def main(args):
annotation_file = args.annotation_file
log_dir = os.path.join('logs', '000')
classes_path = args.classes_path
class_names = get_classes(classes_path)
num_classes = len(class_names)
print('classes_path =', classes_path)
print('class_names = ', class_names)
print('num_classes = ', num_classes)
anchors = get_anchors(args.anchors_path)
num_anchors = len(anchors)
# get freeze level according to CLI option
if args.weights_path:
freeze_level = 0
else:
freeze_level = 1
if args.freeze_level is not None:
freeze_level = args.freeze_level
# callbacks for training process
logging = TensorBoard(log_dir=log_dir,
histogram_freq=0,
write_graph=False,
write_grads=False,
write_images=False,
update_freq='batch')
checkpoint = ModelCheckpoint(os.path.join(
log_dir, 'ep{epoch:03d}-loss{loss:.3f}-val_loss{val_loss:.3f}.h5'),
monitor='val_loss',
verbose=1,
save_weights_only=False,
save_best_only=True,
period=1)
reduce_lr = ReduceLROnPlateau(monitor='val_loss',
factor=0.5,
patience=10,
verbose=1,
cooldown=0,
min_lr=1e-10)
early_stopping = EarlyStopping(monitor='val_loss',
min_delta=0,
patience=50,
verbose=1)
terminate_on_nan = TerminateOnNaN()
callbacks = [
logging, checkpoint, reduce_lr, early_stopping, terminate_on_nan
]
# get train&val dataset
dataset = get_dataset(annotation_file)
if args.val_annotation_file:
val_dataset = get_dataset(args.val_annotation_file)
num_train = len(dataset)
print('num_train = ', num_train)
num_val = len(val_dataset)
dataset.extend(val_dataset)
else:
val_split = args.val_split
num_val = int(len(dataset) * val_split)
num_train = len(dataset) - num_val
# assign multiscale interval
if args.multiscale:
rescale_interval = args.rescale_interval
else:
rescale_interval = -1 #Doesn't rescale
# model input shape check
input_shape = args.model_image_size
assert (input_shape[0] % 32 == 0
and input_shape[1] % 32 == 0), 'Multiples of 32 required'
# get different model type & train&val data generator
if num_anchors == 9:
# YOLOv3 use 9 anchors
get_train_model = get_yolo3_train_model
data_generator = yolo3_data_generator_wrapper
# tf.keras.Sequence style data generator
#train_data_generator = Yolo3DataGenerator(dataset[:num_train], args.batch_size, input_shape, anchors, num_classes, args.enhance_augment, rescale_interval)
#val_data_generator = Yolo3DataGenerator(dataset[num_train:], args.batch_size, input_shape, anchors, num_classes)
tiny_version = False
elif num_anchors == 6:
# Tiny YOLOv3 use 6 anchors
get_train_model = get_yolo3_train_model
data_generator = yolo3_data_generator_wrapper
# tf.keras.Sequence style data generator
#train_data_generator = Yolo3DataGenerator(dataset[:num_train], args.batch_size, input_shape, anchors, num_classes, args.enhance_augment, rescale_interval)
#val_data_generator = Yolo3DataGenerator(dataset[num_train:], args.batch_size, input_shape, anchors, num_classes)
tiny_version = True
elif num_anchors == 5:
# YOLOv2 use 5 anchors
get_train_model = get_yolo2_train_model
data_generator = yolo2_data_generator_wrapper
# tf.keras.Sequence style data generator
#train_data_generator = Yolo2DataGenerator(dataset[:num_train], args.batch_size, input_shape, anchors, num_classes, args.enhance_augment, rescale_interval)
#val_data_generator = Yolo2DataGenerator(dataset[num_train:], args.batch_size, input_shape, anchors, num_classes)
tiny_version = False
else:
raise ValueError('Unsupported anchors number')
# prepare online evaluation callback
if args.eval_online:
eval_callback = EvalCallBack(
args.model_type,
dataset[num_train:],
anchors,
class_names,
args.model_image_size,
args.model_pruning,
log_dir,
eval_epoch_interval=args.eval_epoch_interval,
save_eval_checkpoint=args.save_eval_checkpoint)
callbacks.append(eval_callback)
# prepare train/val data shuffle callback
if args.data_shuffle:
shuffle_callback = DatasetShuffleCallBack(dataset)
callbacks.append(shuffle_callback)
# prepare model pruning config
pruning_end_step = np.ceil(1.0 * num_train / args.batch_size).astype(
np.int32) * args.total_epoch
if args.model_pruning:
pruning_callbacks = [
sparsity.UpdatePruningStep(),
sparsity.PruningSummaries(log_dir=log_dir, profile_batch=0)
]
callbacks = callbacks + pruning_callbacks
# prepare optimizer
optimizer = get_optimizer(args.optimizer,
args.learning_rate,
decay_type=None)
# get train model
model = get_train_model(args.model_type,
anchors,
num_classes,
weights_path=args.weights_path,
freeze_level=freeze_level,
optimizer=optimizer,
label_smoothing=args.label_smoothing,
model_pruning=args.model_pruning,
pruning_end_step=pruning_end_step)
# support multi-gpu training
template_model = None
if args.gpu_num >= 2:
# keep the template model for saving result
template_model = model
model = multi_gpu_model(model, gpus=args.gpu_num)
# recompile multi gpu model
model.compile(optimizer=optimizer,
loss={
'yolo_loss': lambda y_true, y_pred: y_pred
})
model.summary()
# Transfer training some epochs with frozen layers first if needed, to get a stable loss.
initial_epoch = args.init_epoch #####################################################################################################
epochs = initial_epoch + args.transfer_epoch
print("Transfer training stage")
print(
'Train on {} samples, val on {} samples, with batch size {}, input_shape {}.'
.format(num_train, num_val, args.batch_size, input_shape))
#model.fit_generator(train_data_generator,
model.fit_generator(
data_generator(dataset[:num_train], args.batch_size, input_shape,
anchors, num_classes, args.enhance_augment),
steps_per_epoch=max(1, num_train // args.batch_size),
#validation_data=val_data_generator,
validation_data=data_generator(dataset[num_train:], args.batch_size,
input_shape, anchors, num_classes),
validation_steps=max(1, num_val // args.batch_size),
epochs=epochs,
initial_epoch=initial_epoch,
#verbose=1,
workers=1,
use_multiprocessing=False,
max_queue_size=10,
callbacks=callbacks)
# Wait 2 seconds for next stage
time.sleep(2)
if args.decay_type:
# rebuild optimizer to apply learning rate decay, only after
# unfreeze all layers
callbacks.remove(reduce_lr)
steps_per_epoch = max(1, num_train // args.batch_size)
decay_steps = steps_per_epoch * (args.total_epoch - args.init_epoch -
args.transfer_epoch)
optimizer = get_optimizer(args.optimizer,
args.learning_rate,
decay_type=args.decay_type,
decay_steps=decay_steps)
# Unfreeze the whole network for further tuning
# NOTE: more GPU memory is required after unfreezing the body
print("Unfreeze and continue training, to fine-tune.")
for i in range(len(model.layers)):
model.layers[i].trainable = True
model.compile(optimizer=optimizer,
loss={
'yolo_loss': lambda y_true, y_pred: y_pred
}) # recompile to apply the change
print(
'Train on {} samples, val on {} samples, with batch size {}, input_shape {}.'
.format(num_train, num_val, args.batch_size, input_shape))
#model.fit_generator(train_data_generator,
model.fit_generator(
data_generator(dataset[:num_train], args.batch_size, input_shape,
anchors, num_classes, args.enhance_augment,
rescale_interval),
steps_per_epoch=max(1, num_train // args.batch_size),
#validation_data=val_data_generator,
validation_data=data_generator(dataset[num_train:], args.batch_size,
input_shape, anchors, num_classes),
validation_steps=max(1, num_val // args.batch_size),
epochs=args.total_epoch,
initial_epoch=epochs,
#verbose=1,
workers=1,
use_multiprocessing=False,
max_queue_size=10,
callbacks=callbacks)
# Finally store model
if args.model_pruning:
if template_model is not None:
template_model = sparsity.strip_pruning(template_model)
else:
model = sparsity.strip_pruning(model)
if template_model is not None:
template_model.save(os.path.join(log_dir, 'trained_final.h5'))
else:
model.save(os.path.join(log_dir, 'trained_final.h5'))
def main(args):
#데이터 annotation 파일 경로
annotation_file = args.annotation_file
# 결과 log 및 weight가 저장될 경로
log_dir = os.path.join('logs', '000')
#클래스 파일 경로
classes_path = args.classes_path
class_names = get_classes(classes_path)
num_classes = len(class_names)
# anchors 받아오는 라인
anchors = get_anchors(args.anchors_path)
num_anchors = len(anchors)
# get freeze level according to CLI option
if args.weights_path:
freeze_level = 0
else:
freeze_level = 1
if args.freeze_level is not None:
freeze_level = args.freeze_level
# callbacks for training process
logging = TensorBoard(log_dir=log_dir,
histogram_freq=0,
write_graph=False,
write_grads=False,
write_images=False,
update_freq='batch')
checkpoint = ModelCheckpoint(os.path.join(
log_dir, 'ep{epoch:03d}-loss{loss:.3f}-val_loss{val_loss:.3f}.h5'),
monitor='val_loss',
mode='min',
verbose=1,
save_weights_only=False,
save_best_only=True,
period=1)
reduce_lr = ReduceLROnPlateau(monitor='val_loss',
factor=0.5,
mode='min',
patience=10,
verbose=1,
cooldown=0,
min_lr=1e-10)
early_stopping = EarlyStopping(monitor='val_loss',
min_delta=0,
patience=50,
verbose=1,
mode='min')
terminate_on_nan = TerminateOnNaN()
callbacks = [
logging, checkpoint, reduce_lr, early_stopping, terminate_on_nan
]
# 데이터셋 로딩
dataset = get_dataset(annotation_file)
if args.val_annotation_file:
val_dataset = get_dataset(args.val_annotation_file)
num_train = len(dataset)
num_val = len(val_dataset)
dataset.extend(val_dataset)
else:
val_split = args.val_split
num_val = int(len(dataset) * val_split)
num_train = len(dataset) - num_val
# assign multiscale interval
if args.multiscale:
rescale_interval = args.rescale_interval
else:
rescale_interval = -1 #Doesn't rescale
# model input shape check
input_shape = args.model_image_size
assert (input_shape[0] % 32 == 0 and input_shape[1] % 32
== 0), 'model_image_size should be multiples of 32'
# 모델종류에 따른 data generator 및 모델 생성
if num_anchors == 9:
# YOLOv3 use 9 anchors
get_train_model = get_yolo3_train_model
data_generator = yolo3_data_generator_wrapper
tiny_version = False
elif num_anchors == 6:
# Tiny YOLOv3 use 6 anchors
get_train_model = get_yolo3_train_model
data_generator = yolo3_data_generator_wrapper
tiny_version = True
elif num_anchors == 5:
# YOLOv2 use 5 anchors
get_train_model = get_yolo2_train_model
data_generator = yolo2_data_generator_wrapper
tiny_version = False
else:
raise ValueError('Unsupported anchors number')
# prepare online evaluation callback
if args.eval_online:
eval_callback = EvalCallBack(
args.model_type,
dataset[num_train:],
anchors,
class_names,
args.model_image_size,
args.model_pruning,
log_dir,
eval_epoch_interval=args.eval_epoch_interval,
save_eval_checkpoint=args.save_eval_checkpoint,
elim_grid_sense=args.elim_grid_sense)
callbacks.append(eval_callback)
# prepare train/val data shuffle callback
if args.data_shuffle:
shuffle_callback = DatasetShuffleCallBack(dataset)
callbacks.append(shuffle_callback)
# prepare model pruning config
pruning_end_step = np.ceil(1.0 * num_train / args.batch_size).astype(
np.int32) * args.total_epoch
if args.model_pruning:
pruning_callbacks = [
sparsity.UpdatePruningStep(),
sparsity.PruningSummaries(log_dir=log_dir, profile_batch=0)
]
callbacks = callbacks + pruning_callbacks
# prepare optimizer
optimizer = get_optimizer(args.optimizer,
args.learning_rate,
decay_type=None)
# support multi-gpu training
if args.gpu_num >= 2:
# devices_list=["/gpu:0", "/gpu:1"]
devices_list = ["/gpu:{}".format(n) for n in range(args.gpu_num)]
strategy = tf.distribute.MirroredStrategy(devices=devices_list)
print('Number of devices: {}'.format(strategy.num_replicas_in_sync))
with strategy.scope():
# get multi-gpu train model
model = get_train_model(args.model_type,
anchors,
num_classes,
weights_path=args.weights_path,
freeze_level=freeze_level,
optimizer=optimizer,
label_smoothing=args.label_smoothing,
elim_grid_sense=args.elim_grid_sense,
model_pruning=args.model_pruning,
pruning_end_step=pruning_end_step)
else:
# get normal train model
model = get_train_model(args.model_type,
anchors,
num_classes,
weights_path=args.weights_path,
freeze_level=freeze_level,
optimizer=optimizer,
label_smoothing=args.label_smoothing,
elim_grid_sense=args.elim_grid_sense,
model_pruning=args.model_pruning,
pruning_end_step=pruning_end_step)
model.summary()
# Transfer training some epochs with frozen layers first if needed, to get a stable loss.
initial_epoch = args.init_epoch
epochs = initial_epoch + args.transfer_epoch
print("Transfer training stage")
print(
'Train on {} samples, val on {} samples, with batch size {}, input_shape {}.'
.format(num_train, num_val, args.batch_size, input_shape))
# 성능향상을 위해 초반 일부 epoch은 Transfer Learning 진행 (Initial Epoch ~ Transfer Epoch)
model.fit_generator(
data_generator(dataset[:num_train],
args.batch_size,
input_shape,
anchors,
num_classes,
args.enhance_augment,
rescale_interval,
multi_anchor_assign=args.multi_anchor_assign),
steps_per_epoch=max(1, num_train // args.batch_size),
#validation_data=val_data_generator,
validation_data=data_generator(
dataset[num_train:],
args.batch_size,
input_shape,
anchors,
num_classes,
multi_anchor_assign=args.multi_anchor_assign),
validation_steps=max(1, num_val // args.batch_size),
epochs=epochs,
initial_epoch=initial_epoch,
#verbose=1,
workers=1,
use_multiprocessing=False,
max_queue_size=10,
callbacks=callbacks)
# Wait 2 seconds for next stage
time.sleep(2)
if args.decay_type:
# rebuild optimizer to apply learning rate decay, only after
# unfreeze all layers
callbacks.remove(reduce_lr)
steps_per_epoch = max(1, num_train // args.batch_size)
decay_steps = steps_per_epoch * (args.total_epoch - args.init_epoch -
args.transfer_epoch)
optimizer = get_optimizer(args.optimizer,
args.learning_rate,
decay_type=args.decay_type,
decay_steps=decay_steps)
# Unfreeze the whole network for further tuning
# NOTE: more GPU memory is required after unfreezing the body
print("Unfreeze and continue training, to fine-tune.")
if args.gpu_num >= 2:
with strategy.scope():
for i in range(len(model.layers)):
model.layers[i].trainable = True
model.compile(optimizer=optimizer,
loss={
'yolo_loss': lambda y_true, y_pred: y_pred
}) # recompile to apply the change
else:
for i in range(len(model.layers)):
model.layers[i].trainable = True
model.compile(optimizer=optimizer,
loss={
'yolo_loss': lambda y_true, y_pred: y_pred
}) # recompile to apply the change
print(
'Train on {} samples, val on {} samples, with batch size {}, input_shape {}.'
.format(num_train, num_val, args.batch_size, input_shape))
# Transfer Learning 이후 나머지 Epoch에 대하여 학습 진행 (Transfer Epoch ~ Total Epoch)
# 이 부분이 필요없거나 학습 시간이 너무 오래 걸릴 경우 Total Epoch을 Transfer와 동일하게 두고, 아래 학습을 진행하지 않고 넘어갈 수 있음
# 본인 컴퓨터 사양에 맞춰서 진행
model.fit_generator(
data_generator(dataset[:num_train],
args.batch_size,
input_shape,
anchors,
num_classes,
args.enhance_augment,
rescale_interval,
multi_anchor_assign=args.multi_anchor_assign),
steps_per_epoch=max(1, num_train // args.batch_size),
#validation_data=val_data_generator,
validation_data=data_generator(
dataset[num_train:],
args.batch_size,
input_shape,
anchors,
num_classes,
multi_anchor_assign=args.multi_anchor_assign),
validation_steps=max(1, num_val // args.batch_size),
epochs=args.total_epoch,
initial_epoch=epochs,
#verbose=1,
workers=1,
use_multiprocessing=False,
max_queue_size=10,
callbacks=callbacks)
# Finally store model
if args.model_pruning:
model = sparsity.strip_pruning(model)
model.save(os.path.join(log_dir, 'trained_final.h5'))
def main(args):
annotation_file = args.annotation_file
log_dir = os.path.join('logs', '000')
classes_path = args.classes_path
class_names = get_classes(classes_path)
num_classes = len(class_names)
anchors = get_anchors(args.anchors_path)
num_anchors = len(anchors)
# get freeze level according to CLI option
if args.weights_path:
freeze_level = 0
else:
freeze_level = 1
if args.freeze_level is not None:
freeze_level = args.freeze_level
# callbacks for training process
logging = TensorBoard(log_dir=log_dir,
histogram_freq=0,
write_graph=False,
write_grads=False,
write_images=False,
update_freq='batch')
checkpoint = ModelCheckpoint(os.path.join(
log_dir, 'ep{epoch:03d}-loss{loss:.3f}-val_loss{val_loss:.3f}.h5'),
monitor='val_loss',
mode='min',
verbose=1,
save_weights_only=False,
save_best_only=True,
period=1)
reduce_lr = ReduceLROnPlateau(monitor='val_loss',
factor=0.5,
mode='min',
patience=10,
verbose=1,
cooldown=0,
min_lr=1e-10)
early_stopping = EarlyStopping(monitor='val_loss',
min_delta=0,
patience=50,
verbose=1,
mode='min')
terminate_on_nan = TerminateOnNaN()
callbacks = [
logging, checkpoint, reduce_lr, early_stopping, terminate_on_nan
]
# get train&val dataset
dataset = get_dataset(annotation_file)
if args.val_annotation_file:
val_dataset = get_dataset(args.val_annotation_file)
num_train = len(dataset)
num_val = len(val_dataset)
dataset.extend(val_dataset)
else:
val_split = args.val_split
num_val = int(len(dataset) * val_split)
num_train = len(dataset) - num_val
# assign multiscale interval
if args.multiscale:
rescale_interval = args.rescale_interval
else:
rescale_interval = -1 #Doesn't rescale
# model input shape check
input_shape = args.model_image_size
assert (input_shape[0] % 32 == 0 and input_shape[1] % 32
== 0), 'model_image_size should be multiples of 32'
# get different model type & train&val data generator
if args.model_type.startswith(
'scaled_yolo4_') or args.model_type.startswith('yolo5_'):
# Scaled-YOLOv4 & YOLOv5 entrance, use yolo5 submodule but now still yolo3 data generator
# TODO: create new yolo5 data generator to apply YOLOv5 anchor assignment
get_train_model = get_yolo5_train_model
data_generator = yolo5_data_generator_wrapper
# tf.keras.Sequence style data generator
#train_data_generator = Yolo5DataGenerator(dataset[:num_train], args.batch_size, input_shape, anchors, num_classes, args.enhance_augment, rescale_interval, args.multi_anchor_assign)
#val_data_generator = Yolo5DataGenerator(dataset[num_train:], args.batch_size, input_shape, anchors, num_classes, multi_anchor_assign=args.multi_anchor_assign)
tiny_version = False
elif args.model_type.startswith('yolo3_') or args.model_type.startswith(
'yolo4_'):
#if num_anchors == 9:
# YOLOv3 & v4 entrance, use 9 anchors
get_train_model = get_yolo3_train_model
data_generator = yolo3_data_generator_wrapper
# tf.keras.Sequence style data generator
#train_data_generator = Yolo3DataGenerator(dataset[:num_train], args.batch_size, input_shape, anchors, num_classes, args.enhance_augment, rescale_interval, args.multi_anchor_assign)
#val_data_generator = Yolo3DataGenerator(dataset[num_train:], args.batch_size, input_shape, anchors, num_classes, multi_anchor_assign=args.multi_anchor_assign)
tiny_version = False
elif args.model_type.startswith(
'tiny_yolo3_') or args.model_type.startswith('tiny_yolo4_'):
#elif num_anchors == 6:
# Tiny YOLOv3 & v4 entrance, use 6 anchors
get_train_model = get_yolo3_train_model
data_generator = yolo3_data_generator_wrapper
# tf.keras.Sequence style data generator
#train_data_generator = Yolo3DataGenerator(dataset[:num_train], args.batch_size, input_shape, anchors, num_classes, args.enhance_augment, rescale_interval, args.multi_anchor_assign)
#val_data_generator = Yolo3DataGenerator(dataset[num_train:], args.batch_size, input_shape, anchors, num_classes, multi_anchor_assign=args.multi_anchor_assign)
tiny_version = True
elif args.model_type.startswith('yolo2_') or args.model_type.startswith(
'tiny_yolo2_'):
#elif num_anchors == 5:
# YOLOv2 & Tiny YOLOv2 use 5 anchors
get_train_model = get_yolo2_train_model
data_generator = yolo2_data_generator_wrapper
# tf.keras.Sequence style data generator
#train_data_generator = Yolo2DataGenerator(dataset[:num_train], args.batch_size, input_shape, anchors, num_classes, args.enhance_augment, rescale_interval)
#val_data_generator = Yolo2DataGenerator(dataset[num_train:], args.batch_size, input_shape, anchors, num_classes)
tiny_version = False
else:
raise ValueError('Unsupported model type')
# prepare online evaluation callback
if args.eval_online:
eval_callback = EvalCallBack(
args.model_type,
dataset[num_train:],
anchors,
class_names,
args.model_image_size,
args.model_pruning,
log_dir,
eval_epoch_interval=args.eval_epoch_interval,
save_eval_checkpoint=args.save_eval_checkpoint,
elim_grid_sense=args.elim_grid_sense)
callbacks.append(eval_callback)
# prepare train/val data shuffle callback
if args.data_shuffle:
shuffle_callback = DatasetShuffleCallBack(dataset)
callbacks.append(shuffle_callback)
# prepare model pruning config
pruning_end_step = np.ceil(1.0 * num_train / args.batch_size).astype(
np.int32) * args.total_epoch
if args.model_pruning:
pruning_callbacks = [
sparsity.UpdatePruningStep(),
sparsity.PruningSummaries(log_dir=log_dir, profile_batch=0)
]
callbacks = callbacks + pruning_callbacks
# prepare optimizer
optimizer = get_optimizer(args.optimizer,
args.learning_rate,
average_type=None,
decay_type=None)
# support multi-gpu training
if args.gpu_num >= 2:
# devices_list=["/gpu:0", "/gpu:1"]
devices_list = ["/gpu:{}".format(n) for n in range(args.gpu_num)]
strategy = tf.distribute.MirroredStrategy(devices=devices_list)
print('Number of devices: {}'.format(strategy.num_replicas_in_sync))
with strategy.scope():
# get multi-gpu train model
model = get_train_model(args.model_type,
anchors,
num_classes,
weights_path=args.weights_path,
freeze_level=freeze_level,
optimizer=optimizer,
label_smoothing=args.label_smoothing,
elim_grid_sense=args.elim_grid_sense,
model_pruning=args.model_pruning,
pruning_end_step=pruning_end_step)
else:
# get normal train model
model = get_train_model(args.model_type,
anchors,
num_classes,
weights_path=args.weights_path,
freeze_level=freeze_level,
optimizer=optimizer,
label_smoothing=args.label_smoothing,
elim_grid_sense=args.elim_grid_sense,
model_pruning=args.model_pruning,
pruning_end_step=pruning_end_step)
model.summary()
# Transfer training some epochs with frozen layers first if needed, to get a stable loss.
initial_epoch = args.init_epoch
epochs = initial_epoch + args.transfer_epoch
print("Transfer training stage")
print(
'Train on {} samples, val on {} samples, with batch size {}, input_shape {}.'
.format(num_train, num_val, args.batch_size, input_shape))
#model.fit_generator(train_data_generator,
model.fit_generator(
data_generator(dataset[:num_train],
args.batch_size,
input_shape,
anchors,
num_classes,
args.enhance_augment,
rescale_interval,
multi_anchor_assign=args.multi_anchor_assign),
steps_per_epoch=max(1, num_train // args.batch_size),
#validation_data=val_data_generator,
validation_data=data_generator(
dataset[num_train:],
args.batch_size,
input_shape,
anchors,
num_classes,
multi_anchor_assign=args.multi_anchor_assign),
validation_steps=max(1, num_val // args.batch_size),
epochs=epochs,
initial_epoch=initial_epoch,
#verbose=1,
workers=1,
use_multiprocessing=False,
max_queue_size=10,
callbacks=callbacks)
# Wait 2 seconds for next stage
time.sleep(2)
if args.decay_type or args.average_type:
# rebuild optimizer to apply learning rate decay or weights averager,
# only after unfreeze all layers
if args.decay_type:
callbacks.remove(reduce_lr)
if args.average_type == 'ema' or args.average_type == 'swa':
# weights averager need tensorflow-addons,
# which request TF 2.x and have version compatibility
import tensorflow_addons as tfa
callbacks.remove(checkpoint)
avg_checkpoint = tfa.callbacks.AverageModelCheckpoint(
filepath=os.path.join(
log_dir,
'ep{epoch:03d}-loss{loss:.3f}-val_loss{val_loss:.3f}.h5'),
update_weights=True,
monitor='val_loss',
mode='min',
verbose=1,
save_weights_only=False,
save_best_only=True,
period=1)
callbacks.append(avg_checkpoint)
steps_per_epoch = max(1, num_train // args.batch_size)
decay_steps = steps_per_epoch * (args.total_epoch - args.init_epoch -
args.transfer_epoch)
optimizer = get_optimizer(args.optimizer,
args.learning_rate,
average_type=args.average_type,
decay_type=args.decay_type,
decay_steps=decay_steps)
# Unfreeze the whole network for further tuning
# NOTE: more GPU memory is required after unfreezing the body
print("Unfreeze and continue training, to fine-tune.")
if args.gpu_num >= 2:
with strategy.scope():
for i in range(len(model.layers)):
model.layers[i].trainable = True
model.compile(optimizer=optimizer,
loss={
'yolo_loss': lambda y_true, y_pred: y_pred
}) # recompile to apply the change
else:
for i in range(len(model.layers)):
model.layers[i].trainable = True
model.compile(optimizer=optimizer,
loss={
'yolo_loss': lambda y_true, y_pred: y_pred
}) # recompile to apply the change
print(
'Train on {} samples, val on {} samples, with batch size {}, input_shape {}.'
.format(num_train, num_val, args.batch_size, input_shape))
#model.fit_generator(train_data_generator,
model.fit_generator(
data_generator(dataset[:num_train],
args.batch_size,
input_shape,
anchors,
num_classes,
args.enhance_augment,
rescale_interval,
multi_anchor_assign=args.multi_anchor_assign),
steps_per_epoch=max(1, num_train // args.batch_size),
#validation_data=val_data_generator,
validation_data=data_generator(
dataset[num_train:],
args.batch_size,
input_shape,
anchors,
num_classes,
multi_anchor_assign=args.multi_anchor_assign),
validation_steps=max(1, num_val // args.batch_size),
epochs=args.total_epoch,
initial_epoch=epochs,
#verbose=1,
workers=1,
use_multiprocessing=False,
max_queue_size=10,
callbacks=callbacks)
# Finally store model
if args.model_pruning:
model = sparsity.strip_pruning(model)
model.save(os.path.join(log_dir, 'trained_final.h5'))
def main(args):
annotation_file = args.annotation_file
classes_path = args.classes_path
class_names = get_classes(classes_path)
num_classes = len(class_names)
anchors = get_anchors(args.anchors_path)
num_anchors = len(anchors)
log_dir_path = args.log_directory
try:
log_dir = os.path.join('logs', log_dir_path)
except TypeError:
date_now = datetime.now()
log_dir_folder_name = f'{date_now.strftime("%Y_%m_%d_%H%M%S")}_{args.model_type}_TransferEp_{args.transfer_epoch}_TotalEP_{args.total_epoch}'
log_dir = os.path.realpath(os.path.join(
'logs',
log_dir_folder_name
))
# get freeze level according to CLI option
if args.weights_path:
freeze_level = 0
else:
freeze_level = 1
if args.freeze_level is not None:
freeze_level = args.freeze_level
# How many percentage of layers to unfreeze in fine tuning
unfreeze_level = args.unfreeze_level
# callbacks for training process
logging = TensorBoard(log_dir=log_dir, histogram_freq=0, write_graph=False, write_grads=False, write_images=False,
update_freq='batch')
checkpoint = ModelCheckpoint(
filepath=log_dir + os.sep + 'ep{epoch:03d}-loss{loss:.3f}-val_loss{val_loss:.3f}.h5',
monitor='val_loss',
mode='min',
verbose=1,
save_weights_only=False,
save_best_only=True,
period=1
)
reduce_lr = ReduceLROnPlateau(
monitor='val_loss',
factor=0.5, mode='min',
patience=10,
verbose=1,
cooldown=0,
min_lr=1e-10
)
early_stopping = EarlyStopping(monitor='val_loss', min_delta=0, patience=50, verbose=1, mode='min')
terminate_on_nan = TerminateOnNaN()
callbacks = [logging, checkpoint, reduce_lr, early_stopping, terminate_on_nan]
# get train&val dataset
dataset = get_dataset(annotation_file)
if args.val_annotation_file:
val_dataset = get_dataset(args.val_annotation_file)
num_train = len(dataset)
num_val = len(val_dataset)
dataset.extend(val_dataset)
else:
val_split = args.val_split
num_val = int(len(dataset) * val_split)
num_train = len(dataset) - num_val
# assign multiscale interval
if args.multiscale:
rescale_interval = args.rescale_interval
else:
rescale_interval = -1 # Doesn't rescale
# model input shape check
input_shape = args.model_image_size
assert (input_shape[0] % 32 == 0 and input_shape[1] % 32 == 0), 'model_image_size should be multiples of 32'
# get different model type & train&val data generator
if num_anchors == 9:
# YOLOv3 use 9 anchors
get_train_model = get_yolo3_train_model
data_generator = yolo3_data_generator_wrapper
# tf.keras.Sequence style data generator
# train_data_generator = Yolo3DataGenerator(dataset[:num_train], args.batch_size, input_shape, anchors, num_classes, args.enhance_augment, rescale_interval, args.multi_anchor_assign)
# val_data_generator = Yolo3DataGenerator(dataset[num_train:], args.batch_size, input_shape, anchors, num_classes, multi_anchor_assign=args.multi_anchor_assign)
tiny_version = False
elif num_anchors == 6:
# Tiny YOLOv3 use 6 anchors
get_train_model = get_yolo3_train_model
data_generator = yolo3_data_generator_wrapper
# tf.keras.Sequence style data generator
# train_data_generator = Yolo3DataGenerator(dataset[:num_train], args.batch_size, input_shape, anchors, num_classes, args.enhance_augment, rescale_interval, args.multi_anchor_assign)
# val_data_generator = Yolo3DataGenerator(dataset[num_train:], args.batch_size, input_shape, anchors, num_classes, multi_anchor_assign=args.multi_anchor_assign)
tiny_version = True
elif num_anchors == 5:
# YOLOv2 use 5 anchors
get_train_model = get_yolo2_train_model
data_generator = yolo2_data_generator_wrapper
# tf.keras.Sequence style data generator
# train_data_generator = Yolo2DataGenerator(dataset[:num_train], args.batch_size, input_shape, anchors, num_classes, args.enhance_augment, rescale_interval)
# val_data_generator = Yolo2DataGenerator(dataset[num_train:], args.batch_size, input_shape, anchors, num_classes)
tiny_version = False
else:
raise ValueError('Unsupported anchors number')
# prepare online evaluation callback
if args.eval_online:
eval_callback = EvalCallBack(
model_type=args.model_type,
annotation_lines=dataset[num_train:],
anchors=anchors,
class_names=class_names,
model_image_size=args.model_image_size,
model_pruning=args.model_pruning,
log_dir=log_dir,
eval_epoch_interval=args.eval_epoch_interval,
save_eval_checkpoint=args.save_eval_checkpoint,
elim_grid_sense=args.elim_grid_sense
)
callbacks.append(eval_callback)
# prepare train/val data shuffle callback
if args.data_shuffle:
shuffle_callback = DatasetShuffleCallBack(dataset)
callbacks.append(shuffle_callback)
# prepare model pruning config
pruning_end_step = np.ceil(1.0 * num_train / args.batch_size).astype(np.int32) * args.total_epoch
if args.model_pruning:
pruning_callbacks = [sparsity.UpdatePruningStep(), sparsity.PruningSummaries(log_dir=log_dir, profile_batch=0)]
callbacks = callbacks + pruning_callbacks
# prepare optimizer
optimizer = get_optimizer(args.optimizer, args.learning_rate, decay_type=None)
# support multi-gpu training
if args.gpu_num >= 2:
# devices_list=["/gpu:0", "/gpu:1"]
devices_list = ["/gpu:{}".format(n) for n in range(args.gpu_num)]
strategy = tf.distribute.MirroredStrategy(devices=devices_list)
print('Number of devices: {}'.format(strategy.num_replicas_in_sync))
with strategy.scope():
# get multi-gpu train model
model = get_train_model(
model_type=args.model_type,
anchors=anchors,
num_classes=num_classes,
weights_path=args.weights_path,
freeze_level=freeze_level,
optimizer=optimizer,
label_smoothing=args.label_smoothing,
elim_grid_sense=args.elim_grid_sense,
model_pruning=args.model_pruning,
pruning_end_step=pruning_end_step
)
else:
# get normal train model
model = get_train_model(
model_type=args.model_type,
anchors=anchors,
num_classes=num_classes,
weights_path=args.weights_path,
freeze_level=freeze_level,
optimizer=optimizer,
label_smoothing=args.label_smoothing,
elim_grid_sense=args.elim_grid_sense,
model_pruning=args.model_pruning,
pruning_end_step=pruning_end_step
)
if args.show_history:
model.summary()
layers_count = len(model.layers)
print(f'Total layers: {layers_count}')
# Transfer training some epochs with frozen layers first if needed, to get a stable loss.
initial_epoch = args.init_epoch
epochs = initial_epoch + args.transfer_epoch
print("Transfer training stage")
print('Train on {} samples, val on {} samples, with batch size {}, input_shape {}.'.format(num_train, num_val,
args.batch_size,
input_shape))
# model.fit_generator(train_data_generator,
"""
Transfer training steps, train with freeze layers
"""
model.fit(
data_generator(
annotation_lines=dataset[:num_train],
batch_size=args.batch_size,
input_shape=input_shape,
anchors=anchors,
num_classes=num_classes,
enhance_augment=args.enhance_augment,
rescale_interval=rescale_interval,
multi_anchor_assign=args.multi_anchor_assign
),
steps_per_epoch=max(1, num_train // args.batch_size),
# validation_data=val_data_generator,
validation_data=data_generator(
annotation_lines=dataset[num_train:],
batch_size=args.batch_size,
input_shape=input_shape,
anchors=anchors,
num_classes=num_classes,
multi_anchor_assign=args.multi_anchor_assign
),
validation_steps=max(1, num_val // args.batch_size),
epochs=epochs,
initial_epoch=initial_epoch,
# verbose=1,
workers=1,
use_multiprocessing=False,
max_queue_size=10,
callbacks=callbacks
)
# Wait 2 seconds for next stage
time.sleep(2)
if args.decay_type:
# rebuild optimizer to apply learning rate decay, only after
# unfreeze all layers
callbacks.remove(reduce_lr)
steps_per_epoch = max(1, num_train // args.batch_size)
decay_steps = steps_per_epoch * (args.total_epoch - args.init_epoch - args.transfer_epoch)
optimizer = get_optimizer(args.optimizer, args.learning_rate, decay_type=args.decay_type,
decay_steps=decay_steps)
# Unfreeze the whole network for further tuning
# NOTE: more GPU memory is required after unfreezing the body
fine_tune_layers = int(layers_count * unfreeze_level)
print(f"Unfreeze {unfreeze_level * 100}% of layers and continue training, to fine-tune.")
print(f"Unfroze {fine_tune_layers} layers of {layers_count}")
if args.gpu_num >= 2:
with strategy.scope():
for i in range(layers_count - fine_tune_layers, layers_count):
model.layers[i].trainable = True
model.compile(optimizer=optimizer,
loss={'yolo_loss': lambda y_true, y_pred: y_pred}) # recompile to apply the change
else:
for i in range(layers_count - fine_tune_layers, layers_count):
model.layers[i].trainable = True
model.compile(optimizer=optimizer,
loss={'yolo_loss': lambda y_true, y_pred: y_pred}) # recompile to apply the change
print('Train on {} samples, val on {} samples, with batch size {}, input_shape {}.'.format(num_train, num_val,
args.batch_size,
input_shape))
"""
Fine-tuning steps, more memory will be used. LR (Learning Rate) will be decayed
"""
# model.fit_generator(train_data_generator,
model.fit(
# The YOLO data augmentation generator tool
data_generator(
annotation_lines=dataset[:num_train],
batch_size=args.batch_size,
input_shape=input_shape,
anchors=anchors,
num_classes=num_classes,
enhance_augment=args.enhance_augment,
rescale_interval=rescale_interval,
multi_anchor_assign=args.multi_anchor_assign
),
steps_per_epoch=max(1, num_train // args.batch_size),
# validation_data=val_data_generator,
# Validation generator
validation_data=data_generator(
annotation_lines=dataset[num_train:],
batch_size=args.batch_size,
input_shape=input_shape,
anchors=anchors,
num_classes=num_classes,
multi_anchor_assign=args.multi_anchor_assign
),
validation_steps=max(1, num_val // args.batch_size),
epochs=args.total_epoch,
initial_epoch=epochs,
# verbose=1,
workers=1,
use_multiprocessing=False,
max_queue_size=10,
callbacks=callbacks
)
# Finally store model
if args.model_pruning:
model = sparsity.strip_pruning(model)
model.save(os.path.join(log_dir, 'trained_final.h5'))
def main(args):
log_dir = 'logs/000/'
# get class info, add background class to match model & GT
class_names = get_classes(args.classes_path)
assert len(class_names) < 254, 'PNG image label only support less than 254 classes.'
class_names = ['background'] + class_names
num_classes = len(class_names)
# callbacks for training process
monitor = 'Jaccard'
tensorboard = TensorBoard(log_dir=log_dir, histogram_freq=0, write_graph=False, write_grads=False, write_images=False, update_freq='batch')
checkpoint = ModelCheckpoint(os.path.join(log_dir, 'ep{epoch:03d}-loss{loss:.3f}-Jaccard{Jaccard:.3f}-val_loss{val_loss:.3f}-val_Jaccard{val_Jaccard:.3f}.h5'),
monitor='val_{}'.format(monitor),
mode='max',
verbose=1,
save_weights_only=False,
save_best_only=True,
period=1)
reduce_lr = ReduceLROnPlateau(monitor='val_{}'.format(monitor), factor=0.5, mode='max',
patience=5, verbose=1, cooldown=0, min_lr=1e-6)
early_stopping = EarlyStopping(monitor='val_{}'.format(monitor), min_delta=0, patience=100, verbose=1, mode='max')
terminate_on_nan = TerminateOnNaN()
callbacks=[tensorboard, checkpoint, reduce_lr, early_stopping, terminate_on_nan]
# get train&val dataset
dataset = get_data_list(args.dataset_file)
if args.val_dataset_file:
val_dataset = get_data_list(args.val_dataset_file)
num_train = len(dataset)
num_val = len(val_dataset)
dataset.extend(val_dataset)
else:
val_split = args.val_split
num_val = int(len(dataset)*val_split)
num_train = len(dataset) - num_val
# prepare train&val data generator
train_generator = SegmentationGenerator(args.dataset_path, dataset[:num_train],
args.batch_size,
num_classes,
resize_shape=args.model_input_shape[::-1],
crop_shape=None,
weighted_type=args.weighted_type,
augment=True,
do_ahisteq=False)
valid_generator = SegmentationGenerator(args.dataset_path, dataset[num_train:],
args.batch_size,
num_classes,
resize_shape=args.model_input_shape[::-1],
crop_shape=None,
weighted_type=args.weighted_type,
augment=False,
do_ahisteq=False)
# prepare online evaluation callback
if args.eval_online:
eval_callback = EvalCallBack(args.dataset_path, dataset[num_train:], class_names, args.model_input_shape, args.model_pruning, log_dir, eval_epoch_interval=args.eval_epoch_interval, save_eval_checkpoint=args.save_eval_checkpoint)
callbacks.append(eval_callback)
# prepare optimizer
#optimizer = Adam(lr=7e-4, epsilon=1e-8, decay=1e-6)
optimizer = get_optimizer(args.optimizer, args.learning_rate, decay_type=None)
# prepare loss according to loss type & weigted type
if args.weighted_type == 'balanced':
classes_weights_path = os.path.join(args.dataset_path, 'classes_weights.txt')
if os.path.isfile(classes_weights_path):
weights = load_class_weights(classes_weights_path)
else:
weights = calculate_weigths_labels(train_generator, num_classes, save_path=args.dataset_path)
losses = WeightedSparseCategoricalCrossEntropy(weights)
sample_weight_mode = None
elif args.weighted_type == 'adaptive':
losses = sparse_crossentropy
sample_weight_mode = 'temporal'
elif args.weighted_type == None:
losses = sparse_crossentropy
sample_weight_mode = None
else:
raise ValueError('invalid weighted_type {}'.format(args.weighted_type))
if args.loss == 'focal':
warnings.warn("Focal loss doesn't support weighted class balance, will ignore related config")
losses = softmax_focal_loss
sample_weight_mode = None
elif args.loss == 'crossentropy':
# using crossentropy will keep the weigted type setting
pass
else:
raise ValueError('invalid loss type {}'.format(args.loss))
# prepare metric
#metrics = {'pred_mask' : [Jaccard, sparse_accuracy_ignoring_last_label]}
metrics = {'pred_mask' : Jaccard}
# support multi-gpu training
if args.gpu_num >= 2:
# devices_list=["/gpu:0", "/gpu:1"]
devices_list=["/gpu:{}".format(n) for n in range(args.gpu_num)]
strategy = tf.distribute.MirroredStrategy(devices=devices_list)
print ('Number of devices: {}'.format(strategy.num_replicas_in_sync))
with strategy.scope():
# get multi-gpu train model
model = get_deeplabv3p_model(args.model_type, num_classes, args.model_input_shape, args.output_stride, args.freeze_level, weights_path=args.weights_path)
# compile model
model.compile(optimizer=optimizer, sample_weight_mode=sample_weight_mode,
loss = losses, metrics = metrics)
else:
# get normal train model
model = get_deeplabv3p_model(args.model_type, num_classes, args.model_input_shape, args.output_stride, args.freeze_level, weights_path=args.weights_path)
# compile model
model.compile(optimizer=optimizer, sample_weight_mode=sample_weight_mode,
loss = losses, metrics = metrics)
model.summary()
# Transfer training some epochs with frozen layers first if needed, to get a stable loss.
initial_epoch = args.init_epoch
epochs = initial_epoch + args.transfer_epoch
print("Transfer training stage")
print('Train on {} samples, val on {} samples, with batch size {}, input_shape {}.'.format(num_train, num_val, args.batch_size, args.model_input_shape))
model.fit_generator(generator=train_generator,
steps_per_epoch=len(train_generator),
validation_data=valid_generator,
validation_steps=len(valid_generator),
epochs=epochs,
initial_epoch=initial_epoch,
verbose=1,
workers=1,
use_multiprocessing=False,
max_queue_size=10,
callbacks = callbacks)
# Wait 2 seconds for next stage
time.sleep(2)
if args.decay_type:
# rebuild optimizer to apply learning rate decay, only after
# unfreeze all layers
callbacks.remove(reduce_lr)
steps_per_epoch = max(1, len(train_generator))
decay_steps = steps_per_epoch * (args.total_epoch - args.init_epoch - args.transfer_epoch)
optimizer = get_optimizer(args.optimizer, args.learning_rate, decay_type=args.decay_type, decay_steps=decay_steps)
# Unfreeze the whole network for further tuning
# NOTE: more GPU memory is required after unfreezing the body
print("Unfreeze and continue training, to fine-tune.")
if args.gpu_num >= 2:
with strategy.scope():
for i in range(len(model.layers)):
model.layers[i].trainable = True
model.compile(optimizer=optimizer, sample_weight_mode=sample_weight_mode,
loss = losses, metrics = metrics) # recompile to apply the change
else:
for i in range(len(model.layers)):
model.layers[i].trainable = True
model.compile(optimizer=optimizer, sample_weight_mode=sample_weight_mode,
loss = losses, metrics = metrics) # recompile to apply the change
print('Train on {} samples, val on {} samples, with batch size {}, input_shape {}.'.format(num_train, num_val, args.batch_size, args.model_input_shape))
model.fit_generator(generator=train_generator,
steps_per_epoch=len(train_generator),
validation_data=valid_generator,
validation_steps=len(valid_generator),
epochs=args.total_epoch,
initial_epoch=epochs,
verbose=1,
workers=1,
use_multiprocessing=False,
max_queue_size=10,
callbacks = callbacks)
# Finally store model
model.save(os.path.join(log_dir, 'trained_final.h5'))