def test_horovod_learning_rate_schedule(self):
import horovod
major, minor, patch = horovod.__version__.split(".")
larger_major = int(major) > 0
larger_minor = int(major) == 0 and int(minor) > 19
larger_patch = int(major) == 0 and int(minor) == 19 and int(patch) >= 2
if larger_major or larger_minor or larger_patch:
ray_ctx = RayContext.get()
batch_size = 32
workers_per_node = 4
global_batch_size = batch_size * workers_per_node
config = {"batch_size": global_batch_size, "lr": 0.8}
trainer = Estimator(model_creator=simple_model,
compile_args_creator=compile_args,
verbose=True,
config=config,
backend="horovod",
workers_per_node=workers_per_node)
import horovod.tensorflow.keras as hvd
callbacks = [
hvd.callbacks.LearningRateWarmupCallback(warmup_epochs=5,
initial_lr=0.4,
verbose=True),
hvd.callbacks.LearningRateScheduleCallback(start_epoch=5,
end_epoch=10,
multiplier=1.,
initial_lr=0.4),
hvd.callbacks.LearningRateScheduleCallback(start_epoch=10,
end_epoch=15,
multiplier=1e-1,
initial_lr=0.4),
hvd.callbacks.LearningRateScheduleCallback(start_epoch=15,
end_epoch=20,
multiplier=1e-2,
initial_lr=0.4),
hvd.callbacks.LearningRateScheduleCallback(start_epoch=20,
multiplier=1e-3,
initial_lr=0.4),
LRChecker()
]
for i in range(30):
trainer.fit(create_train_datasets,
epochs=1,
callbacks=callbacks)
else:
# skip tests in horovod lower version
pass
def test_changing_config_during_evaluate(self):
train_data_shard = XShards.partition({
"x":
np.random.randn(100, 1),
"y":
np.random.randint(0, 1, size=(100, ))
})
config = {"lr": 0.8}
trainer = Estimator(model_creator=model_creator,
verbose=True,
config=config,
workers_per_node=2)
trainer.evaluate(train_data_shard, data_config={"batch_size": 8})
def test_require_batch_size(self):
train_data_shard = XShards.partition({
"x":
np.random.randn(100, 1),
"y":
np.random.randint(0, 1, size=(100, ))
})
config = {"lr": 0.8}
trainer = Estimator(model_creator=model_creator,
verbose=True,
config=config,
workers_per_node=2)
with pytest.raises(ray.exceptions.RayTaskError,
match=r".*batch_size must be set in config*."):
trainer.fit(train_data_shard, epochs=1, steps_per_epoch=25)
def test_num_part_data_diff_val_data(self):
sc = init_nncontext()
rdd = sc.range(200, numSlices=10)
val_rdd = sc.range(60, numSlices=8)
from pyspark.sql import SparkSession
spark = SparkSession(sc)
from pyspark.ml.linalg import DenseVector
df = rdd.map(lambda x:
(DenseVector(np.random.randn(1, ).astype(np.float)),
int(np.random.randint(0, 1, size=())))).toDF(
["feature", "label"])
val_df = val_rdd.map(lambda x: (DenseVector(np.random.randn(1,).astype(np.float)),
int(np.random.randint(0, 1, size=()))))\
.toDF(["feature", "label"])
config = {"lr": 0.8}
trainer = Estimator.from_keras(model_creator=model_creator,
verbose=True,
config=config,
workers_per_node=2)
assert df.rdd.getNumPartitions() > trainer.num_workers
assert df.rdd.getNumPartitions() != val_df.rdd.getNumPartitions()
trainer.fit(df,
epochs=1,
batch_size=4,
steps_per_epoch=25,
validation_data=val_df,
validation_steps=1,
feature_cols=["feature"],
label_cols=["label"])
def test_dataframe_shard_size(self):
from zoo.orca import OrcaContext
OrcaContext._shard_size = 3
sc = init_nncontext()
rdd = sc.range(0, 10)
from pyspark.sql import SparkSession
spark = SparkSession(sc)
from pyspark.ml.linalg import DenseVector
df = rdd.map(lambda x:
(DenseVector(np.random.randn(1, ).astype(np.float)),
int(np.random.randint(0, 1, size=())))).toDF(
["feature", "label"])
config = {"lr": 0.8}
trainer = Estimator.from_keras(model_creator=model_creator,
verbose=True,
config=config,
workers_per_node=2)
trainer.fit(df,
epochs=1,
batch_size=4,
steps_per_epoch=25,
feature_cols=["feature"],
label_cols=["label"])
trainer.evaluate(df,
batch_size=4,
num_steps=25,
feature_cols=["feature"],
label_cols=["label"])
trainer.predict(df, feature_cols=["feature"]).collect()
def test_partition_num_less_than_workers(self):
sc = init_nncontext()
rdd = sc.range(200, numSlices=1)
assert rdd.getNumPartitions() == 1
from pyspark.sql import SparkSession
spark = SparkSession(sc)
from pyspark.ml.linalg import DenseVector
df = rdd.map(lambda x:
(DenseVector(np.random.randn(1, ).astype(np.float)),
int(np.random.randint(0, 1, size=())))).toDF(
["feature", "label"])
config = {"lr": 0.8}
trainer = Estimator.from_keras(model_creator=model_creator,
verbose=True,
config=config,
workers_per_node=2)
assert df.rdd.getNumPartitions() < trainer.num_workers
trainer.fit(df,
epochs=1,
batch_size=4,
steps_per_epoch=25,
validation_data=df,
validation_steps=1,
feature_cols=["feature"],
label_cols=["label"])
trainer.evaluate(df,
batch_size=4,
num_steps=25,
feature_cols=["feature"],
label_cols=["label"])
trainer.predict(df, feature_cols=["feature"]).collect()
def test_pandas_dataframe(self):
def model_creator(config):
import tensorflow as tf
input1 = tf.keras.layers.Input(shape=(1, ))
input2 = tf.keras.layers.Input(shape=(1, ))
concatenation = tf.concat([input1, input2], axis=-1)
outputs = tf.keras.layers.Dense(
units=1, activation='softmax')(concatenation)
model = tf.keras.Model(inputs=[input1, input2], outputs=outputs)
model.compile(**compile_args(config))
return model
file_path = os.path.join(resource_path, "orca/learn/ncf2.csv")
train_data_shard = zoo.orca.data.pandas.read_csv(file_path)
config = {"lr": 0.8}
trainer = Estimator.from_keras(model_creator=model_creator,
verbose=True,
config=config,
workers_per_node=1)
trainer.fit(train_data_shard,
epochs=1,
batch_size=4,
steps_per_epoch=25,
feature_cols=["user", "item"],
label_cols=["label"])
trainer.evaluate(train_data_shard,
batch_size=4,
num_steps=25,
feature_cols=["user", "item"],
label_cols=["label"])
trainer.predict(train_data_shard, feature_cols=["user",
"item"]).collect()
def test_dataframe_with_empty_partition(self):
from zoo.orca import OrcaContext
sc = OrcaContext.get_spark_context()
rdd = sc.range(0, 10)
rdd_with_empty = rdd.repartition(4).\
mapPartitionsWithIndex(lambda idx, part: [] if idx == 0 else part)
from pyspark.sql import SparkSession
spark = SparkSession(sc)
from pyspark.ml.linalg import DenseVector
df = rdd_with_empty.map(lambda x: (DenseVector(np.random.randn(1,).astype(np.float)),
int(np.random.randint(0, 1, size=()))))\
.toDF(["feature", "label"])
config = {"lr": 0.8}
trainer = Estimator.from_keras(model_creator=model_creator,
verbose=True,
config=config,
workers_per_node=2)
trainer.fit(df,
epochs=1,
batch_size=4,
steps_per_epoch=25,
feature_cols=["feature"],
label_cols=["label"])
trainer.evaluate(df,
batch_size=4,
num_steps=25,
feature_cols=["feature"],
label_cols=["label"])
trainer.predict(df, feature_cols=["feature"]).collect()
def test_parquet_images_training(self):
from zoo.orca.learn.tf2 import Estimator
temp_dir = tempfile.mkdtemp()
try:
ParquetDataset.write("file://" + temp_dir, images_generator(),
images_schema)
path = "file://" + temp_dir
output_types = {
"id": tf.string,
"image": tf.string,
"label": tf.float32
}
output_shapes = {"id": (), "image": (), "label": ()}
def data_creator(config, batch_size):
dataset = read_parquet("tf_dataset",
input_path=path,
output_types=output_types,
output_shapes=output_shapes)
dataset = dataset.shuffle(10)
dataset = dataset.map(lambda data_dict:
(data_dict["image"], data_dict["label"]))
dataset = dataset.map(parse_data_train)
dataset = dataset.batch(batch_size)
return dataset
ray_ctx = RayContext.get()
trainer = Estimator.from_keras(model_creator=model_creator)
trainer.fit(data=data_creator, epochs=1, batch_size=2)
finally:
shutil.rmtree(temp_dir)
def test_predict_xshards(self):
train_data_shard = XShards.partition({
"x":
np.random.randn(100, 1),
"y":
np.random.randint(0, 1, size=(100, ))
})
expected = train_data_shard.collect()
expected = [shard["x"] for shard in expected]
for x in expected:
print(x.shape)
expected = np.concatenate(expected)
config = {}
trainer = Estimator.from_keras(model_creator=identity_model_creator,
verbose=True,
config=config,
workers_per_node=2)
result_shards = trainer.predict(train_data_shard,
batch_size=10).collect()
result = [shard["prediction"] for shard in result_shards]
expected_result = [shard["x"] for shard in result_shards]
result = np.concatenate(result)
assert np.allclose(expected, result)
def test_sparkxshards_with_inbalanced_data(self):
train_data_shard = XShards.partition({
"x":
np.random.randn(100, 1),
"y":
np.random.randint(0, 1, size=(100))
})
def random_pad(data):
import numpy as np
import random
times = random.randint(1, 10)
data["x"] = np.concatenate([data["x"]] * times)
data["y"] = np.concatenate([data["y"]] * times)
return data
train_data_shard = train_data_shard.transform_shard(random_pad)
config = {"lr": 0.8}
trainer = Estimator.from_keras(model_creator=model_creator,
verbose=True,
config=config,
workers_per_node=2)
trainer.fit(train_data_shard,
epochs=1,
batch_size=4,
steps_per_epoch=25)
trainer.evaluate(train_data_shard, batch_size=4, num_steps=25)
def test_auto_shard_tf(self):
# file 1 contains all 0s, file 2 contains all 1s
# If shard by files, then each model will
# see the same records in the same batch.
# If shard by records, then each batch
# will have different records.
# The loss func is constructed such that
# the former case will return 0, and the latter
# case will return non-zero.
ray_ctx = RayContext.get()
trainer = Estimator(
model_creator=auto_shard_model_creator,
verbose=True,
config={"batch_size": 4},
backend="tf", workers_per_node=2)
stats = trainer.fit(create_auto_shard_datasets, epochs=1, steps_per_epoch=2)
assert stats["train_loss"] == 0.0
def test_save_and_load(self):
def model_creator(config):
import tensorflow as tf
model = tf.keras.Sequential([
tf.keras.layers.Conv2D(64,
kernel_size=(3, 3),
strides=(1, 1),
activation='relu',
padding='valid'),
tf.keras.layers.BatchNormalization(),
tf.keras.layers.MaxPooling2D(pool_size=(2, 2),
strides=(2, 2),
padding='valid'),
tf.keras.layers.Conv2D(64,
kernel_size=(3, 3),
strides=(1, 1),
activation='relu',
padding='valid'),
tf.keras.layers.MaxPooling2D(pool_size=(2, 2),
strides=(2, 2),
padding='valid'),
tf.keras.layers.Flatten(),
tf.keras.layers.Dense(10, activation='softmax')
])
model.compile(optimizer=tf.keras.optimizers.RMSprop(),
loss='sparse_categorical_crossentropy',
metrics=['accuracy'])
return model
def train_data_creator(config, batch_size):
dataset = tf.data.Dataset.from_tensor_slices(
(np.random.randn(100, 28, 28,
3), np.random.randint(0, 10, (100, 1))))
dataset = dataset.repeat()
dataset = dataset.shuffle(1000)
dataset = dataset.batch(batch_size)
return dataset
batch_size = 320
try:
est = Estimator.from_keras(model_creator=model_creator,
workers_per_node=2)
history = est.fit(train_data_creator,
epochs=1,
batch_size=batch_size,
steps_per_epoch=5)
print("start saving")
est.save("/tmp/cifar10_keras.ckpt")
est.load("/tmp/cifar10_keras.ckpt")
print("save success")
finally:
os.remove("/tmp/cifar10_keras.ckpt")
def test_dataframe_predict(self):
sc = init_nncontext()
rdd = sc.parallelize(range(20))
df = rdd.map(lambda x: ([float(x)] * 5,
[int(np.random.randint(0, 2, size=()))])).toDF(
["feature", "label"])
estimator = Estimator.from_keras(model_creator=identity_model_creator,
verbose=True,
config={},
workers_per_node=2)
result = estimator.predict(df, batch_size=4, feature_cols=["feature"])
expr = "sum(cast(feature <> to_array(prediction) as int)) as error"
assert result.selectExpr(expr).first()["error"] == 0
def test_sparkxshards(self):
train_data_shard = XShards.partition({"x": np.random.randn(100, 1),
"y": np.random.randint(0, 1, size=(100))})
config = {
"lr": 0.8
}
trainer = Estimator.from_keras(
model_creator=model_creator,
verbose=True,
config=config,
workers_per_node=2)
trainer.fit(train_data_shard, epochs=1, batch_size=4, steps_per_epoch=25)
trainer.evaluate(train_data_shard, batch_size=4, num_steps=25)
def main(max_epoch):
batch_size = 320
config = {"batch_size": batch_size}
est = Estimator.from_keras(model_creator,
config=config,
workers_per_node=2)
stats = est.fit(train_data_creator,
epochs=max_epoch,
steps_per_epoch=60000 // batch_size,
validation_data_creator=val_data_creator,
validation_steps=10000 // batch_size)
print(stats)
est.save("/tmp/mnist_keras.ckpt")
est.restore("/tmp/mnist_keras.ckpt")
stats = est.evaluate(val_data_creator, steps=10000 // batch_size)
print(stats)
def test_string_input(self):
def model_creator(config):
import tensorflow as tf
vectorize_layer = tf.keras.layers.experimental.preprocessing.TextVectorization(
max_tokens=10, output_mode='int', output_sequence_length=4)
model = tf.keras.models.Sequential()
model.add(tf.keras.Input(shape=(1, ), dtype=tf.string))
model.add(vectorize_layer)
return model
from zoo.orca import OrcaContext
from pyspark.sql.types import StructType, StructField, StringType
spark = OrcaContext.get_spark_session()
schema = StructType([StructField("input", StringType(), True)])
input_data = [["foo qux bar"], ["qux baz"]]
input_df = spark.createDataFrame(input_data, schema)
estimator = Estimator.from_keras(model_creator=model_creator)
output_df = estimator.predict(input_df,
batch_size=1,
feature_cols=["input"])
output = output_df.collect()
print(output)
def impl_test_fit_and_evaluate(self, backend):
import tensorflow as tf
ray_ctx = RayContext.get()
batch_size = 32
global_batch_size = batch_size * ray_ctx.num_ray_nodes
config = {"batch_size": global_batch_size}
trainer = Estimator(model_creator=simple_model,
compile_args_creator=compile_args,
verbose=True,
config=config,
backend=backend)
# model baseline performance
start_stats = trainer.evaluate(create_test_dataset,
steps=NUM_TEST_SAMPLES //
global_batch_size)
print(start_stats)
def scheduler(epoch):
if epoch < 2:
return 0.001
else:
return 0.001 * tf.math.exp(0.1 * (2 - epoch))
scheduler = tf.keras.callbacks.LearningRateScheduler(scheduler,
verbose=1)
# train for 2 epochs
trainer.fit(create_train_datasets, epochs=2, callbacks=[scheduler])
trainer.fit(create_train_datasets, epochs=2, callbacks=[scheduler])
# model performance after training (should improve)
end_stats = trainer.evaluate(create_test_dataset,
steps=NUM_TEST_SAMPLES //
global_batch_size)
print(end_stats)
# sanity check that training worked
dloss = end_stats["validation_loss"] - start_stats["validation_loss"]
dmse = (end_stats["validation_mean_squared_error"] -
start_stats["validation_mean_squared_error"])
print(f"dLoss: {dloss}, dMSE: {dmse}")
assert dloss < 0 and dmse < 0, "training sanity check failed. loss increased!"
lr_schdule = tf.keras.callbacks.LearningRateScheduler(
lambda epoch: schedule(epoch, lr_multiplier), verbose=1)
config = {
"momentum": 0.9,
"wd": 0.00005,
"batch_size": args.batch_size_per_worker,
"val_batch_size": args.batch_size_per_worker,
"warmup_epoch": 5,
"num_worker": args.worker_num,
"data_dir": args.data_dir,
}
trainer = Estimator(model_creator=model_creator,
compile_args_creator=compile_args_creator,
verbose=True,
config=config,
backend="horovod")
results = trainer.fit(
data_creator=train_data_creator,
epochs=90,
validation_data_creator=val_data_creator,
steps_per_epoch=_NUM_IMAGES['train'] // global_batch_size,
callbacks=[lr_schdule],
validation_steps=_NUM_IMAGES['validation'] // global_batch_size,
)
print(results)
stop_orca_context()
def main():
anchors = yolo_anchors
anchor_masks = yolo_anchor_masks
parser = argparse.ArgumentParser()
parser.add_argument("--data_dir",
dest="data_dir",
help="Required. The path where data locates.")
parser.add_argument(
"--output_data",
dest="output_data",
default=tempfile.mkdtemp(),
help="Required. The path where voc parquet data locates.")
parser.add_argument("--data_year",
dest="data_year",
default="2009",
help="Required. The voc data date.")
parser.add_argument("--split_name_train",
dest="split_name_train",
default="train",
help="Required. Split name.")
parser.add_argument("--split_name_test",
dest="split_name_test",
default="val",
help="Required. Split name.")
parser.add_argument("--names",
dest="names",
help="Required. The path where class names locates.")
parser.add_argument("--weights",
dest="weights",
default="./checkpoints/yolov3.weights",
help="Required. The path where weights locates.")
parser.add_argument("--checkpoint",
dest="checkpoint",
default="./checkpoints/yolov3.tf",
help="Required. The path where checkpoint locates.")
parser.add_argument(
"--checkpoint_folder",
dest="checkpoint_folder",
default="./checkpoints",
help="Required. The path where saved checkpoint locates.")
parser.add_argument("--epochs",
dest="epochs",
type=int,
default=2,
help="Required. epochs.")
parser.add_argument("--batch_size",
dest="batch_size",
type=int,
default=16,
help="Required. epochs.")
parser.add_argument("--cluster_mode",
dest="cluster_mode",
default="local",
help="Required. Run on local/yarn/k8s mode.")
parser.add_argument("--class_num",
dest="class_num",
type=int,
default=20,
help="Required. class num.")
parser.add_argument(
"--worker_num",
type=int,
default=1,
help="The number of slave nodes to be used in the cluster."
"You can change it depending on your own cluster setting.")
parser.add_argument(
"--cores",
type=int,
default=4,
help="The number of cpu cores you want to use on each node. "
"You can change it depending on your own cluster setting.")
parser.add_argument(
"--memory",
type=str,
default="20g",
help="The memory you want to use on each node. "
"You can change it depending on your own cluster setting.")
parser.add_argument(
"--object_store_memory",
type=str,
default="10g",
help="The memory you want to use on each node. "
"You can change it depending on your own cluster setting.")
parser.add_argument('--k8s_master',
type=str,
default="",
help="The k8s master. "
"It should be k8s://https://<k8s-apiserver-host>: "
"<k8s-apiserver-port>.")
parser.add_argument("--container_image",
type=str,
default="",
help="The runtime k8s image. ")
parser.add_argument('--k8s_driver_host',
type=str,
default="",
help="The k8s driver localhost.")
parser.add_argument('--k8s_driver_port',
type=str,
default="",
help="The k8s driver port.")
options = parser.parse_args()
# convert yolov3 weights
yolo = YoloV3(classes=80)
load_darknet_weights(yolo, options.weights)
yolo.save_weights(options.checkpoint)
def model_creator(config):
model = YoloV3(DEFAULT_IMAGE_SIZE,
training=True,
classes=options.class_num)
anchors = yolo_anchors
anchor_masks = yolo_anchor_masks
model_pretrained = YoloV3(DEFAULT_IMAGE_SIZE,
training=True,
classes=80)
model_pretrained.load_weights(options.checkpoint)
model.get_layer('yolo_darknet').set_weights(
model_pretrained.get_layer('yolo_darknet').get_weights())
freeze_all(model.get_layer('yolo_darknet'))
optimizer = tf.keras.optimizers.Adam(lr=1e-3)
loss = [
YoloLoss(anchors[mask], classes=options.class_num)
for mask in anchor_masks
]
model.compile(optimizer=optimizer, loss=loss, run_eagerly=False)
return model
# prepare data
class_map = {
name: idx
for idx, name in enumerate(open(options.names).read().splitlines())
}
dataset_path = os.path.join(options.data_dir, "VOCdevkit")
voc_train_path = os.path.join(options.output_data, "train_dataset")
voc_val_path = os.path.join(options.output_data, "val_dataset")
write_parquet(format="voc",
voc_root_path=dataset_path,
output_path="file://" + voc_train_path,
splits_names=[(options.data_year, options.split_name_train)],
classes=class_map)
write_parquet(format="voc",
voc_root_path=dataset_path,
output_path="file://" + voc_val_path,
splits_names=[(options.data_year, options.split_name_test)],
classes=class_map)
output_types = {
"image": tf.string,
"label": tf.float32,
"image_id": tf.string
}
output_shapes = {"image": (), "label": (None, 5), "image_id": ()}
def train_data_creator(config, batch_size):
train_dataset = read_parquet(format="tf_dataset",
path=voc_train_path,
output_types=output_types,
output_shapes=output_shapes)
train_dataset = train_dataset.map(
lambda data_dict: (data_dict["image"], data_dict["label"]))
train_dataset = train_dataset.map(parse_data_train)
train_dataset = train_dataset.shuffle(buffer_size=512)
train_dataset = train_dataset.batch(batch_size)
train_dataset = train_dataset.map(lambda x, y: (
transform_images(x, DEFAULT_IMAGE_SIZE),
transform_targets(y, anchors, anchor_masks, DEFAULT_IMAGE_SIZE)))
train_dataset = train_dataset.prefetch(
buffer_size=tf.data.experimental.AUTOTUNE)
return train_dataset
def val_data_creator(config, batch_size):
val_dataset = read_parquet(format="tf_dataset",
path=voc_val_path,
output_types=output_types,
output_shapes=output_shapes)
val_dataset = val_dataset.map(lambda data_dict:
(data_dict["image"], data_dict["label"]))
val_dataset = val_dataset.map(parse_data_train)
val_dataset = val_dataset.batch(batch_size)
val_dataset = val_dataset.map(lambda x, y: (
transform_images(x, DEFAULT_IMAGE_SIZE),
transform_targets(y, anchors, anchor_masks, DEFAULT_IMAGE_SIZE)))
return val_dataset
callbacks = [
ReduceLROnPlateau(verbose=1),
EarlyStopping(patience=3, verbose=1),
ModelCheckpoint(options.checkpoint_folder + '/yolov3_train_{epoch}.tf',
verbose=1,
save_weights_only=True),
TensorBoard(log_dir='logs')
]
if options.cluster_mode == "local":
init_orca_context(cluster_mode="local",
cores=options.cores,
num_nodes=options.worker_num,
memory=options.memory,
init_ray_on_spark=True,
enable_numa_binding=False,
object_store_memory=options.object_store_memory)
elif options.cluster_mode == "k8s":
init_orca_context(cluster_mode="k8s",
master=options.k8s_master,
container_image=options.container_image,
init_ray_on_spark=True,
enable_numa_binding=False,
num_nodes=options.worker_num,
cores=options.cores,
memory=options.memory,
object_store_memory=options.object_store_memory,
conf={
"spark.driver.host": options.driver_host,
"spark.driver.port": options.driver_port
})
elif options.cluster_mode == "yarn":
init_orca_context(cluster_mode="yarn-client",
cores=options.cores,
num_nodes=options.worker_num,
memory=options.memory,
init_ray_on_spark=True,
enable_numa_binding=False,
object_store_memory=options.object_store_memory)
trainer = Estimator.from_keras(model_creator=model_creator)
trainer.fit(train_data_creator,
epochs=options.epochs,
batch_size=options.batch_size,
steps_per_epoch=3473 // options.batch_size,
callbacks=callbacks,
validation_data=val_data_creator,
validation_steps=3581 // options.batch_size)
stop_orca_context()
callbacks = get_lr_schedule_callbacks(initial_lr)
config = {
"wd": 0.00005,
"momentum": 0.9,
"batch_size": global_batch_size,
"warmup_epoch": 5,
"num_worker": args.worker_num,
"data_dir": args.data_dir,
"bf16": args.use_bf16,
"lr": initial_lr,
}
trainer = Estimator(model_creator=model_creator,
compile_args_creator=compile_args_creator,
verbose=True,
config=config,
backend="horovod")
if args.benchmark:
trainer.fit(
data_creator=train_data_creator
if not args.use_dummy_data else dummy_data_creator,
epochs=3,
steps_per_epoch=20,
callbacks=callbacks,
)
else:
epoch = 0
for i in range(5):
dummy = args.use_dummy_data
initial_lr = 0.1 * lr_multiplier
callbacks = get_lr_schedule_callbacks(initial_lr)
config = {
"wd": 0.00005,
"momentum": 0.9,
"warmup_epoch": 5,
"num_worker": args.worker_num,
"data_dir": args.data_dir,
"bf16": args.use_bf16,
"lr": initial_lr,
}
trainer = Estimator.from_keras(model_creator=model_creator,
compile_args_creator=compile_args_creator,
verbose=True,
config=config,
backend="horovod")
if args.benchmark:
trainer.fit(
data=train_data_creator
if not args.use_dummy_data else dummy_data_creator,
epochs=3,
batch_size=global_batch_size,
steps_per_epoch=20,
callbacks=callbacks,
)
else:
epoch = 0
for i in range(5):
