def main():
args = parse_args()
_, log_file_exist = initialize_logging(
logging_dir_path=args.save_dir,
logging_file_name=args.logging_file_name,
script_args=args,
log_packages=args.log_packages,
log_pip_packages=args.log_pip_packages)
ctx, batch_size = prepare_mx_context(num_gpus=args.num_gpus,
batch_size=args.batch_size)
net = prepare_model(model_name=args.model,
use_pretrained=args.use_pretrained,
pretrained_model_file_path=args.resume.strip(),
dtype=args.dtype,
classes=args.num_classes,
in_channels=args.in_channels,
do_hybridize=(not args.calc_flops),
ctx=ctx)
assert (hasattr(net, "in_size"))
input_image_size = net.in_size
ds_metainfo = get_dataset_metainfo(dataset_name=args.dataset)
val_data = get_val_data_source(dataset_metainfo=ds_metainfo,
dataset_dir=args.data_dir,
batch_size=batch_size,
num_workers=args.num_workers,
input_image_size=input_image_size,
resize_inv_factor=args.resize_inv_factor)
batch_fn = get_batch_fn(use_imgrec=ds_metainfo.use_imgrec)
assert (args.use_pretrained or args.resume.strip() or args.calc_flops_only)
test(
net=net,
val_data=val_data,
batch_fn=batch_fn,
data_source_needs_reset=ds_metainfo.use_imgrec,
val_metric=get_composite_metric(ds_metainfo.val_metric_names),
dtype=args.dtype,
ctx=ctx,
input_image_size=input_image_size,
in_channels=args.in_channels,
# calc_weight_count=(not log_file_exist),
calc_weight_count=True,
calc_flops=args.calc_flops,
calc_flops_only=args.calc_flops_only,
extended_log=True)
def parse_args():
parser = argparse.ArgumentParser(
description="Evaluate a model for image classification (Gluon/CIFAR)",
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument(
"--dataset",
type=str,
default="CIFAR10",
help="dataset name. options are CIFAR10, CIFAR100, and SVHN")
args, _ = parser.parse_known_args()
dataset_metainfo = get_dataset_metainfo(dataset_name=args.dataset)
work_dir_path = os.path.join("..", "imgclsmob_data")
dataset_metainfo.add_dataset_parser_arguments(parser, work_dir_path)
add_eval_cls_parser_arguments(parser)
args = parser.parse_args()
return args
def main():
args = parse_args()
args.seed = init_rand(seed=args.seed)
_, log_file_exist = initialize_logging(
logging_dir_path=args.save_dir,
logging_file_name=args.logging_file_name,
script_args=args,
log_packages=args.log_packages,
log_pip_packages=args.log_pip_packages)
ctx, batch_size = prepare_mx_context(num_gpus=args.num_gpus,
batch_size=args.batch_size)
net = prepare_model(model_name=args.model,
use_pretrained=args.use_pretrained,
pretrained_model_file_path=args.resume.strip(),
dtype=args.dtype,
tune_layers=args.tune_layers,
classes=args.num_classes,
in_channels=args.in_channels,
do_hybridize=(not args.not_hybridize),
ctx=ctx)
assert (hasattr(net, "classes"))
num_classes = net.classes if hasattr(net, "classes") else 10
ds_metainfo = get_dataset_metainfo(dataset_name=args.dataset)
train_data = get_train_data_source(dataset_metainfo=ds_metainfo,
dataset_dir=args.data_dir,
batch_size=batch_size,
num_workers=args.num_workers)
val_data = get_val_data_source(dataset_metainfo=ds_metainfo,
dataset_dir=args.data_dir,
batch_size=batch_size,
num_workers=args.num_workers)
batch_fn = get_batch_fn(use_imgrec=ds_metainfo.use_imgrec)
num_training_samples = len(
train_data._dataset
) if not ds_metainfo.use_imgrec else ds_metainfo.num_training_samples
trainer, lr_scheduler = prepare_trainer(
net=net,
optimizer_name=args.optimizer_name,
wd=args.wd,
momentum=args.momentum,
lr_mode=args.lr_mode,
lr=args.lr,
lr_decay_period=args.lr_decay_period,
lr_decay_epoch=args.lr_decay_epoch,
lr_decay=args.lr_decay,
target_lr=args.target_lr,
poly_power=args.poly_power,
warmup_epochs=args.warmup_epochs,
warmup_lr=args.warmup_lr,
warmup_mode=args.warmup_mode,
batch_size=batch_size,
num_epochs=args.num_epochs,
num_training_samples=num_training_samples,
dtype=args.dtype,
gamma_wd_mult=args.gamma_wd_mult,
beta_wd_mult=args.beta_wd_mult,
bias_wd_mult=args.bias_wd_mult,
state_file_path=args.resume_state)
if args.save_dir and args.save_interval:
param_names = ds_metainfo.val_metric_capts + ds_metainfo.train_metric_capts + [
"Train.Loss", "LR"
]
lp_saver = TrainLogParamSaver(
checkpoint_file_name_prefix="{}_{}".format(args.dataset.lower(),
args.model),
last_checkpoint_file_name_suffix="last",
best_checkpoint_file_name_suffix=None,
last_checkpoint_dir_path=args.save_dir,
best_checkpoint_dir_path=None,
last_checkpoint_file_count=2,
best_checkpoint_file_count=2,
checkpoint_file_save_callback=save_params,
checkpoint_file_exts=(".params", ".states"),
save_interval=args.save_interval,
num_epochs=args.num_epochs,
param_names=param_names,
acc_ind=ds_metainfo.saver_acc_ind,
# bigger=[True],
# mask=None,
score_log_file_path=os.path.join(args.save_dir, "score.log"),
score_log_attempt_value=args.attempt,
best_map_log_file_path=os.path.join(args.save_dir, "best_map.log"))
else:
lp_saver = None
train_net(batch_size=batch_size,
num_epochs=args.num_epochs,
start_epoch1=args.start_epoch,
train_data=train_data,
val_data=val_data,
batch_fn=batch_fn,
data_source_needs_reset=ds_metainfo.use_imgrec,
dtype=args.dtype,
net=net,
trainer=trainer,
lr_scheduler=lr_scheduler,
lp_saver=lp_saver,
log_interval=args.log_interval,
mixup=args.mixup,
mixup_epoch_tail=args.mixup_epoch_tail,
label_smoothing=args.label_smoothing,
num_classes=num_classes,
grad_clip_value=args.grad_clip,
batch_size_scale=args.batch_size_scale,
val_metric=get_composite_metric(ds_metainfo.val_metric_names),
train_metric=get_composite_metric(
ds_metainfo.train_metric_names),
opt_metric_name=ds_metainfo.val_metric_names[
ds_metainfo.saver_acc_ind],
ctx=ctx)
def parse_args():
parser = argparse.ArgumentParser(
description='Train a model for image classification (Gluon/CIFAR)',
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--dataset',
type=str,
default="CIFAR10",
help='dataset name. options are CIFAR10 and CIFAR100')
args, _ = parser.parse_known_args()
dataset_metainfo = get_dataset_metainfo(dataset_name=args.dataset)
work_dir_path = os.path.join("..", "imgclsmob_data")
dataset_metainfo.add_dataset_parser_arguments(parser, work_dir_path)
parser.add_argument(
'--model',
type=str,
required=True,
help='type of model to use. see model_provider for options.')
parser.add_argument('--use-pretrained',
action='store_true',
help='enable using pretrained model from gluon.')
parser.add_argument('--dtype',
type=str,
default='float32',
help='data type for training')
parser.add_argument('--not-hybridize',
action='store_true',
help='do not hybridize model')
parser.add_argument(
'--resume',
type=str,
default='',
help='resume from previously saved parameters if not None')
parser.add_argument(
'--resume-state',
type=str,
default='',
help='resume from previously saved optimizer state if not None')
parser.add_argument('--num-gpus',
type=int,
default=0,
help='number of gpus to use.')
parser.add_argument('-j',
'--num-data-workers',
dest='num_workers',
default=4,
type=int,
help='number of preprocessing workers')
parser.add_argument('--batch-size',
type=int,
default=128,
help='training batch size per device (CPU/GPU).')
parser.add_argument('--batch-size-scale',
type=int,
default=1,
help='manual batch-size increasing factor.')
parser.add_argument('--num-epochs',
type=int,
default=200,
help='number of training epochs.')
parser.add_argument(
'--start-epoch',
type=int,
default=1,
help='starting epoch for resuming, default is 1 for new training')
parser.add_argument('--attempt',
type=int,
default=1,
help='current number of training')
parser.add_argument('--optimizer-name',
type=str,
default='nag',
help='optimizer name')
parser.add_argument('--lr', type=float, default=0.1, help='learning rate')
parser.add_argument(
'--lr-mode',
type=str,
default='cosine',
help='learning rate scheduler mode. options are step, poly and cosine')
parser.add_argument('--lr-decay',
type=float,
default=0.1,
help='decay rate of learning rate')
parser.add_argument(
'--lr-decay-period',
type=int,
default=0,
help=
'interval for periodic learning rate decays. default is 0 to disable.')
parser.add_argument('--lr-decay-epoch',
type=str,
default='40,60',
help='epoches at which learning rate decays')
parser.add_argument('--target-lr',
type=float,
default=1e-8,
help='ending learning rate')
parser.add_argument('--poly-power',
type=float,
default=2,
help='power value for poly LR scheduler')
parser.add_argument('--warmup-epochs',
type=int,
default=0,
help='number of warmup epochs.')
parser.add_argument('--warmup-lr',
type=float,
default=1e-8,
help='starting warmup learning rate')
parser.add_argument(
'--warmup-mode',
type=str,
default='linear',
help=
'learning rate scheduler warmup mode. options are linear, poly and constant'
)
parser.add_argument('--momentum',
type=float,
default=0.9,
help='momentum value for optimizer')
parser.add_argument('--wd',
type=float,
default=0.0001,
help='weight decay rate')
parser.add_argument('--gamma-wd-mult',
type=float,
default=1.0,
help='weight decay multiplier for batchnorm gamma')
parser.add_argument('--beta-wd-mult',
type=float,
default=1.0,
help='weight decay multiplier for batchnorm beta')
parser.add_argument('--bias-wd-mult',
type=float,
default=1.0,
help='weight decay multiplier for bias')
parser.add_argument('--grad-clip',
type=float,
default=None,
help='max_norm for gradient clipping')
parser.add_argument('--label-smoothing',
action='store_true',
help='use label smoothing')
parser.add_argument('--mixup',
action='store_true',
help='use mixup strategy')
parser.add_argument(
'--mixup-epoch-tail',
type=int,
default=20,
help='number of epochs without mixup at the end of training')
parser.add_argument('--log-interval',
type=int,
default=200,
help='number of batches to wait before logging.')
parser.add_argument(
'--save-interval',
type=int,
default=4,
help='saving parameters epoch interval, best model will always be saved'
)
parser.add_argument('--save-dir',
type=str,
default='',
help='directory of saved models and log-files')
parser.add_argument('--logging-file-name',
type=str,
default='train.log',
help='filename of training log')
parser.add_argument('--seed',
type=int,
default=-1,
help='Random seed to be fixed')
parser.add_argument('--log-packages',
type=str,
default='mxnet',
help='list of python packages for logging')
parser.add_argument('--log-pip-packages',
type=str,
default='mxnet-cu100',
help='list of pip packages for logging')
parser.add_argument('--tune-layers',
type=str,
default='',
help='Regexp for selecting layers for fine tuning')
args = parser.parse_args()
return args
def parse_args():
parser = argparse.ArgumentParser(
description="Train a model for image classification (Gluon/ImageNet1K)",
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument(
"--dataset",
type=str,
default="ImageNet1K_rec",
help="dataset name. options are ImageNet1K and ImageNet1K_rec")
args, _ = parser.parse_known_args()
dataset_metainfo = get_dataset_metainfo(dataset_name=args.dataset)
work_dir_path = os.path.join("..", "imgclsmob_data")
dataset_metainfo.add_dataset_parser_arguments(parser, work_dir_path)
parser.add_argument(
"--model",
type=str,
required=True,
help="type of model to use. see model_provider for options")
parser.add_argument(
"--use-pretrained",
action="store_true",
help="enable using pretrained model from gluon")
parser.add_argument(
"--dtype",
type=str,
default="float32",
help="data type for training")
parser.add_argument(
"--resume",
type=str,
default="",
help="resume from previously saved parameters if not None")
parser.add_argument(
"--resume-state",
type=str,
default="",
help="resume from previously saved optimizer state if not None")
parser.add_argument(
"--num-gpus",
type=int,
default=0,
help="number of gpus to use")
parser.add_argument(
"-j",
"--num-data-workers",
dest="num_workers",
default=4,
type=int,
help="number of preprocessing workers")
parser.add_argument(
"--batch-size",
type=int,
default=512,
help="training batch size per device (CPU/GPU)")
parser.add_argument(
"--batch-size-scale",
type=int,
default=1,
help="manual batch-size increasing factor")
parser.add_argument(
"--num-epochs",
type=int,
default=120,
help="number of training epochs")
parser.add_argument(
"--start-epoch",
type=int,
default=1,
help="starting epoch for resuming, default is 1 for new training")
parser.add_argument(
"--attempt",
type=int,
default=1,
help="current attempt number for training")
parser.add_argument(
"--optimizer-name",
type=str,
default="nag",
help="optimizer name")
parser.add_argument(
"--lr",
type=float,
default=0.1,
help="learning rate")
parser.add_argument(
"--lr-mode",
type=str,
default="cosine",
help="learning rate scheduler mode. options are step, poly and cosine")
parser.add_argument(
"--lr-decay",
type=float,
default=0.1,
help="decay rate of learning rate")
parser.add_argument(
"--lr-decay-period",
type=int,
default=0,
help="interval for periodic learning rate decays. default is 0 to disable")
parser.add_argument(
"--lr-decay-epoch",
type=str,
default="40,60",
help="epoches at which learning rate decays")
parser.add_argument(
"--target-lr",
type=float,
default=1e-8,
help="ending learning rate")
parser.add_argument(
"--poly-power",
type=float,
default=2,
help="power value for poly LR scheduler")
parser.add_argument(
"--warmup-epochs",
type=int,
default=0,
help="number of warmup epochs")
parser.add_argument(
"--warmup-lr",
type=float,
default=1e-8,
help="starting warmup learning rate")
parser.add_argument(
"--warmup-mode",
type=str,
default="linear",
help="learning rate scheduler warmup mode. options are linear, poly and constant")
parser.add_argument(
"--momentum",
type=float,
default=0.9,
help="momentum value for optimizer")
parser.add_argument(
"--wd",
type=float,
default=0.0001,
help="weight decay rate")
parser.add_argument(
"--gamma-wd-mult",
type=float,
default=1.0,
help="weight decay multiplier for batchnorm gamma")
parser.add_argument(
"--beta-wd-mult",
type=float,
default=1.0,
help="weight decay multiplier for batchnorm beta")
parser.add_argument(
"--bias-wd-mult",
type=float,
default=1.0,
help="weight decay multiplier for bias")
parser.add_argument(
"--grad-clip",
type=float,
default=None,
help="max_norm for gradient clipping")
parser.add_argument(
"--label-smoothing",
action="store_true",
help="use label smoothing")
parser.add_argument(
"--mixup",
action="store_true",
help="use mixup strategy")
parser.add_argument(
"--mixup-epoch-tail",
type=int,
default=20,
help="number of epochs without mixup at the end of training")
parser.add_argument(
"--log-interval",
type=int,
default=50,
help="number of batches to wait before logging")
parser.add_argument(
"--save-interval",
type=int,
default=4,
help="saving parameters epoch interval, best model will always be saved")
parser.add_argument(
"--save-dir",
type=str,
default="",
help="directory of saved models and log-files")
parser.add_argument(
"--logging-file-name",
type=str,
default="train.log",
help="filename of training log")
parser.add_argument(
"--seed",
type=int,
default=-1,
help="random seed to be fixed")
parser.add_argument(
"--log-packages",
type=str,
default="mxnet",
help="list of python packages for logging")
parser.add_argument(
"--log-pip-packages",
type=str,
default="mxnet-cu100",
help="list of pip packages for logging")
parser.add_argument(
"--tune-layers",
type=str,
default="",
help="regexp for selecting layers for fine tuning")
args = parser.parse_args()
return args