force_v2_in_keras_compile is on longer
# a valid arg for this model. Also remove as a valid flag.
metrics = (['sparse_categorical_accuracy'])
metrics.append('sparse_top_k_categorical_accuracy')
if flags_obj.force_v2_in_keras_compile is not None:
model.compile(
loss='sparse_categorical_crossentropy',
optimizer=optimizer,
metrics=metrics,
run_eagerly=flags_obj.run_eagerly,
experimental_run_tf_function=flags_obj.force_v2_in_keras_compile)
else:
model.compile(loss='sparse_categorical_crossentropy',
optimizer=optimizer,
metrics=metrics,
run_eagerly=flags_obj.run_eagerly)
# adjust number of steps
cluster_size = current_cluster_size()
steps_per_epoch = (imagenet_preprocessing.NUM_IMAGES['train'] //
flags_obj.batch_size)
steps_per_epoch = steps_per_epoch // cluster_size
train_epochs = flags_obj.train_epochs
callbacks = common.get_callbacks(steps_per_epoch, current_rank(),
cluster_size,
common.learning_rate_schedule)
# Broadcast variables for KungFu
callbacks.append(BroadcastGlobalVariablesCallback())
# Checkpoint callback only on worker 0
if flags_obj.enable_checkpoint_and_export and current_rank() == 0:
ckpt_full_path = os.path.join(flags_obj.model_dir,
'model.ckpt-{epoch:04d}')
callbacks.append(
tf.keras.callbacks.ModelCheckpoint(ckpt_full_path,
save_weights_only=True))
if flags_obj.train_steps:
steps_per_epoch = min(flags_obj.train_steps, steps_per_epoch)
num_eval_steps = (imagenet_preprocessing.NUM_IMAGES['validation'] //
flags_obj.batch_size)
validation_data = eval_input_dataset
if flags_obj.skip_eval:
# Only build the training graph. This reduces memory usage introduced by
# control flow ops in layers that have different implementations for
# training and inference (e.g., batch norm).
if flags_obj.set_learning_phase_to_train:
# TODO(haoyuzhang): Understand slowdown of setting learning phase when
# not using distribution strategy.
tf.keras.backend.set_learning_phase(1)
num_eval_steps = None
validation_data = None
history = model.fit(train_input_dataset,
epochs=train_epochs,
steps_per_epoch=steps_per_epoch,
callbacks=callbacks,
validation_steps=num_eval_steps,
validation_data=validation_data,
validation_freq=flags_obj.epochs_between_evals,
verbose=2)
# Checkpoint only on 0th worker
if flags_obj.enable_checkpoint_and_export and current_rank() == 0:
if dtype == tf.bfloat16:
logging.warning(
"Keras model.save does not support bfloat16 dtype.")
else:
# Keras model.save assumes a float32 input designature.
export_path = os.path.join(flags_obj.model_dir, 'saved_model')
model.save(export_path, include_optimizer=False)
eval_output = None
if not flags_obj.skip_eval:
eval_output = model.evaluate(eval_input_dataset,
steps=num_eval_steps,
verbose=2)
stats = common.build_stats(history, eval_output, callbacks)
return stats
mode=experimental.Mode(args.mode),
context_dir=str(pathlib.Path.cwd()))
lr_schedule = ctx.get_hparam("learning_rate")
if ctx.get_data_config().get("use_tensor_lr", False):
lr_schedule = common.PiecewiseConstantDecayWithWarmup(
batch_size=ctx.get_per_slot_batch_size(),
epoch_size=imagenet_preprocessing.NUM_IMAGES["train"],
warmup_epochs=common.LR_SCHEDULE[0][1],
boundaries=[p[1] for p in common.LR_SCHEDULE[1:]],
multipliers=[p[0] for p in common.LR_SCHEDULE],
compute_lr_on_cpu=True,
)
optimizer = common.get_optimizer(lr_schedule)
model = resnet_model.resnet50(
num_classes=imagenet_preprocessing.NUM_CLASSES)
model = ctx.wrap_model(model)
model.compile(
loss="sparse_categorical_crossentropy",
optimizer=optimizer,
metrics=(["sparse_categorical_accuracy"]),
)
data_shape = (
ctx.get_per_slot_batch_size(),
imagenet_preprocessing.DEFAULT_IMAGE_SIZE,
imagenet_preprocessing.DEFAULT_IMAGE_SIZE,
imagenet_preprocessing.NUM_CHANNELS,
)
labels_shape = (ctx.get_per_slot_batch_size(), )