def init_gpu(args, logger):
hvd.init()
init_logger(
full=hvd.rank() == 0,
args=args,
logger=logger
)
if args.affinity != 'disabled':
gpu_id = hvd.local_rank()
affinity = set_affinity(
gpu_id=gpu_id,
nproc_per_node=hvd.size(),
mode=args.affinity
)
logger.warning(f'{gpu_id}: thread affinity: {affinity}')
gpus = tf.config.experimental.list_physical_devices('GPU')
tf.config.experimental.set_visible_devices(gpus[hvd.local_rank()], 'GPU')
if args.amp:
policy = tf.keras.mixed_precision.experimental.Policy("mixed_float16")
tf.keras.mixed_precision.experimental.set_policy(policy)
if args.xla:
tf.config.optimizer.set_jit(True)
def build(self) -> tf.data.Dataset:
"""Construct a dataset end-to-end and return it.
Args:
input_context: An optional context provided by `tf.distribute` for
cross-replica training.
Returns:
A TensorFlow dataset outputting batched images and labels.
"""
if self._use_dali:
print("Using dali for {train} dataloading".format(
train="training" if self.is_training else "validation"))
tfrec_filenames = sorted(
tf.io.gfile.glob(
os.path.join(self._data_dir, '%s-*' % self._split)))
tfrec_idx_filenames = sorted(
tf.io.gfile.glob(
os.path.join(self._index_file, '%s-*' % self._split)))
# # Create pipeline
dali_pipeline = Dali.DaliPipeline(
tfrec_filenames=tfrec_filenames,
tfrec_idx_filenames=tfrec_idx_filenames,
height=self._image_size,
width=self._image_size,
batch_size=self.local_batch_size,
num_threads=1,
device_id=hvd.local_rank(),
shard_id=hvd.rank(),
num_gpus=hvd.size(),
num_classes=self.num_classes,
deterministic=False,
dali_cpu=False,
training=self.is_training)
# Define shapes and types of the outputs
shapes = ((self.local_batch_size, self._image_size,
self._image_size, 3), (self.local_batch_size,
self._num_classes))
dtypes = (tf.float32, tf.float32)
# Create dataset
dataset = dali_tf.DALIDataset(pipeline=dali_pipeline,
batch_size=self.local_batch_size,
output_shapes=shapes,
output_dtypes=dtypes,
device_id=hvd.local_rank())
# if self.is_training and self._augmenter:
# print('Augmenting with {}'.format(self._augmenter))
# dataset.unbatch().map(self.augment_pipeline, num_parallel_calls=tf.data.experimental.AUTOTUNE).batch(self.local_batch_size)
return dataset
else:
print("Using tf native pipeline for {train} dataloading".format(
train="training" if self.is_training else "validation"))
dataset = self.load_records()
dataset = self.pipeline(dataset)
return dataset
def check_tf_1(aggregation_frequency: int,
average_aggregated_gradients: bool) -> None:
config = tf.compat.v1.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.visible_device_list = str(hvd.local_rank())
K.set_session(tf.Session(config=config))
session = tf.compat.v1.keras.backend.get_session(op_input_list=())
hvd_optimizer = hvd.DistributedOptimizer(
optimizer=CustomOptimizer("mine"),
backward_passes_per_step=aggregation_frequency,
average_aggregated_gradients=average_aggregated_gradients,
)
iterations = hvd_optimizer.iterations
session.run(iterations.initializer)
grads = [tf.constant([float(hvd.rank())])]
variables = [tf.Variable([0.0])]
session.run(variables[0].initializer)
allreduce_op = hvd_optimizer._allreduce(grads)
grads_and_vars = [(allreduce_op[0], variables[0])]
apply_grads_op = hvd_optimizer.apply_gradients(grads_and_vars)
for idx in range(10):
_ = session.run(apply_grads_op)
expected_value = compute_expected_variable_value(
idx, aggregation_frequency, average_aggregated_gradients)
assert idx + 1 == session.run(hvd_optimizer.iterations)
assert expected_value == session.run(variables[0].read_value())
def check_tf_2(aggregation_frequency: int,
average_aggregated_gradients: bool) -> None:
gpus = tf.config.experimental.list_physical_devices("GPU")
for gpu in gpus:
tf.config.experimental.set_memory_growth(gpu, True)
if gpus:
tf.config.experimental.set_visible_devices(gpus[hvd.local_rank()],
"GPU")
hvd_optimizer = hvd.DistributedOptimizer(
optimizer=CustomOptimizer("mine"),
backward_passes_per_step=aggregation_frequency,
average_aggregated_gradients=average_aggregated_gradients,
)
_ = hvd_optimizer.iterations
gradients = [tf.constant([float(hvd.rank())])]
variables = [tf.Variable([0.0])]
for idx in range(10):
if _PRE_TF_2_4_0:
# In TF < 2.4 `_aggregate_gradients()` is called outside of `apply_gradients()`.
updated_gradients = hvd_optimizer._aggregate_gradients(
zip(gradients, variables))
hvd_optimizer.apply_gradients(
zip(updated_gradients, variables),
experimental_aggregate_gradients=False)
else:
hvd_optimizer.apply_gradients(zip(gradients, variables))
updated_variable_value = variables[0][0].numpy()
expected_value = compute_expected_variable_value(
idx, aggregation_frequency, average_aggregated_gradients)
assert expected_value == updated_variable_value
assert idx + 1 == hvd_optimizer.iterations.numpy()
def init_workers(distributed=False):
"""Initialize distributed worker"""
rank, local_rank, n_ranks = 0, 0, 1
if distributed:
hvd.init()
rank, local_rank, n_ranks = hvd.rank(), hvd.local_rank(), hvd.size()
return rank, local_rank, n_ranks
def generate_stats_name(model, root):
# Generates the name of the output stats file.
# If Horovod distribution is enabled, the node and GPU ID are appended to the end
return ('%s_%s%s%s.csv' %
(model, root,
('_%s' % os.environ["HOSTNAME"] if hvd else ""),
(('_gpu%s' % hvd.local_rank()) if hvd else "")))
def run_model(args):
if args.lms:
tf.config.experimental.set_lms_enabled(True)
image_dim = args.image_size
opt = tf.keras.optimizers.RMSprop()
if hvd:
# Horovod: pin GPU to be used to process local rank (one GPU per process)
gpus = tf.config.list_physical_devices('GPU')
tf.config.experimental.set_memory_growth(gpus[hvd.local_rank()], True)
tf.config.set_visible_devices(gpus[hvd.local_rank()], 'GPU')
# Horovod: add Horovod DistributedOptimizer.
opt = hvd.DistributedOptimizer(opt)
steps_per_epoch = max(1, args.steps // hvd.size())
experimental_run_tf_function = False
else:
steps_per_epoch = args.steps
experimental_run_tf_function = True
if args.channels_last:
K.set_image_data_format('channels_last')
input_shape = (image_dim, image_dim, 3)
else:
K.set_image_data_format('channels_first')
input_shape = (3, image_dim, image_dim)
num_classes = 15
batch_size = args.batch_size
model_class = model_choices.get(args.model)
model = model_class(weights=None,
include_top=True,
input_shape=input_shape,
classes=num_classes)
model.compile(optimizer=opt,
loss='categorical_crossentropy',
experimental_run_tf_function=experimental_run_tf_function)
random_generator = random_image_generator(batch_size, num_classes,
input_shape)
model.fit(random_generator,
steps_per_epoch=steps_per_epoch,
verbose=1 if not hvd or hvd.rank() == 0 else 0,
epochs=args.epochs,
callbacks=get_callbacks(args))
def init_workers(distributed=False):
if distributed:
hvd.init()
return SimpleNamespace(rank=hvd.rank(), size=hvd.size(),
local_rank=hvd.local_rank(),
local_size=hvd.local_size())
else:
return SimpleNamespace(rank=0, size=1, local_rank=0, local_size=1)
def __init__(self, *args, **kwargs):
super(TfKerasTests, self).__init__(*args, **kwargs)
warnings.simplefilter('module')
hvd.init()
self.config = tf.compat.v1.ConfigProto()
self.config.gpu_options.allow_growth = True
self.config.gpu_options.visible_device_list = str(hvd.local_rank())
def main(_):
hvd.init()
print("After hvd init")
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.visible_device_list = str(hvd.local_rank())
# K.set_session(tf.Session(config=config))
print("After gpu_options visible_device_list")
tf.enable_eager_execution(config=config)
epochs = 20
steps_per_epoch = 2
batch_size = 32
num_classes = 10
full_model = 'image'
image_model = 'efficientnet'
image_training_type = 'finetuning'
text_model = 'cnn'
combined_embeddings = 'stack'
learning_rate = 0.005
width = 150
height = 150
input_shape = (height, width, 3)
input_size = (224, 224, 3)
train_tfrecord = tf.data.TFRecordDataset(filenames=['tfrecords/train.tfrecords'])
print(train_tfrecord)
val_tfrecord = tf.data.TFRecordDataset(filenames=['tfrecords/val.tfrecords'])
test_tfrecord = tf.data.TFRecordDataset(filenames=['tfrecords/test.tfrecords'])
def read_tfrecord(serialized_example):
feature_description = {
'image_raw': tf.io.FixedLenFeature([], tf.string),
'label': tf.io.FixedLenFeature([], tf.int64),
}
example = tf.io.parse_single_example(serialized_example, feature_description)
input_2 = tf.image.decode_png(example['image_raw'], channels=3, dtype=tf.dtypes.uint8)
input_2 = tf.image.resize(input_2, [600, 600])
return (input_2, example['label'])
train_parsed_dataset = train_tfrecord.map(read_tfrecord)
val_parsed_dataset = val_tfrecord.map(read_tfrecord)
test_parsed_dataset = test_tfrecord.map(read_tfrecord)
tf.keras.backend.clear_session()
baseModel = EfficientNetB7(weights='imagenet', include_top=True)
probs = baseModel.layers.pop()
top_droput = probs.input
headModel = layers.Dense(10, activation='softmax')(top_droput)
model = models.Model(inputs=baseModel.input, outputs=headModel)
SGD = optimizers.SGD(lr=0.01, decay=4e-05, momentum=0.9)
model.compile(loss='sparse_categorical_crossentropy', optimizer=adapt_optimizer(SGD), metrics=['accuracy'])
train_dataset = train_parsed_dataset.batch(2).repeat()
val_dataset = val_parsed_dataset.batch(2).repeat()
test_dataset = test_parsed_dataset.batch(2).repeat()
model.fit(train_dataset, epochs=adapt_epochs(epochs), steps_per_epoch=400, validation_data=val_dataset, validation_steps=100, verbose=(1 if (hvd.rank() == 0) else 0), callbacks=adapt_callbacks([], True))
if (hvd.rank() == 0):
model.save('saved_model.h5')
if (hvd.rank() == 0):
(test_loss, test_acc) = model.evaluate(test_dataset, verbose=0, steps=1241)
print('Test loss =', test_loss)
print('Test acc =', test_acc)
def initialize_horovod():
hvd.init()
gpus = tf.config.experimental.list_physical_devices("GPU")
for gpu in gpus:
tf.config.experimental.set_memory_growth(gpu, True)
if gpus:
tf.config.experimental.set_visible_devices(gpus[hvd.local_rank()],
"GPU")
return hvd.size()
def __init__(self, *args, **kwargs):
super(Tf2KerasTests, self).__init__(*args, **kwargs)
warnings.simplefilter('module')
hvd.init()
gpus = tf.config.experimental.list_physical_devices('GPU')
for gpu in gpus:
tf.config.experimental.set_memory_growth(gpu, True)
if gpus:
tf.config.experimental.set_visible_devices(gpus[hvd.local_rank()], 'GPU')
def config_gpu():
""""
Setup the GPUs to support for running with more memory as opposed to pre allocated memory.
Controls the percentage use of memory. Sets up process such that, typically one process acts on one gpu.
"""
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.visible_device_list = str(hvd.local_rank())
config.gpu_options.per_process_gpu_memory_fraction = FLAGS.gpu_memory_fraction
session = tf.Session(config=config)
K.set_session(session)
def gpu_config():
try:
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.visible_device_list = str(hvd.local_rank())
K.set_session(tf.Session(config=config))
except:
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
# config.gpu_options.allow_soft_placement = True
session = tf.Session(config=config)
K.set_session(session)
def connect_GPU_to_horovod():
import horovod.tensorflow.keras as hvd
import tensorflow as tf
tf.keras.backend.clear_session()
hvd.init()
# Horovod: pin GPU to be used to process local rank (one GPU per process)
gpus = tf.config.experimental.list_physical_devices('GPU')
for gpu in gpus:
tf.config.experimental.set_memory_growth(gpu, True)
if gpus:
tf.config.experimental.set_visible_devices(gpus[hvd.local_rank()],
'GPU')
def train_hvd(learning_rate=1.0):
# Tensorflow has given up on pickling. We need to explicitly import its modules inside workers
from tensorflow.keras import backend as K
from tensorflow.keras.models import Sequential
import tensorflow as tf
from tensorflow import keras
import horovod.tensorflow.keras as hvd
# Horovod: initialize Horovod.
hvd.init()
# Horovod: pin GPU to be used to process local rank (one GPU per process)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.visible_device_list = str(hvd.local_rank())
K.set_session(tf.Session(config=config))
(x_train, y_train), (x_test, y_test) = get_dataset(num_classes, hvd.rank(),
hvd.size())
model = get_model(num_classes)
# Horovod: adjust learning rate based on number of GPUs.
optimizer = keras.optimizers.Adadelta(lr=learning_rate * hvd.size())
# Horovod: add Horovod Distributed Optimizer.
optimizer = hvd.DistributedOptimizer(optimizer)
model.compile(optimizer=optimizer,
loss='categorical_crossentropy',
metrics=['accuracy'])
callbacks = [
# Horovod: broadcast initial variable states from rank 0 to all other processes.
# This is necessary to ensure consistent initialization of all workers when
# training is started with random weights or restored from a checkpoint.
hvd.callbacks.BroadcastGlobalVariablesCallback(0),
]
# Horovod: save checkpoints only on worker 0 to prevent other workers from corrupting them.
if hvd.rank() == 0:
callbacks.append(
keras.callbacks.ModelCheckpoint(checkpoint_dir +
'/checkpoint-{epoch}.ckpt',
save_weights_only=True))
model.fit(x_train,
y_train,
batch_size=batch_size,
callbacks=callbacks,
epochs=epochs,
verbose=2,
validation_data=(x_test, y_test))
def setup(args, report):
"""Set up environment variables given the type of partition."""
# Initialize Horovod
hvd.init()
# Set environment variable necessary to use h5py for file read/write
os.putenv("HDF5_USE_FILE_LOCKING", "FALSE")
os.system("export $HDF5_USE_FILE_LOCKING")
# Pin GPU to be used to process local rank (one GPU per process)
config = tf.ConfigProto()
config.log_device_placement = False
config.gpu_options.allow_growth = True
config.gpu_options.visible_device_list = str(hvd.local_rank())
K.set_session(tf.compat.v1.Session(config=config))
np.random.seed(args.random_seed)
print('Rank ' + str(hvd.rank()) + ' session configured')
def train_evaluate():
# Initialize Horovod
hvd.init()
# Horovod: pin GPU to be used to process local rank (one GPU per process)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.visible_device_list = str(hvd.local_rank())
tf.keras.backend.set_session(tf.Session(config=config))
train_dataset, eval_dataset = prepare_datasets()
model = toy_resnet_model()
# Wrap an optimizer in Horovod
optimizer = hvd.DistributedOptimizer(optimizers.Adadelta())
model.compile(optimizer=optimizer,
loss="categorical_crossentropy",
metrics=["accuracy"])
callbacks = [
# Horovod: broadcast initial variable states from rank 0 to all other processes.
# This is necessary to ensure consistent initialization of all workers when
# training is started with loaded weights.
hvd.callbacks.BroadcastGlobalVariablesCallback(0),
# Horovod: average metrics among workers at the end of every epoch.
#
# Note: This callback must be in the list before the ReduceLROnPlateau,
# TensorBoard, or other metrics-based callbacks.
hvd.callbacks.MetricAverageCallback()
]
# Horovod: save checkpoints only on worker 0 (master) to prevent other workers from corrupting them.
# Configure Tensorboard and Azure ML Tracking
if hvd.rank() == 0:
#callbacks.append(tf.keras.callbacks.ModelCheckpoint('./checkpoint-{epoch}.h5'))
callbacks.append(
tf.keras.callbacks.TensorBoard(log_dir=FLAGS['job-dir'].value,
update_freq='epoch'))
model.fit(train_dataset,
epochs=FLAGS.epochs,
steps_per_epoch=1000,
callbacks=callbacks,
validation_data=eval_dataset,
validation_steps=200)
def init_gpu(args, logger):
hvd.init()
init_logger(full=hvd.rank() == 0, args=args, logger=logger)
if args.affinity != "disabled":
gpu_id = hvd.local_rank()
affinity = set_affinity(gpu_id=gpu_id,
nproc_per_node=hvd.size(),
mode=args.affinity)
logger.warning(f"{gpu_id}: thread affinity: {affinity}")
if args.amp:
tf.keras.mixed_precision.set_global_policy("mixed_float16")
if args.xla:
tf.config.optimizer.set_jit(True)
def init(en_mem_growth=False, set_visible_dev=False):
""" This initializes the horovod package.
:param en_mem_growth:
:param set_visible_dev:
"""
if hvd is not None:
hvd.init()
gpus = tf.config.experimental.list_physical_devices('GPU')
for gpu in gpus:
if en_mem_growth:
tf.config.experimental.set_memory_growth(gpu, True)
if gpus and set_visible_dev:
tf.config.experimental.set_visible_devices(gpus[hvd.local_rank()],
'GPU')
else:
print("Horovod not supported on this system!")
def handle_distribution_strategy(distribution_strategy):
""" Create distribution strategy. """
strategy = None
if distribution_strategy:
strategy = distribution_strategy
if isinstance(distribution_strategy, dict):
strategy = distribution_strategy.get("distribution_strategy", None)
if isinstance(distribution_strategy, str):
strategy = distribution_strategy.lower()
if is_third_party_allreduce(strategy):
if strategy == "horovod":
import horovod.tensorflow.keras as hvd
else:
import byteps.tensorflow.keras as hvd
logging.info("import {} as hvd backend.".format(strategy))
hvd.init()
# Horovod: pin GPU to be used to process local rank (one GPU per process)
gpus = tf.config.experimental.list_physical_devices('GPU')
for gpu in gpus:
tf.config.experimental.set_memory_growth(gpu, True)
if gpus:
tf.config.experimental.set_visible_devices(
gpus[hvd.local_rank()], 'GPU')
compat.register_distributed_worker_setting(hvd.rank(), hvd.size(),
strategy)
if hvd.rank() != 0:
logging.set_verbosity(logging.ERROR)
else:
if isinstance(distribution_strategy, str):
strategy = distribution_utils.get_distribution_strategy(
distribution_strategy=distribution_strategy)
elif isinstance(distribution_strategy, dict):
strategy = distribution_utils.get_distribution_strategy(
**distribution_strategy)
if strategy is None:
logging.info("No distribution strategy was used.")
else:
try:
logging.info(
"Using distribution strategy: {} with num_replicas_in_sync={}".
format(strategy, strategy.num_replicas_in_sync))
except Exception:
pass
return strategy
def main() -> None:
"""
Start training Seq2Seq model.
:return: None
"""
# Horovod: initialize Horovod.
hvd.init()
# Pin GPU to be used to process local rank (one GPU per process)
gpu_list = tf.config.experimental.list_physical_devices("GPU")
for gpu in gpu_list:
tf.config.experimental.set_memory_growth(gpu, True)
if gpu_list:
print("Visible GPUs detected.")
tf.config.experimental.set_visible_devices(gpu_list[hvd.local_rank()], "GPU")
print("Loading input data.")
subject_list, body_list = load_data()
config: Seq2SeqConfig = fit_text(body_list, subject_list)
summarizer: Seq2SeqSummarizer = Seq2SeqSummarizer(config)
if not CONFIG.is_dev:
if tf.io.gfile.exists(LOCAL_MODEL_WEIGHTS):
summarizer.load_weights(weight_file_path=LOCAL_MODEL_WEIGHTS)
else:
Path(CONFIG.bucket_summarization_model).mkdir(parents=True, exist_ok=True)
if Path(LOCAL_MODEL_WEIGHTS).exists():
summarizer.load_weights(weight_file_path=LOCAL_MODEL_WEIGHTS)
body_train, body_test, subject_train, subject_test = train_test_split(body_list, subject_list, test_size=0.2)
print("Starting training.")
summarizer.fit(
body_train=body_train,
subject_train=subject_train,
body_test=body_test,
subject_test=subject_test,
epochs=int(math.ceil(100 / hvd.size())),
batch_size=128,
)
def __init__(self,
filenames,
idx_filenames,
height,
width,
batch_size,
num_threads,
dtype=tf.uint8,
dali_cpu=True,
deterministic=False,
training=False):
device_id = hvd.local_rank()
shard_id = hvd.rank()
num_gpus = hvd.size()
self.pipe = get_dali_pipeline(
tfrec_filenames=filenames,
tfrec_idx_filenames=idx_filenames,
height=height,
width=width,
batch_size=batch_size,
num_threads=num_threads,
device_id=device_id,
shard_id=shard_id,
num_gpus=num_gpus,
dali_cpu=dali_cpu,
training=training,
seed=7 * (1 + hvd.rank()) if deterministic else None)
self.daliop = dali_tf.DALIIterator()
self.batch_size = batch_size
self.height = height
self.width = width
self.device_id = device_id
self.dalidataset = dali_tf.DALIDataset(
pipeline=self.pipe,
output_shapes=((batch_size, height, width, 3), (batch_size)),
batch_size=batch_size,
output_dtypes=(tf.float32, tf.int64),
device_id=device_id)
def main(argv=None):
tf.reset_default_graph()
# init horovod
hvd.init()
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.visible_device_list = str(hvd.local_rank())
keras.backend.set_session(tf.Session(config=config))
image, label = _get_dataset()
model_input = keras.layers.Input(tensor=image)
model_output = keras.layers.Flatten(input_shape=(-1, 299, 299, 3))(model_input)
model_output = keras.layers.Dense(5, activation='relu')(model_output)
model = keras.models.Model(inputs=model_input, outputs=model_output)
# Horovod:
opt = keras.optimizers.Adadelta(1.0 * hvd.size())
opt = hvd.DistributedOptimizer(opt)
model.compile(optimizer=opt,loss='categorical_crossentropy',
metrics=['accuracy'],target_tensors=[label])
# callback
t_callback = keras.callbacks.TensorBoard(log_dir='./logs') # fit model
callbacks = [
hvd.callbacks.BroadcastGlobalVariablesCallback(0),
t_callback,
]
epochs = int(math.ceil(FLAGS.num_epochs / hvd.size()))
model.fit(epochs=epochs,steps_per_epoch=FLAGS.steps_one_epoch,callbacks=callbacks)
# save to h5
h5file = os.path.join(FLAGS.model_path,'model.h5')
if hvd.rank() == 0:
keras.models.save_model(model, h5file)
def __init__(self, config):
self.config = config
self.checkpoint_path = config.get_attribute('checkpoint_path')
self.epochs = config.get_attribute('epochs')
self.checkpoint_save_period = config.get_attribute(
'checkpoint_save_period')
self.checkpoint_format = 'checkpoint-{epoch}.h5'
self.learning_rate = config.get_attribute('learning_rate')
self.models_train = []
self.models_eval = []
self.train_steps_per_epoch = 1
self.eval_steps_per_epoch = 1
self.resume_from_epoch = 0
self.verbose = 1
self.cur_epoch = 0
hvd.init()
# Horovod: pin GPU to be used to process local rank (one GPU per process)
gpus = tf.config.experimental.list_physical_devices('GPU')
for gpu in gpus:
tf.config.experimental.set_memory_growth(gpu, True)
if gpus:
tf.config.experimental.set_visible_devices(gpus[hvd.local_rank()],
'GPU')
self.verbose = 1 if hvd.rank() == 0 else 0
origin_train_model = get_model(config, is_training=True)
origin_eval_model = get_model(config, is_training=False)
self.models_train.append(origin_train_model)
self.models_eval.append(origin_eval_model)
train_model = tf.keras.models.clone_model(origin_train_model)
eval_model = tf.keras.models.clone_model(origin_eval_model)
self.models_train.append(train_model)
self.models_eval.append(eval_model)
self.train_dataset, self.eval_dataset, self.train_dataset_distill, self.eval_dataset_distill = \
self.build_dataset()
self.build_train()
self.build_eval()
self.load_model()
self.save_model_path = config.get_attribute('checkpoint_eval_path')
self.callbacks = []
def check_tf_1(aggregation_frequency: int, average_aggregated_gradients: bool) -> None:
config = tf.compat.v1.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.visible_device_list = str(hvd.local_rank())
K.set_session(tf.Session(config=config))
session = tf.compat.v1.keras.backend.get_session(op_input_list=())
hvd_optimizer = hvd.DistributedOptimizer(
optimizer=CustomOptimizer("mine"),
aggregation_frequency=aggregation_frequency,
average_aggregated_gradients=average_aggregated_gradients,
)
constant_multiplier = 4.0
grads = [tf.constant([hvd.rank() * constant_multiplier])]
op = hvd_optimizer._allreduce(grads)
for idx in range(10):
value = session.run(op)[0][0]
expected_value = compute_expected_value(
idx, aggregation_frequency, constant_multiplier, average_aggregated_gradients, False
)
assert expected_value == value
def __init__(self, benchmark, output_dir):
self._benchmark = benchmark
self._output_dir = output_dir
Path(self._output_dir).mkdir(parents=True, exist_ok=True)
host_spec = HostSpec()
self._node_name = host_spec.node_name
# Log system information if on local rank 0
if hvd.local_rank() == 0:
# Log host information
file_name = '{}_host.json'.format(self._node_name)
db = TrackingClient(Path(self._output_dir) / file_name)
host_info = {
'name': host_spec.name,
'node_name': host_spec.node_name,
'ip': host_spec.node_name,
'num_cores': host_spec.num_cores,
'release': host_spec.release,
'system': host_spec.system,
'cpu_info': host_spec.cpu_info,
}
db.log_tag('host_info', host_info)
# Log device information
device_specs = DeviceSpecs()
file_name = '{}_devices.json'.format(self._node_name)
db = TrackingClient(Path(self._output_dir) / file_name)
device_info = {}
device_info['gpu_count'] = device_specs.device_count
device_info.update({'gpu_{}'.format(i): device_specs.get_device_info(i) for i in range(device_specs.device_count)})
db.log_tag('device_info', device_info)
def check_tf_2(aggregation_frequency: int, average_aggregated_gradients: bool) -> None:
gpus = tf.config.experimental.list_physical_devices("GPU")
for gpu in gpus:
tf.config.experimental.set_memory_growth(gpu, True)
if gpus:
tf.config.experimental.set_visible_devices(gpus[hvd.local_rank()], "GPU")
hvd_optimizer = hvd.DistributedOptimizer(
optimizer=CustomOptimizer("mine"),
aggregation_frequency=aggregation_frequency,
average_aggregated_gradients=average_aggregated_gradients,
)
constant_multiplier = 4.0
grads_and_vars = [(tf.constant([hvd.rank() * constant_multiplier]), None)]
for idx in range(10):
grads = hvd_optimizer._aggregate_gradients(grads_and_vars)
value = grads[0][0].numpy()
expected_value = compute_expected_value(
idx, aggregation_frequency, constant_multiplier, average_aggregated_gradients, True
)
assert expected_value == value
RANDOM_STATE = mlctx.get_param("random_state", 1)
TEST_SIZE = mlctx.get_param("test_size", 0.2)
# kubeflow outputs/inputs
categories_map = str(mlctx.get_input("categories_map").get())
df = pd.read_csv(str(mlctx.get_input("file_categories")))
# Horovod: initialize Horovod.
hvd.init()
# if gpus found, pin GPU to be used to process local rank (one GPU per process)
gpus = tf.config.experimental.list_physical_devices("GPU")
if gpus:
for gpu in gpus:
tf.config.experimental.set_memory_growth(gpu, True)
tf.config.experimental.set_visible_devices(gpus[hvd.local_rank()], "GPU")
else:
os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
if hvd.rank() == 0:
mlctx.logger.info(
f"Validating paths:\nData_path:\t{DATA_PATH}\nModel_dir:\t{MODEL_DIR}\n"
)
mlctx.logger.info(f"Categories map:{categories_map}")
mlctx.logger.info(f"Got {df.shape[0]} files in {DATA_PATH}")
mlctx.logger.info(f"Training data has {df.size} samples")
mlctx.logger.info(df.category.value_counts())
# artifact folders (deprecate these)
os.makedirs(DATA_PATH, exist_ok=True)
os.makedirs(CHECKPOINTS_DIR, exist_ok=True)
def train(model_func, params):
image_width = params['image_width']
image_height = params['image_height']
image_format = params['image_format']
distort_color = params['distort_color']
momentum = params['momentum']
loss_scale = params['loss_scale']
data_dir = params['data_dir']
data_idx_dir = params['data_idx_dir']
batch_size = params['batch_size']
num_iter = params['num_iter']
iter_unit = params['iter_unit']
log_dir = params['log_dir']
export_dir = params['export_dir']
tensorboard_dir = params['tensorboard_dir']
display_every = params['display_every']
precision = params['precision']
dali_mode = params['dali_mode']
use_xla = params['use_xla']
if data_dir is not None:
file_format = os.path.join(data_dir, '%s-*')
train_files = sorted(tf.io.gfile.glob(file_format % 'train'))
valid_files = sorted(tf.io.gfile.glob(file_format % 'validation'))
num_train_samples = common.get_num_records(train_files)
num_valid_samples = common.get_num_records(valid_files)
else:
num_train_samples = 1281982
num_valid_samples = 5000
train_idx_files = None
valid_idx_files = None
if data_idx_dir is not None:
file_format = os.path.join(data_idx_dir, '%s-*')
train_idx_files = sorted(tf.io.gfile.glob(file_format % 'train'))
valid_idx_files = sorted(tf.io.gfile.glob(file_format % 'validation'))
if iter_unit.lower() == 'epoch':
num_epochs = num_iter
nstep_per_epoch = num_train_samples // (batch_size * hvd.size())
nstep_per_valid = num_valid_samples // (batch_size * hvd.size())
else:
assert iter_unit.lower() == 'batch'
num_epochs = 1
nstep_per_epoch = min(num_iter,
num_train_samples // (batch_size * hvd.size()))
nstep_per_valid = min(10,
num_valid_samples // (batch_size * hvd.size()))
initial_epoch = 0
if log_dir:
# We save check points only when using the real data.
assert data_dir, "--data_dir cannot be empty when using --log_dir"
assert os.path.exists(log_dir)
ckpt_format = log_dir + "/model-{epoch:02d}-{val_top1:.2f}.hdf5"
# Looks for the most recent checkpoint and sets the initial epoch from it.
for filename in os.listdir(log_dir):
if filename.startswith('model-'):
initial_epoch = max(int(re.findall(r'\d+', filename)[0]),
initial_epoch)
if tensorboard_dir:
assert os.path.exists(tensorboard_dir)
if export_dir:
assert os.path.exists(export_dir)
save_format = export_dir + "/saved_model_rn50.h5"
if use_xla:
tf.config.optimizer.set_jit(True)
# Horovod: pin GPU to be used to process local rank (one GPU per process)
gpus = tf.config.experimental.list_physical_devices('GPU')
for gpu in gpus:
tf.config.experimental.set_memory_growth(gpu, True)
if gpus:
tf.config.experimental.set_visible_devices(gpus[hvd.local_rank()],
'GPU')
if precision == 'fp16':
policy = keras.mixed_precision.experimental.Policy(
'mixed_float16', loss_scale)
keras.mixed_precision.experimental.set_policy(policy)
lr_schedule = common.create_piecewise_constant_decay_with_warmup(
batch_size=batch_size * hvd.size(),
epoch_size=num_train_samples,
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)
opt = keras.optimizers.SGD(learning_rate=lr_schedule, momentum=momentum)
# Horovod: add Horovod DistributedOptimizer. We use a modified version to
# support the custom learning rate schedule.
opt = hvd_patch.DistributedOptimizer(opt)
backend.set_image_data_format(image_format)
dtype = 'float16' if precision == 'fp16' else 'float32'
backend.set_floatx(dtype)
model = model_func(num_classes=image_processing.NUM_CLASSES)
loss_func = 'sparse_categorical_crossentropy',
top5 = tf.keras.metrics.SparseTopKCategoricalAccuracy(k=5, name='top5')
top1 = tf.keras.metrics.SparseTopKCategoricalAccuracy(k=1, name='top1')
# Horovod: Specify `experimental_run_tf_function=False` to ensure TensorFlow
# uses hvd.DistributedOptimizer() to compute gradients. However, this option
# will disable the overlapping of the data loading and compute and hurt the
# performace if the model is not under the scope of distribution strategy
# scope.
model.compile(optimizer=opt,
loss=loss_func,
metrics=[top1, top5],
experimental_run_tf_function=False)
training_hooks = []
training_hooks.append(hvd.callbacks.BroadcastGlobalVariablesCallback(0))
training_hooks.append(_ProfileKerasFitCallback(batch_size, display_every))
if log_dir and hvd.rank() == 0:
ckpt_callback = keras.callbacks.ModelCheckpoint(
ckpt_format,
monitor='val_top1',
verbose=1,
save_best_only=False,
save_weights_only=False,
save_frequency=1)
training_hooks.append(ckpt_callback)
if tensorboard_dir and hvd.rank() == 0:
tensorboard_callback = tf.keras.callbacks.TensorBoard(
log_dir=tensorboard_dir)
training_hooks.append(tensorboard_callback)
if data_dir is not None:
num_preproc_threads = params['dali_threads'] if dali_mode else 10
train_input = image_processing.image_set(
train_files,
batch_size,
image_height,
image_width,
training=True,
distort_color=distort_color,
deterministic=False,
num_threads=num_preproc_threads,
use_dali=dali_mode,
idx_filenames=train_idx_files)
valid_input = image_processing.image_set(
valid_files,
batch_size,
image_height,
image_width,
training=False,
distort_color=False,
deterministic=False,
num_threads=num_preproc_threads,
use_dali=dali_mode,
idx_filenames=valid_idx_files)
if dali_mode:
train_input = train_input.get_device_dataset()
valid_input = valid_input.get_device_dataset()
valid_params = {
'validation_data': valid_input,
'validation_steps': nstep_per_valid,
'validation_freq': 1
}
else:
train_input = image_processing.fake_image_set(batch_size, image_height,
image_width)
valid_params = {}
try:
verbose = 2 if hvd.rank() == 0 else 0
model.fit(train_input,
epochs=num_epochs,
callbacks=training_hooks,
steps_per_epoch=nstep_per_epoch,
verbose=verbose,
initial_epoch=initial_epoch,
**valid_params)
except KeyboardInterrupt:
print("Keyboard interrupt")
if export_dir and hvd.rank() == 0:
model.save(save_format)
print(f"The model is saved to {save_format}")
# Data, model, and output directories. These are required.
parser.add_argument('--output-dir', type=str, default=os.environ['SM_OUTPUT_DIR'])
parser.add_argument('--model_dir', type=str)
parser.add_argument('--train', type=str, default=os.environ['SM_CHANNEL_TRAIN'])
parser.add_argument('--test', type=str, default=os.environ['SM_CHANNEL_TEST'])
args, _ = parser.parse_known_args()
# Horovod: initialize Horovod.
hvd.init()
# Horovod: pin GPU to be used to process local rank (one GPU per process)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.visible_device_list = str(hvd.local_rank())
K.set_session(tf.Session(config=config))
batch_size = 128
num_classes = 10
# Horovod: adjust number of epochs based on number of GPUs.
epochs = int(math.ceil(12.0 / hvd.size()))
# Input image dimensions
img_rows, img_cols = 28, 28
# The data, shuffled and split between train and test sets
x_train = np.load(os.path.join(args.train, 'train.npz'))['data']
y_train = np.load(os.path.join(args.train, 'train.npz'))['labels']
