def getModel(net_settings, num_classes=1):
'''
Should be modified with model type as input and returns the desired model
'''
if net_settings['model_type'] == 'resnet':
base_model = resnet50.ResNet50(include_top=True, weights='imagenet')
finetuning = Dense(1, activation='sigmoid',
name='predictions')(base_model.layers[-2].output)
model = Model(input=base_model.input, output=finetuning)
## Adjust learning rate based on number of GPUs
hv_lr = net_settings['lr'] * hvd.size()
opt = optimizers.SGD(lr=hv_lr, momentum=0.9, decay=1e-6, nesterov=True)
## Adding Horovod DistributedOptimizer
opt = hvd.DistributedOptimizer(opt)
model.compile(loss=net_settings['loss'],
optimizer=opt,
metrics=['accuracy'])
callbacks = [
hvd.callbacks.BroadcastGlobalVariablesCallback(0),
]
if hvd.rank() == 0:
callbacks.append(
keras.callbacks.ModelCheckpoint('./checkpoint-{epoch}.h5'))
return model
elif net_settings['model_type'] == 'resnet101':
model = resnet101_model(224, 224, 3, 1)
## Adjust learning rate based on number of GPUs
hv_lr = net_settings['lr'] * hvd.size()
opt = optimizers.SGD(lr=hv_lr, momentum=0.9, decay=1e-6, nesterov=True)
## Adding Horovod DistributedOptimizer
opt = hvd.DistributedOptimizer(opt)
model.compile(loss=net_settings['loss'],
optimizer=opt,
metrics=['accuracy'])
callbacks = [
hvd.callbacks.BroadcastGlobalVariablesCallback(0),
]
if hvd.rank() == 0:
callbacks.append(
keras.callbacks.ModelCheckpoint('./checkpoint-{epoch}.h5'))
return model
else:
print '[models] Ugggh. Not ready for this yet.'
exit(0)
return None
def _get_hooks(is_distributed=_DISTRIBUTED, verbose=1):
logger = _get_logger()
if is_distributed:
logger.info("Rank: {} Cluster Size {}".format(hvd.local_rank(),
hvd.size()))
return [
# 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: 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: using `lr = 1.0 * hvd.size()` from the very beginning leads to worse final
# accuracy. Scale the learning rate `lr = 1.0` ---> `lr = 1.0 * hvd.size()` during
# the first five epochs. See https://arxiv.org/abs/1706.02677 for details.
hvd.callbacks.LearningRateWarmupCallback(
warmup_epochs=_WARMUP_EPOCHS, verbose=verbose),
# Horovod: after the warmup reduce learning rate by 10 on the 30th, 60th and 80th epochs.
hvd.callbacks.LearningRateScheduleCallback(
start_epoch=_WARMUP_EPOCHS, end_epoch=30, multiplier=1.0),
hvd.callbacks.LearningRateScheduleCallback(start_epoch=30,
end_epoch=60,
multiplier=1e-1),
hvd.callbacks.LearningRateScheduleCallback(start_epoch=60,
end_epoch=80,
multiplier=1e-2),
hvd.callbacks.LearningRateScheduleCallback(start_epoch=80,
multiplier=1e-3),
]
else:
return []
def setup_callbacks(params, callbacks, encoder, decoder, prop_pred):
import horovod.keras as hvd
# model checkpointing
if params.checkpoint_period and hvd.rank() == 0:
model_checkpoint_callback = model_checkpoint(
encoder,
decoder,
prop_pred,
params.checkpoint_path,
nepochs=params.checkpoint_period,
overwrite=params.overwrite_checkpoint)
callbacks.append(model_checkpoint_callback)
# LR scheduler
if params.lr_schedule_patience:
lr_callback = ReduceLROnPlateau(monitor=params.lr_schedule_prop,
factor=0.5,
patience=params.lr_schedule_patience,
min_lr=params.lr_schedule_min *
hvd.size(),
cooldown=params.lr_schedule_cooldown,
verbose=(hvd.rank() == 0))
callbacks.append(lr_callback)
if hvd.rank() == 0:
callbacks.append(print_loss())
if params.enable_tensorboard:
callbacks.append(TensorBoard(params.checkpoint_path))
def load_data(self, data_fn, test_size=0.3, random=True):
if not self.distributed_training:
self.logger.info(
'Loading the full dataset since distributed training is disabled ...'
)
# X, Y = self.data_io.load_all(data_fn, dset_name_pattern=dset_name_pattern, camera_pos=camera_pos)
X, Y = self.data_io.load_all(data_fn)
else:
self.logger.info(
'Loading part of the dataset since distributed training is enabled ...'
)
X, Y = self.data_io.load_partial(data_fn, hvd.size(), hvd.rank())
self.logger.debug('Shape of X: %s' % str(X.shape))
self.logger.debug('Shape of Y: %s' % str(Y.shape))
# update the input_shape setting according to the loaded data
self.input_shape = X.shape[1:]
if test_size > 0:
x_train, x_test, y_train, y_test = train_test_split(
X, Y, test_size=test_size, random_state=42)
self.x_train = x_train
self.x_test = x_test
self.y_train = y_train
self.y_test = y_test
else:
self.x_train = X
self.y_train = Y
self.num_classes = np.unique(Y).shape[0]
print("shapes:", self.x_train.shape, self.x_test.shape,
self.y_train.shape, self.y_test.shape)
self.logger.debug('Number of classes: %d' % self.num_classes)
def save_model(self):
if self.distributed_training is True:
if hvd.rank() == 0:
if self.use_noise is True:
self.model.save('model_hvd_bw_%d_B0_with_noise_n_p_%d.h5' %
(self.input_shape[0], hvd.size()))
else:
self.model.save('model_hvd_bw_%d_B0_no_noise_%d_nodes.h5' %
(self.input_shape[0], hvd.size()))
else:
if self.use_noise is True:
self.model.save('model_bw_%d_B0_with_noise.h5' %
(self.input_shape[0]))
else:
self.model.save('model_bw_%d_B0_no_noise.h5' %
(self.input_shape[0]))
def test_elastic_state(self):
with self.test_session(config=self.config) as sess:
K.set_session(sess)
v = 1.0 if hvd.rank() == 0 else 2.0
model1 = keras.models.Sequential([
keras.layers.Dense(2, activation='softmax')
])
model1.build((2, 2))
model1.set_weights(
[np.array([[v, v], [v, v]], dtype=np.float32),
np.array([v, v], dtype=np.float32)])
model2 = keras.models.Sequential([
keras.layers.Dense(2, activation='softmax')
])
model2.build((2, 2))
model2.set_weights(
[np.array([[1.0, 2.0], [3.0, 4.0]], dtype=np.float32),
np.array([0.0, 0.0], dtype=np.float32)])
optimizer = keras.optimizers.Adam(0.001 * hvd.size())
state = hvd.elastic.KerasState(model1, optimizer, batch=20 + hvd.rank(), epoch=10 + hvd.rank())
state.sync()
model1_weights = model1.get_weights()
model2_weights = model2.get_weights()
# After sync, all values should match the root rank
for w in state.model.get_weights():
self.assertAllClose(w, np.ones_like(w))
assert state.batch == 20
assert state.epoch == 10
# Partially modify then restore
model1.set_weights(model2_weights)
state.batch = 21
state.epoch = 11
state.restore()
for w1, w2 in zip(model1.get_weights(), model1_weights):
self.assertAllClose(w1, w2)
assert state.batch == 20
assert state.epoch == 10
# Partially modify then commit
model1.set_weights(model2_weights)
state.batch = 21
state.epoch = 11
state.commit()
state.restore()
for w1, w2 in zip(model1.get_weights(), model2_weights):
self.assertAllClose(w1, w2)
assert state.batch == 21
assert state.epoch == 11
def get_batch_sharing_solution(self, batch_patch_info):
n_workers = hvd.size()
n_imgs_per_worker = batch_patch_info.shape[1] // hvd.size()
worker_batch_delta = np.zeros((hvd.size(), 2), np.int32)
for wi in range(n_workers):
worker_batch_delta[wi, 0] = np.sum(
batch_patch_info[0, :] == wi) - n_imgs_per_worker
worker_batch_delta[wi, 1] = wi
sw = worker_batch_delta[worker_batch_delta[:, 0].argsort(
), :] # sorted by decreasing nb of missing patchs
# print("***SHARING SOLUTION***")
# print("INITIAL OFFERINGS")
# print(sw)
transfers = np.zeros((n_workers, n_workers), np.int32)
i = 0
j = n_workers - 1
while i < j and sw[
i, 0] < 0: # where there are missing patches for a worker
if sw[i, 0] < 0 and sw[
j,
0] > 0: # if patches missing for a worker and too much for another
init_i = sw[i, 0]
init_j = sw[j, 0]
if -sw[i, 0] < sw[
j,
0]: #worker j having images can fullfill request of i and still have some images left
transfers[sw[i, 1], sw[j, 1]] = sw[i, 0]
transfers[sw[j, 1], sw[i, 1]] = -sw[i, 0]
sw[j, 0] += sw[i, 0]
sw[i, 0] = 0
i += 1
else: #worker i having images can can get all images of i and (may) still need some more
transfers[sw[i, 1], sw[j, 1]] = -sw[j, 0]
transfers[sw[j, 1], sw[i, 1]] = sw[j, 0]
sw[i, 0] += sw[j, 0]
sw[j, 0] = 0
if -init_i == init_j: #if both are fullfilled continue
i += 1
j -= 1
if not np.sum(sw[:, 0]) == 0:
raise Exception(
"Error in sharing solution, check source code !!!!")
return transfers
def save_model(self):
if self.distributed_training is True:
if hvd.rank() == 0:
if self.noise_stddev > 0 is True:
self.model.save('model_%d_%s_noise_np_%d.h5' %
(self.input_shape[0], self.base_model_name,
hvd.size()))
else:
self.model.save('model_%d_%s_np_%d.h5' %
(self.input_shape[0], self.base_model_name,
hvd.size()))
else:
if self.noise_stddev > 0 is True:
self.model.save('model_%d_%s_noise.h5' %
(self.input_shape[0], self.base_model_name))
else:
self.model.save('model_%d_%s.h5' %
(self.input_shape[0], self.base_model_name))
def build(self):
from keras.optimizers import deserialize
opt_config = {'class_name': self.name, 'config': self.config}
opt = deserialize(opt_config)
if self.horovod_wrapper:
import horovod.keras as hvd
if hasattr(opt, 'lr'):
opt.lr *= hvd.size()
opt = hvd.DistributedOptimizer(opt)
return opt
def _get_optimizer(params, is_distributed=_DISTRIBUTED):
if is_distributed:
# Horovod: adjust learning rate based on number of GPUs.
opt = keras.optimizers.SGD(lr=params["learning_rate"] * hvd.size(),
momentum=params["momentum"])
# Horovod: add Horovod Distributed Optimizer.
return hvd.DistributedOptimizer(opt)
else:
return keras.optimizers.SGD(lr=params["learning_rate"],
momentum=params["momentum"])
def train_evaluate():
# Generate training and validation data generators
def get_image_list(data_dir):
dataset = []
for folder in os.listdir(data_dir):
for image in os.listdir(os.path.join(data_dir, folder)):
dataset.append((os.path.join(data_dir, folder, image), folder))
return dataset
training_data = ImageSequence(get_image_list(os.path.join(FLAGS.data_dir, 'train')), FLAGS.batch_size, True)
validation_data = ImageSequence(get_image_list(os.path.join(FLAGS.data_dir, 'test')), FLAGS.batch_size, False)
# Horovod: Initialize Horovod
hvd.init()
# Horvod: 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))
# Create a model
model = network_model(FLAGS.hidden_units)
loss = 'categorical_crossentropy'
# Horovod: Adjust learning rate based on number of GPUs
optimizer = Adadelta(lr=1.0 * hvd.size())
# Horovod: add Horovod Distributed Optimizer
optimizer = hvd.DistributedOptimizer(optimizer)
metrics = ['acc']
model.compile(optimizer, loss, metrics)
# Set up callbacks
callbacks = [
# Broadcast initial variable states from rank 0 to all other processes
hvd.callbacks.BroadcastGlobalVariablesCallback(0),
]
# Horovod: save logs only on worker 0
if hvd.rank() == 0:
callbacks.append(tf.keras.callbacks.TensorBoard(log_dir=FLAGS.log_dir))
# Start training
model.fit_generator(generator = training_data,
validation_data = validation_data,
epochs = FLAGS.epochs,
use_multiprocessing = True,
workers = 4,
callbacks = callbacks,
verbose = 1)
# Save the model
model.save(FLAGS.save_model_path)
def __init__(self, filename, batch_size):
self.f_array = h5py.File(filename, "r")
x = self.f_array["images"]
y = self.f_array["masks"]
self.batch_size = batch_size
node_array_size = int(np.ceil(len(x) / hvd.size()))
self.init_array = hvd.rank() * node_array_size
self.end_array = self.init_array + node_array_size
self.x = x
self.y = y
print("calculating size")
print("size", len(self))
def lr_schedule(epoch):
"""Learning Rate Schedule
Learning rate is scheduled to be reduced after 80, 120, 160, 180 epochs.
Called automatically every epoch as part of callbacks during training.
# Arguments
epoch (int): The number of epochs
# Returns
lr (float32): learning rate
"""
if epoch <= 20: #was 5
# bypass to the warmup callback
return K.get_value(model.optimizer.lr)
if epoch <= 40:
return 0.08 * hvd.size() #was 80, 0.01
if epoch <= 60:
return 0.01 * hvd.size() #was 120 0.002
if epoch <= 70:
return 0.002 * hvd.size() #was 160 0.0004
return 0.0004 * hvd.size()
def create_model():
opt = keras.optimizers.SGD(lr=0.01 * hvd.size(), momentum=0.9)
opt = hvd.DistributedOptimizer(opt)
model = keras.models.Sequential()
model.add(keras.layers.Dense(2, input_shape=(3,)))
model.add(keras.layers.RepeatVector(3))
model.add(keras.layers.TimeDistributed(keras.layers.Dense(3)))
model.compile(loss=keras.losses.MSE,
optimizer=opt,
metrics=[keras.metrics.categorical_accuracy],
sample_weight_mode='temporal')
return model
def create_model():
opt = keras.optimizers.SGD(lr=0.01 * hvd.size(), momentum=0.9)
opt = hvd.DistributedOptimizer(opt)
model = keras.models.Sequential()
model.add(keras.layers.Dense(2, input_shape=(3,)))
model.add(keras.layers.RepeatVector(3))
model.add(keras.layers.TimeDistributed(keras.layers.Dense(3)))
model.compile(loss=keras.losses.MSE,
optimizer=opt,
metrics=[keras.metrics.categorical_accuracy],
sample_weight_mode='temporal')
return model
def data_generator(file_path, batch_size, seq_len=512, predict=False):
# Trick the code into thinking we're only running 1 process for prediction when running `Metrics`.
if predict:
size = 1
else:
size = hvd.size()
total_batch_size = batch_size * size
print(total_batch_size)
rank = hvd.rank()
print(rank)
range_start = batch_size * rank
range_end = range_start + batch_size
print(range_start, range_end)
while True:
with xopen(file_path, "rt") as f:
_, label_dim = json.loads(f.readline())
text = []
labels = []
for line in f:
if len(text) == total_batch_size:
text = text[range_start:range_end]
labels = labels[range_start:range_end]
print(text[0])
# Fun fact: the 2 inputs must be in a list, *not* a tuple. Why.
yield ([np.asarray(text), np.zeros_like(text)], np.asarray(labels))
text = []
labels = []
line = json.loads(line)
# First sublist is token ids.
text.append(np.asarray(line[0])[0:seq_len])
# Second sublist is positive label indices.
label_line = np.zeros(label_dim, dtype='b')
label_line[line[1]] = 1
labels.append(label_line)
# Yield what is left as the last batch when file has been read to its end.
# Split the remaining examples, duplicating with `ceil()` if they don't split evenly.
leftover_batch_start = ceil(len(text) / size) * rank
leftover_batch_end = leftover_batch_start + ceil(len(text) / size)
text = text[leftover_batch_start:leftover_batch_end]
labels = labels[leftover_batch_start:leftover_batch_end]
yield ([np.asarray(text), np.zeros_like(text)], np.asarray(labels))
def create_inception_model(self, number_categories, dense_layer_sizes, dropout_fraction, unfrozen_layers, focal_loss=False):
hvd.init()
config = tf.compat.v1.ConfigProto()
config.gpu_options.visible_device_list = str(hvd.local_rank())
opt = hvd.DistributedOptimizer(tf.keras.optimizers.Adam(learning_rate=0.001*hvd.size()))
model = InceptionV3(include_top=False, pooling='avg')
output = model.outputs[0]
for layer_size in dense_layer_sizes:
dense = Dense(layer_size, activation='relu')(output)
dropout = Dropout(dropout_fraction)(dense)
output = BatchNormalization()(dropout)
if number_categories == 1:
output = Dense(1, activation='sigmoid')(output)
else:
output = Dense(number_categories, activation='softmax')(output)
model = Model(inputs=model.inputs, outputs=output)
for index in range(len(model.layers) - unfrozen_layers):
model.layers[index].trainable = False
if number_categories == 1:
the_metrics = [metrics.binary_accuracy]
if focal_loss:
loss = customlosses.focal_binary_crossentropy
else:
loss = 'binary_crossentropy'
else:
the_metrics = [metrics.categorical_accuracy]
if focal_loss:
loss = customlosses.focal_categorical_crossentropy
else:
loss = 'categorical_crossentropy'
model.compile(optimizer=opt, loss=loss, metrics=the_metrics)
model.save(self.model_filename)
self.model = model
def batch_generator(full_sequences, fragment_length, batch_size,
fragment_stride, nb_output_bins, randomize_batch_order,
_rnd):
indices = list(
fragment_indices(full_sequences, fragment_length, batch_size,
fragment_stride, nb_output_bins))
global g_multi_gpu
if g_multi_gpu:
import horovod.keras as hvd
gpu_count = hvd.size()
current_gpu = hvd.rank()
else:
gpu_count = 1
current_gpu = 0
if randomize_batch_order:
_rnd.shuffle(indices)
batches_parted = [batch for batch in partition_all(batch_size, indices)]
start_index = len(batches_parted) // gpu_count * current_gpu
batches_gpu = batches_parted[start_index:]
batches = cycle(batches_gpu)
for batch in batches:
if len(batch) < batch_size:
continue
yield np.array([
one_hot(full_sequences[e[0]][e[1]:e[1] + fragment_length])
for e in batch
],
dtype='uint8'), np.array([
one_hot(full_sequences[e[0]][e[1] + 1:e[1] +
fragment_length + 1])
for e in batch
],
dtype='uint8')
# Load the data files
train_file = os.path.join(input_dir, 'train.h5')
valid_file = os.path.join(input_dir, 'val.h5')
test_file = os.path.join(input_dir, 'test.h5')
train_input, train_labels, train_weights = load_file(train_file, n_train)
valid_input, valid_labels, valid_weights = load_file(valid_file, n_valid)
test_input, test_labels, test_weights = load_file(test_file, n_test)
print('train shape:', train_input.shape, 'Mean label:', train_labels.mean())
print('valid shape:', valid_input.shape, 'Mean label:', valid_labels.mean())
print('test shape: ', test_input.shape, 'Mean label:', test_labels.mean())
# Model config
conv_sizes = [8, 16, 32]
fc_sizes = [64]
optimizer = 'Adam'
lr = 0.01 * hvd.size()
dropout = 0.5
# Training config
batch_size = 32 #128
n_epochs = 8
# Build the model
model = build_model(train_input.shape[1:],
conv_sizes=conv_sizes,
fc_sizes=fc_sizes,
dropout=dropout,
optimizer=optimizer,
lr=lr)
if hvd.rank() == 0:
model.summary()
def main(args):
if 'sourcedir.tar.gz' in args.tensorboard_dir:
tensorboard_dir = re.sub('source/sourcedir.tar.gz', 'model',
args.tensorboard_dir)
else:
tensorboard_dir = args.tensorboard_dir
logging.info("Writing TensorBoard logs to {}".format(tensorboard_dir))
if os.path.isdir(args.checkpoint_path):
logging.info("Checkpointing directory {} exists".format(
args.checkpoint_path))
else:
logging.info("Creating Checkpointing directory {}".format(
args.checkpoint_path))
os.mkdir(args.checkpoint_path)
mpi = False
if 'sagemaker_mpi_enabled' in args.fw_params:
if args.fw_params['sagemaker_mpi_enabled']:
import horovod.keras as hvd
mpi = True
# 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))
else:
hvd = None
logging.info("Running with MPI={}".format(mpi))
logging.info("getting data")
train_dataset = train_input_fn()
eval_dataset = eval_input_fn()
validation_dataset = validation_input_fn()
logging.info("configuring model")
# Load model
if not os.listdir(args.checkpoint_path):
model = keras_model_fn(args.learning_rate, args.weight_decay,
args.optimizer, args.momentum, mpi, hvd)
epoch_number = 0
else:
model, epoch_number = load_checkpoint_model(args.checkpoint_path)
logging.info("Checkpointing to: {}".format(args.checkpoint_path))
callbacks = []
if mpi:
callbacks.append(hvd.callbacks.BroadcastGlobalVariablesCallback(0))
callbacks.append(hvd.callbacks.MetricAverageCallback())
callbacks.append(
hvd.callbacks.LearningRateWarmupCallback(warmup_epochs=5,
verbose=1))
callbacks.append(
keras.callbacks.ReduceLROnPlateau(patience=10, verbose=1))
if hvd.rank() == 0:
callbacks.append(
ModelCheckpoint(args.checkpoint_path +
'/checkpoint-{epoch}.h5'))
callbacks.append(
TensorBoard(log_dir=tensorboard_dir, update_freq='epoch'))
else:
callbacks.append(
keras.callbacks.ReduceLROnPlateau(patience=10, verbose=1))
callbacks.append(
ModelCheckpoint(args.checkpoint_path + '/checkpoint-{epoch}.h5'))
callbacks.append(
TensorBoard(log_dir=tensorboard_dir, update_freq='epoch'))
logging.info("Starting training")
size = 1
if mpi:
size = hvd.size()
model.fit(
x=train_dataset[0],
y=train_dataset[1],
steps_per_epoch=(num_examples_per_epoch('train') // args.batch_size) //
size,
epochs=args.epochs,
initial_epoch=epoch_number,
validation_data=validation_dataset,
validation_steps=(num_examples_per_epoch('validation') //
args.batch_size) // size,
callbacks=callbacks)
score = model.evaluate(eval_dataset[0],
eval_dataset[1],
steps=num_examples_per_epoch('eval') //
args.batch_size,
verbose=0)
logging.info('Test loss:{}'.format(score[0]))
logging.info('Test accuracy:{}'.format(score[1]))
# Horovod: Save model only on worker 0 (i.e. master)
if mpi:
if hvd.rank() == 0:
save_model(model, args.model_output_dir)
else:
save_model(model, args.model_output_dir)
y_train = keras.utils.to_categorical(y_train, num_classes)
y_test = keras.utils.to_categorical(y_test, num_classes)
model = Sequential()
model.add(
Conv2D(32, kernel_size=(3, 3), activation='relu', input_shape=input_shape))
model.add(Conv2D(64, (3, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.25))
model.add(Flatten())
model.add(Dense(128, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(num_classes, activation='softmax'))
# Horovod: adjust learning rate based on number of GPUs.
opt = keras.optimizers.Adadelta(lr=args.lr * hvd.size())
# Horovod: add Horovod Distributed Optimizer.
opt = hvd.DistributedOptimizer(opt)
model.compile(loss=keras.losses.categorical_crossentropy,
optimizer=opt,
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: average metrics among workers at the end of every epoch.
def _run():
import keras
import models
logger = tk.log.get(__name__)
parser = argparse.ArgumentParser()
parser.add_argument('--epochs', help='epoch数。', default=300, type=int)
parser.add_argument('--batch-size', help='バッチサイズ。', default=16, type=int)
parser.add_argument('--warm',
help='models/model.fold{cv_index}.h5を読み込む',
action='store_true',
default=False)
parser.add_argument('--cv-index', help='CVの何番目か。', type=int)
parser.add_argument('--cv-size', help='CVの分割数。', default=5, type=int)
parser.add_argument('--split-seed', help='分割のシード値。', default=123, type=int)
args = parser.parse_args()
assert args.cv_index in range(args.cv_size)
model_path = _MODELS_DIR / 'model.fold{}.h5'.format(args.cv_index)
(X_train,
y_train), (X_val, y_val), _ = data.load_data(args.cv_index, args.cv_size,
args.split_seed)
num_classes = len(np.unique(y_train))
y_train = tk.ml.to_categorical(num_classes)(y_train)
y_val = tk.ml.to_categorical(num_classes)(y_val)
logger.info('len(X_train) = {} len(X_val) = {}'.format(
len(X_train), len(X_val)))
model = models.create_network(num_classes)
# 学習率：
# ・lr 0.5、batch size 256くらいが多いのでその辺を基準に
# ・バッチサイズに比例させるのが良いとのうわさ
lr = 0.5 * args.batch_size / 256 * hvd.size()
opt = keras.optimizers.SGD(lr=lr, momentum=0.9, nesterov=True)
opt = hvd.DistributedOptimizer(opt)
model.compile(opt, 'categorical_crossentropy', ['acc'])
if hvd.rank() == 0 and args.cv_index == 0:
model.summary(print_fn=logger.info)
logger.info('network depth: %d', tk.dl.count_network_depth(model))
if args.warm:
model.load_weights(str(model_path))
logger.info('{} loaded'.format(model_path))
else:
assert not model_path.exists() # 誤操作対策
callbacks = []
if args.warm and args.epochs < 300: # 短縮モード
callbacks.append(tk.dl.learning_rate_callback((0, 0.5)))
else:
callbacks.append(tk.dl.learning_rate_callback())
callbacks.append(hvd.callbacks.BroadcastGlobalVariablesCallback(0))
callbacks.append(hvd.callbacks.MetricAverageCallback())
callbacks.append(
hvd.callbacks.LearningRateWarmupCallback(warmup_epochs=5, verbose=1))
if hvd.rank() == 0:
callbacks.append(tk.dl.tsv_log_callback(_MODELS_DIR / 'history.tsv'))
callbacks.append(tk.dl.freeze_bn_callback(0.95))
gen = models.create_generator((299, 299), mixup=True)
model.fit_generator(
gen.flow(X_train,
y_train,
batch_size=args.batch_size,
data_augmentation=True,
shuffle=True),
steps_per_epoch=gen.steps_per_epoch(len(X_train), args.batch_size) //
hvd.size(),
epochs=args.epochs,
verbose=1 if hvd.rank() == 0 else 0,
validation_data=gen.flow(X_val,
y_val,
batch_size=args.batch_size,
shuffle=True),
validation_steps=gen.steps_per_epoch(len(X_val), args.batch_size) //
hvd.size(), # * 3は省略
callbacks=callbacks)
if hvd.rank() == 0:
model.save(str(model_path))
proba_val = model.predict_generator(
gen.flow(X_val, y_val, batch_size=args.batch_size),
gen.steps_per_epoch(len(X_val), args.batch_size),
verbose=1)
joblib.dump(proba_val,
_MODELS_DIR / 'proba_val.fold{}.pkl'.format(args.cv_index))
pred_val = proba_val.argmax(axis=-1)
logger.info('val_acc: {:.1f}%'.format(
sklearn.metrics.accuracy_score(y_val.argmax(axis=-1), pred_val) *
100))
def __init__(self, config_file, resume_training=True, resume_epoch=None, predict_length=None, multi_gpu=False):
self.config = ConfigParser.ConfigParser(allow_no_value=True)
try:
self.config.readfp(open(config_file))
except:
print('Could not read configuration file {} - exiting.'.format(config_file))
sys.exit(1)
# Get General Configuration
self.train_multi_gpu = multi_gpu
self.resume_training = resume_training
self.resume_epoch = resume_epoch
self.keras_verbose = self.config.getint('general', 'keras_verbose')
self.seed = self.config.getint('general', 'seed')
if self.seed is None:
self.seed = 42
# Get Model Configuration
self.data_dir = self.config.get('model', 'data_dir')
self.data_dir_structure = self.config.get('model', 'data_dir_structure')
self.model_dir = self.config.get('model', 'model_dir')
if len(self.model_dir) == 0:
self.model_dir = None
self.sample_rate = self.config.getint('model', 'sample_rate')
self.debug = self.config.getint('model', 'debug')
# Training Configuration
self.max_epoch = self.config.getint('training', 'max_epoch')
self.test_factor = self.config.getfloat('training', 'test_factor')
self.batch_size = self.config.getint('training', 'batch_size')
self.output_bins = self.config.getint('training', 'output_bins')
self.filters = self.config.getint('training', 'filters')
self.dilation_depth = self.config.getint('training', 'dilation_depth')
self.stacks = self.config.getint('training', 'stacks')
self.use_bias = self.config.getboolean('training', 'use_bias')
self.use_ulaw = self.config.getboolean('training', 'use_ulaw')
self.res_l2 = self.config.getfloat('training', 'res_l2')
self.final_l2 = self.config.getfloat('training', 'final_l2')
self.initial_fragment_length = self.config.getint('training', 'initial_fragment_length')
self.fragment_stride = self.config.getint('training', 'fragment_stride')
self.use_skip_connections = self.config.getboolean('training', 'use_skip_connections')
self.learn_all_outputs = self.config.getboolean('training', 'learn_all_outputs')
self.random_train_batches = self.config.getboolean('training', 'random_train_batches')
self.randomize_batch_order = self.config.getboolean('training', 'randomize_batch_order')
self.train_only_in_receptive_field = self.config.getboolean('training', 'train_only_in_receptive_field')
self.train_with_soft_targets = self.config.getboolean('training', 'train_with_soft_targets')
self.soft_target_stdev = self.config.getfloat('training', 'soft_target_stdev')
self.optimizer = self.config.get('training', 'optimizer')
self.early_stopping_patience = self.config.getint('training', 'early_stopping_patience')
# Prediction Configuration
self.predict_length = self.config.getfloat('prediction', 'predict_length')
# Let's allow the user to overwrite the length via cmd-line, it is more practical :-)
if predict_length is not None:
self.predict_length = predict_length
self.sample_argmax = self.config.getboolean('prediction', 'sample_argmax')
self.sample_temperature = self.config.getfloat('prediction', 'sample_temperature')
if self.sample_temperature < 0.001:
self.sample_temperature = None
self.predict_initial_input = self.config.get('prediction', 'initial_input')
if len(self.predict_initial_input) == 0:
self.predict_initial_input = None
self.predict_use_softmax_as_input = self.config.getboolean('prediction', 'use_softmax_as_input')
self.sample_seed = self.seed
np.random.seed(self.seed)
self.rnd = np.random.RandomState(self.seed)
self.fragment_length = self.initial_fragment_length + self._compute_receptive_field2(self.sample_rate, self.dilation_depth, self.stacks)[0]
# Additional Settings
self.num_gpus = 1
self.train_rank = 0
if self.train_multi_gpu:
self.train_rank = hvd.rank()
self.num_gpus = hvd.size()
print('rank = {}, num_gpu={}'.format(self.train_rank, self.num_gpus))
self.dataset = DataSet(self.config, self.fragment_length, self.num_gpus, self.train_rank)
# Add L2 weight decay & adjust BN settings.
model_config = model.get_config()
for layer, layer_config in zip(model.layers, model_config['layers']):
if hasattr(layer, 'kernel_regularizer'):
regularizer = keras.regularizers.l2(args.wd)
layer_config['config']['kernel_regularizer'] = \
{'class_name': regularizer.__class__.__name__,
'config': regularizer.get_config()}
if type(layer) == keras.layers.BatchNormalization:
layer_config['config']['momentum'] = 0.9
layer_config['config']['epsilon'] = 1e-5
model = keras.models.Model.from_config(model_config)
# Horovod: adjust learning rate based on number of GPUs.
opt = keras.optimizers.SGD(lr=args.base_lr * hvd.size(),
momentum=args.momentum)
# Horovod: add Horovod Distributed Optimizer.
opt = hvd.DistributedOptimizer(opt, compression=compression)
model.compile(loss=keras.losses.categorical_crossentropy,
optimizer=opt,
metrics=['accuracy', 'top_k_categorical_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),
import horovod.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))
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, y_train), (x_test, y_test) = mnist.load_data()
if K.image_data_format() == 'channels_first':
x_train = x_train.reshape(x_train.shape[0], 1, img_rows, img_cols)
x_test = x_test.reshape(x_test.shape[0], 1, img_rows, img_cols)
input_shape = (1, img_rows, img_cols)
else:
x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, 1)
x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, 1)
input_shape = (img_rows, img_cols, 1)
# We'll save the worker logs and models separately but only
# use the logs/saved model from worker 0.
args.saved_model = "./worker{}/3d_unet_decathlon.hdf5".format(hvd.rank())
# Optimize CPU threads for TensorFlow
CONFIG = tf.ConfigProto(inter_op_parallelism_threads=args.interop_threads,
intra_op_parallelism_threads=args.intraop_threads)
SESS = tf.Session(config=CONFIG)
K.backend.set_session(SESS)
model, opt = unet_3d(
use_upsampling=args.use_upsampling,
n_cl_in=args.number_input_channels,
learning_rate=args.lr * hvd.size(),
n_cl_out=1, # single channel (greyscale)
dropout=0.2,
print_summary=print_summary)
opt = hvd.DistributedOptimizer(opt)
model.compile(
optimizer=opt,
# loss=[combined_dice_ce_loss],
loss=[dice_coef_loss],
metrics=[dice_coef, "accuracy", sensitivity, specificity])
if hvd.rank() == 0:
start_time = datetime.datetime.now()
print("Started script on {}".format(start_time))
#initial_model = create_vgg16()
#initial_model.load_weights(model_path) # we may begin from scratch
#x = Dense(batches.num_class, activation='softmax')(initial_model.layers[-2].output)
#model = Model(initial_model.input, x)
# for layer in initial_model.layers: layer.trainable=False # for scratch build
#opt = Adam(lr=0.001)
opt = SGD(lr=0.01)
opt = hvd.DistributedOptimizer(opt)
callbacks = [
hvd.callbacks.BroadcastGlobalVariablesCallback(0),
hvd.callbacks.MetricAverageCallback(),
hvd.callbacks.LearningRateWarmupCallback(warmup_epochs=2, verbose=1),
keras.callbacks.ReduceLROnPlateau(patience=3, verbose=1),
]
model.compile(optimizer=opt,
loss='categorical_crossentropy',
metrics=['accuracy'])
model.fit_generator(batches, steps_per_epoch=batches.samples//batch_size // hvd.size(), nb_epoch=10,
validation_data=valid_batches, validation_steps=valid_batches.samples//batch_size//hvd.size())
if hvd.rank() == 0:
model_json = model.to_json()
with open("model.json",'w') as json_file:
json_file.write(model_json)
model.save_weights("model_first.h5")
print("Saved mode in the first step")
y_test = keras.utils.to_categorical(y_test, num_classes)
model = Sequential()
model.add(Conv2D(32, kernel_size=(3, 3),
activation='relu',
input_shape=input_shape))
model.add(Conv2D(64, (3, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.25))
model.add(Flatten())
model.add(Dense(128, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(num_classes, activation='softmax'))
# Horovod: adjust learning rate based on number of GPUs.
opt = keras.optimizers.Adadelta(lr=1.0 * hvd.size())
# Horovod: add Horovod Distributed Optimizer.
opt = hvd.DistributedOptimizer(opt)
model.compile(loss=keras.losses.categorical_crossentropy,
optimizer=opt,
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: average metrics among workers at the end of every epoch.
"validate_test_split": args.validate_test_split,
"augment": False,
"shuffle": False,
"seed": args.random_seed
}
validation_generator = DataGenerator("validate", args.data_path,
**validation_data_params)
if (hvd.rank() == 0):
validation_generator.print_info()
# Fit the model
# Do at least 3 steps for training and validation
steps_per_epoch = max(
3,
training_generator.get_length() // (args.bz * hvd.size()))
validation_steps = max(
3, 3 * training_generator.get_length() // (args.bz * hvd.size()))
unet_model.model.fit_generator(
training_generator,
steps_per_epoch=steps_per_epoch,
epochs=args.epochs,
verbose=verbose,
validation_data=validation_generator,
#validation_steps=validation_steps,
callbacks=callbacks,
max_queue_size=1, #args.num_prefetched_batches,
workers=1, #args.num_data_loaders,
use_multiprocessing=True)
from prednet import PredNet
from data_utils import SequenceGenerator
from kitti_settings import *
import datetime
import horovod.keras as hvd
import keras
import tensorflow as tf
#Horovod:initialize horovod
hvd.init()
#Horovod: pin GPU to be used for 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))
print("horovode size", hvd.size())
save_model = True # if weights will be saved
weights_file = os.path.join(
WEIGHTS_DIR, 'prednet_kitti_weights.hdf5') # where weights will be saved
json_file = os.path.join(WEIGHTS_DIR, 'prednet_kitti_model.json')
if not os.path.exists(WEIGHTS_DIR): os.mkdir(WEIGHTS_DIR)
# Data files
train_file = os.path.join(DATA_DIR, 'X_train.hkl')
train_sources = os.path.join(DATA_DIR, 'sources_train.hkl')
val_file = os.path.join(DATA_DIR, 'X_val.hkl')
val_sources = os.path.join(DATA_DIR, 'sources_val.hkl')
# Training parameters
nb_epoch = 10 #original: 150; for all tests so far set to 100; t2onlyMax: 150
batch_size = 15
y_test = keras.utils.to_categorical(y_test, num_classes)
model = Sequential()
model.add(Conv2D(32, kernel_size=(3, 3),
activation='relu',
input_shape=input_shape))
model.add(Conv2D(64, (3, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.25))
model.add(Flatten())
model.add(Dense(128, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(num_classes, activation='softmax'))
# Horovod: adjust learning rate based on number of GPUs.
opt = keras.optimizers.Adadelta(1.0 * hvd.size())
# Horovod: add Horovod Distributed Optimizer.
opt = hvd.DistributedOptimizer(opt)
model.compile(loss=keras.losses.categorical_crossentropy,
optimizer=opt,
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),
]
# Add L2 weight decay & adjust BN settings.
model_config = model.get_config()
for layer, layer_config in zip(model.layers, model_config['layers']):
if hasattr(layer, 'kernel_regularizer'):
regularizer = keras.regularizers.l2(args.wd)
layer_config['config']['kernel_regularizer'] = \
{'class_name': regularizer.__class__.__name__,
'config': regularizer.get_config()}
if type(layer) == keras.layers.BatchNormalization:
layer_config['config']['momentum'] = 0.9
layer_config['config']['epsilon'] = 1e-5
model = keras.models.Model.from_config(model_config)
# Horovod: adjust learning rate based on number of GPUs.
opt = keras.optimizers.SGD(lr=args.base_lr * hvd.size(),
momentum=args.momentum)
# Horovod: add Horovod Distributed Optimizer.
opt = hvd.DistributedOptimizer(opt)
model.compile(loss=keras.losses.categorical_crossentropy,
optimizer=opt,
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),
y_train = keras.utils.to_categorical(y_train, num_classes)
y_test = keras.utils.to_categorical(y_test, num_classes)
model = Sequential()
model.add(
Conv2D(32, kernel_size=(3, 3), activation='relu', input_shape=input_shape))
model.add(Conv2D(64, (3, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.25))
model.add(Flatten())
model.add(Dense(128, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(num_classes, activation='softmax'))
# Horovod: adjust learning rate based on number of GPUs.
opt = keras.optimizers.Adadelta(1.0 * hvd.size())
# Horovod: add Horovod Distributed Optimizer.
opt = hvd.DistributedOptimizer(opt)
model.compile(loss=keras.losses.categorical_crossentropy,
optimizer=opt,
metrics=['accuracy'])
log_dir = "../logs/keras-tensorboard-profile/" + datetime.now().strftime(
"%Y%m%d-%H%M%S")
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),
def train(self):
train_data_generator = self.data_loader.get_train_data_generator()
batch_size = self.config.trainer.batch_size
steps_per_epoch = self.data_loader.get_train_data_size() // batch_size
if self.config.trainer.use_horovod:
import horovod.keras as hvd
steps_per_epoch //= hvd.size()
assert steps_per_epoch > 0
valid_data_generator = self.data_loader.get_validation_data_generator()
valid_data_size = self.data_loader.get_validation_data_size()
fake_x_pool = FakeImagePool(self.config.trainer.fake_pool_size)
fake_y_pool = FakeImagePool(self.config.trainer.fake_pool_size)
batch_shape = (self.config.trainer.batch_size,
self.config.dataset.image_size // 8, self.config.dataset.image_size // 8, 1)
fake = np.zeros(shape=batch_shape, dtype=np.float32)
real = np.ones(shape=batch_shape, dtype=np.float32)
epochs = self.config.trainer.num_epochs
start_time = datetime.datetime.now()
self.on_train_begin()
for epoch in range(self.config.trainer.epoch_to_continue, epochs):
self.on_epoch_begin(epoch, {})
epoch_logs = defaultdict(float)
for step in range(1, steps_per_epoch + 1):
batch_logs = {'batch': step, 'size': self.config.trainer.batch_size}
batch_logs = {"batch": step, "size": self.config.trainer.batch_size}
self.on_batch_begin(step, batch_logs)
imgs_x, imgs_y = next(train_data_generator)
fakes_y = self.g_xy.predict(imgs_x)
fakes_x = self.g_yx.predict(imgs_y)
# train discriminator using history of fake images (Shrivastava et al)
fakes_x = fake_x_pool.query(fakes_x)
fakes_y = fake_y_pool.query(fakes_y)
if self.config.trainer.label_smoothing:
fake = np.random.uniform(0, 0.2, size=batch_shape)
real = np.random.uniform(0.8, 1.0, size=batch_shape)
# train discriminator
dx_loss_real = self.d_x.train_on_batch(imgs_x, real)
dx_loss_fake = self.d_x.train_on_batch(fakes_x, fake)
dy_loss_real = self.d_y.train_on_batch(imgs_y, real)
dy_loss_fake = self.d_y.train_on_batch(fakes_y, fake)
# train generator
g_loss = self.combined.train_on_batch([imgs_x, imgs_y], [real, real, imgs_x, imgs_y, imgs_x, imgs_y])
dx_metric_names = self.d_metric_names("x")
dy_metric_names = self.d_metric_names("y")
g_metric_names = self.g_metric_names()
assert len(dx_metric_names) == len(dx_loss_real) == len(dx_loss_fake)
assert len(dy_metric_names) == len(dy_loss_real) == len(dy_loss_fake)
assert len(g_metric_names) == len(g_loss)
metric_logs = {}
for metric_name, metric_value in zip(dx_metric_names + dy_metric_names,
dx_loss_real + dy_loss_real):
metric_logs[f"train/{metric_name}_real"] = \
metric_value * (100 if "accuracy" in metric_name.lower() else 1)
for metric_name, metric_value in zip(dx_metric_names + dy_metric_names,
dx_loss_fake + dy_loss_fake):
metric_logs[f"train/{metric_name}_fake"] = \
metric_value * (100 if "accuracy" in metric_name.lower() else 1)
for metric_name, metric_value in zip(g_metric_names, g_loss):
metric_logs[f"train/{metric_name}"] = metric_value
batch_logs.update(metric_logs)
for metric_name in metric_logs.keys():
if metric_name in epoch_logs:
epoch_logs[metric_name] += metric_logs[metric_name]
else:
epoch_logs[metric_name] = metric_logs[metric_name]
print_str = f"[Epoch {epoch + 1}/{epochs}] [Batch {step}/{steps_per_epoch}]"
deliminator = ' '
for metric_name, metric_value in metric_logs.items():
if 'accuracy' in metric_name:
print_str += f"{deliminator}{metric_name}={metric_value:.1f}%"
elif 'loss' in metric_name:
print_str += f"{deliminator}{metric_name}={metric_value:.4f}"
else:
print_str += f"{deliminator}{metric_name}={metric_value}"
if deliminator == ' ':
deliminator = ',\t'
print_str += f", time: {datetime.datetime.now() - start_time}"
print(print_str, flush=True)
self.on_batch_end(step, batch_logs)
# sum to average
for k in epoch_logs:
epoch_logs[k] /= steps_per_epoch
epoch_logs = dict(epoch_logs)
# additional log
epoch_logs['train/lr/G'] = K.get_value(self.combined.optimizer.lr)
epoch_logs['train/lr/D_x'] = K.get_value(self.d_x.optimizer.lr)
epoch_logs['train/lr/D_y'] = K.get_value(self.d_y.optimizer.lr)
self.on_epoch_end(epoch, epoch_logs)
if (epoch + 1) % self.config.trainer.predict_freq == 0:
self.sample_valid_images(epoch, valid_data_generator, valid_data_size)
self.predict_test_images(epochs)
self.on_train_end()
# Add L2 weight decay & adjust BN settings.
model_config = model.get_config()
for layer, layer_config in zip(model.layers, model_config['layers']):
if hasattr(layer, 'kernel_regularizer'):
regularizer = keras.regularizers.l2(weight_decay)
layer_config['config']['kernel_regularizer'] = \
{'class_name': regularizer.__class__.__name__,
'config': regularizer.get_config()}
if type(layer) == keras.layers.BatchNormalization:
layer_config['config']['momentum'] = 0.9
layer_config['config']['epsilon'] = 1e-5
model = keras.models.Model.from_config(model_config)
# Horovod: adjust learning rate based on number of GPUs.
opt = keras.optimizers.SGD(lr=learning_rate * hvd.size(), momentum=0.9)
# Horovod: add Horovod Distributed Optimizer.
opt = hvd.DistributedOptimizer(opt)
model.compile(loss=keras.losses.categorical_crossentropy,
optimizer=opt,
metrics=['accuracy', 'top_k_categorical_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: average metrics among workers at the end of every epoch.
