def main():
if args.distributed:
init_distributed()
model = se_resnet50(num_classes=1000)
optimizer = optim.SGD(lr=0.6 / 1024 * args.batch_size,
momentum=0.9, weight_decay=1e-4)
scheduler = lr_scheduler.MultiStepLR([50, 70])
train_loader, test_loader = imagenet_loaders(args.root, args.batch_size, distributed=args.distributed,
num_train_samples=args.batch_size * 10 if args.debug else None,
num_test_samples=args.batch_size * 10 if args.debug else None)
c = [callbacks.AccuracyCallback(), callbacks.AccuracyCallback(k=5),
callbacks.LossCallback(),
callbacks.WeightSave('.'),
reporters.TensorboardReporter('.'),
reporters.TQDMReporter(range(args.epochs))]
with SupervisedTrainer(model, optimizer, F.cross_entropy,
callbacks=c,
scheduler=scheduler,
) as trainer:
for _ in c[-1]:
trainer.train(train_loader)
trainer.test(test_loader)
def main():
model = {
"resnet20": resnet20,
"wrn28_10": wrn28_10
}[args.model](num_classes=10)
weight_decay = {"resnet20": 1e-4, "wrn28_10": 5e-4}[args.model]
lr_decay = {"resnet20": 0.1, "wrn28_10": 0.2}[args.model]
train_loader, test_loader = cifar10_loaders(args.batch_size)
optimizer = optim.SGD(lr=1e-1, momentum=0.9, weight_decay=weight_decay)
scheduler = lr_scheduler.MultiStepLR([100, 150], gamma=lr_decay)
tq = reporters.TQDMReporter(range(args.epochs), verb=True)
c = [
callbacks.AccuracyCallback(),
callbacks.LossCallback(),
reporters.IOReporter("."),
reporters.TensorboardReporter("."),
callbacks.WeightSave("."), tq
]
with trainers.SupervisedTrainer(model,
optimizer,
F.cross_entropy,
callbacks=c,
scheduler=scheduler) as trainer:
for _ in tq:
trainer.train(train_loader)
trainer.test(test_loader)
def get_components(cfg):
labeled_loader, unlabeled_loader, val_loader, test_loader = get_dataloader(cfg.data.name,
cfg.data.labeled_size,
cfg.data.unlabeled_size,
cfg.data.val_size,
cfg.data.batch_size,
cfg.data.random_state,
download=cfg.data.download,
pilaugment=cfg.data.get('pilaugment',
False)
)
model = wrn28_2(num_classes=6 if cfg.data.name == "animal" else 10)
optimizer = {'adam': optim.Adam(lr=cfg.optim.lr),
'sgd': optim.SGD(lr=cfg.optim.lr, momentum=0.9)}[cfg.optim.name]
scheduler = {'adam': None,
'sgd': lr_scheduler.CosineAnnealingWithWarmup(cfg.optim.epochs,
4, cfg.optim.epochs // 100)}[cfg.optim.name]
ema_model = partial(EMAModel, ema_rate=cfg.model.ema_rate, weight_decay=cfg.optim.wd * cfg.optim.lr)
num_classes = {"animal": 6, "cifar100": 100, "tinyimagenet": 200}.get(cfg.data.name, 10)
tq = reporters.TQDMReporter(range(cfg.optim.epochs))
_callbacks = [callbacks.AccuracyCallback(),
callbacks.LossCallback(),
reporters.IOReporter("."),
reporters.TensorboardReporter("."), tq]
return PackedLoader(labeled_loader, unlabeled_loader), val_loader, test_loader, model, optimizer, \
scheduler, ema_model, num_classes, tq, _callbacks
def train_and_eval(cfg):
train_loader, val_loader, test_loader, num_classes = get_dataloader(
cfg.data.name, cfg.data.val_size, cfg.data.batch_size,
cfg.data.download, cfg.augment, False)
model = get_model(cfg.model.name, num_classes)
optimizer = optim.SGD(cfg.optim.model.lr,
momentum=0.9,
weight_decay=cfg.optim.model.weight_decay)
scheduler = lr_scheduler.MultiStepLR(cfg.optim.model.steps)
tq = reporters.TQDMReporter(range(cfg.optim.epochs), verb=cfg.verb)
callback = [
callbacks.AccuracyCallback(),
callbacks.LossCallback(),
reporters.TensorboardReporter("."),
reporters.IOReporter("."), tq
]
with trainers.SupervisedTrainer(model,
optimizer,
F.cross_entropy,
callbacks=callback,
scheduler=scheduler) as trainer:
for ep in tq:
trainer.train(train_loader)
trainer.test(val_loader, 'val')
trainer.test(test_loader)
def main():
train_loader, test_loader = imagenet_loaders(
args.root,
args.batch_size,
num_train_samples=args.batch_size * args.train_max_iter,
num_test_samples=args.batch_size * args.test_max_iter)
pretrained_model = resnet50(pretrained=True)
for p in pretrained_model.parameters():
p.requires_grad = False
pretrained_model.eval()
generator = ResNetGenerator(3, 3, args.num_filters)
generator.cuda()
optimizer = optim.Adam(lr=args.lr, betas=(args.beta1, 0.999))
trainer = Trainer({
"generator": generator,
"classifier": pretrained_model
},
optimizer,
reporter.TensorboardReporter([
adv_accuracy, fooling_rate,
callbacks.AccuracyCallback(),
callbacks.LossCallback()
],
save_dir="results"),
noise=torch.randn(3, 224,
224).expand(args.batch_size, -1, -1,
-1))
for ep in range(args.epochs):
trainer.train(train_loader)
trainer.test(test_loader)
def main():
model = se_resnet50(num_classes=1000)
optimizer = optim.SGD(lr=0.6 / 1024 * args.batch_size, momentum=0.9, weight_decay=1e-4)
scheduler = lr_scheduler.MultiStepLR([50, 70])
c = [callbacks.AccuracyCallback(), callbacks.LossCallback()]
r = reporters.TQDMReporter(range(args.epochs), callbacks=c)
tb = reporters.TensorboardReporter(c)
rep = callbacks.CallbackList(r, tb, callbacks.WeightSave("checkpoints"))
if args.distributed:
# DistributedSupervisedTrainer sets up torch.distributed
if args.local_rank == 0:
print("\nuse DistributedDataParallel")
trainer = DistributedSupervisedTrainer(model, optimizer, F.cross_entropy, callbacks=rep, scheduler=scheduler,
init_method=args.init_method, backend=args.backend)
else:
multi_gpus = torch.cuda.device_count() > 1
if multi_gpus:
print("\nuse DataParallel")
trainer = SupervisedTrainer(model, optimizer, F.cross_entropy, callbacks=rep,
scheduler=scheduler, data_parallel=multi_gpus)
# if distributed, need to setup loaders after DistributedSupervisedTrainer
train_loader, test_loader = imagenet_loaders(args.root, args.batch_size, distributed=args.distributed,
num_train_samples=args.batch_size * 10 if args.debug else None,
num_test_samples=args.batch_size * 10 if args.debug else None)
for _ in r:
trainer.train(train_loader)
trainer.test(test_loader)
def main():
if is_distributed():
init_distributed()
model = se_resnet50(num_classes=1000)
optimizer = optim.SGD(lr=0.6 / 1024 * args.batch_size,
momentum=0.9,
weight_decay=1e-4)
scheduler = lr_scheduler.MultiStepLR([50, 70])
train_loader, test_loader = DATASET_REGISTRY("imagenet")(args.batch_size)
c = [
callbacks.AccuracyCallback(),
callbacks.AccuracyCallback(k=5),
callbacks.LossCallback(),
callbacks.WeightSave("."),
reporters.TensorboardReporter("."),
reporters.TQDMReporter(range(args.epochs)),
]
with SupervisedTrainer(
model,
optimizer,
F.cross_entropy,
callbacks=c,
scheduler=scheduler,
) as trainer:
for _ in c[-1]:
trainer.train(train_loader)
trainer.test(test_loader)
def main(cfg):
model = wrn28_2(num_classes=10)
train_loader, test_loader = get_dataloaders(cfg.data.name,
cfg.data.batch_size,
cfg.data.train_size,
cfg.data.random_state)
optimizer = optim.Adam(lr=cfg.optim.lr)
tq = reporters.TQDMReporter(range(cfg.optim.epochs))
c = [
callbacks.AccuracyCallback(),
callbacks.LossCallback(),
reporters.IOReporter("."), tq
]
with SupervisedTrainer(
model,
optimizer,
F.cross_entropy,
callbacks=c,
ema_model=partial(EMAModel,
ema_rate=cfg.model.ema_rate,
weight_decay=cfg.optim.wd * cfg.optim.lr),
) as trainer:
for _ in tq:
trainer.train(train_loader)
trainer.test(test_loader)
trainer.logger.info(
f"test accuracy: {median(c[0].history['test'][-20:])}")
def main():
if args.distributed:
init_distributed()
if args.enable_accimage:
enable_accimage()
model = resnet50()
optimizer = optim.SGD(lr=1e-1 * args.batch_size * get_num_nodes() / 256,
momentum=0.9,
weight_decay=1e-4)
scheduler = lr_scheduler.MultiStepLR([30, 60, 80])
c = [callbacks.AccuracyCallback(), callbacks.LossCallback()]
r = reporters.TQDMReporter(range(args.epochs), callbacks=c)
tb = reporters.TensorboardReporter(c)
rep = callbacks.CallbackList(r, tb, callbacks.WeightSave("checkpoints"))
_train_loader, _test_loader = imagenet_loaders(
args.root,
args.batch_size,
distributed=args.distributed,
num_train_samples=args.batch_size * 10 if args.debug else None,
num_test_samples=args.batch_size * 10 if args.debug else None)
if args.distributed:
# DistributedSupervisedTrainer sets up torch.distributed
if args.local_rank == 0:
print("\nuse DistributedDataParallel\n")
trainer = DistributedSupervisedTrainer(model,
optimizer,
F.cross_entropy,
callbacks=rep,
scheduler=scheduler,
init_method=args.init_method,
backend=args.backend,
enable_amp=args.enable_amp)
else:
use_multi_gpus = torch.cuda.device_count() > 1
if use_multi_gpus:
print("\nuse DataParallel\n")
trainer = SupervisedTrainer(model,
optimizer,
F.cross_entropy,
callbacks=rep,
data_parallel=use_multi_gpus)
for epoch in r:
if args.use_prefetcher:
train_loader = prefetcher.DataPrefetcher(_train_loader)
test_loader = prefetcher.DataPrefetcher(_test_loader)
else:
train_loader, test_loader = _train_loader, _test_loader
# following apex's training scheme
trainer.train(train_loader)
trainer.test(test_loader)
rep.close()
def main(cfg):
model = {
"resnet20": resnet20,
"wrn28_10": wrn28_10
}[cfg.model](num_classes=10)
weight_decay = {"resnet20": 1e-4, "wrn28_10": 5e-4}[cfg.model]
lr_decay = {"resnet20": 0.1, "wrn28_10": 0.2}[cfg.model]
train_loader, test_loader = vision_loaders("cifar10", cfg.batch_size)
optimizer = None if cfg.bn_no_wd else optim.SGD(
lr=1e-1, momentum=0.9, weight_decay=weight_decay)
scheduler = lr_scheduler.MultiStepLR([100, 150], gamma=lr_decay)
tq = reporters.TQDMReporter(range(cfg.epochs), verb=True)
c = [
callbacks.AccuracyCallback(),
callbacks.LossCallback(),
reporters.IOReporter("."),
reporters.TensorboardReporter("."),
callbacks.WeightSave("."), tq
]
if cfg.bn_no_wd:
def set_optimizer(trainer):
bn_params = []
non_bn_parameters = []
for name, p in trainer.model.named_parameters():
if "bn" in name:
bn_params.append(p)
else:
non_bn_parameters.append(p)
optim_params = [
{
"params": bn_params,
"weight_decay": 0
},
{
"params": non_bn_parameters,
"weight_decay": weight_decay
},
]
trainer.optimizer = torch.optim.SGD(optim_params,
lr=1e-1,
momentum=0.9)
trainers.SupervisedTrainer.set_optimizer = set_optimizer
with trainers.SupervisedTrainer(model,
optimizer,
F.cross_entropy,
callbacks=c,
scheduler=scheduler) as trainer:
for _ in tq:
trainer.train(train_loader)
trainer.test(test_loader)
def main(cfg):
if cfg.distributed.enable:
init_distributed(use_horovod=cfg.distributed.use_horovod,
backend=cfg.distributed.backend,
init_method=cfg.distributed.init_method)
if cfg.enable_accimage:
enable_accimage()
model = resnet50()
optimizer = optim.SGD(lr=1e-1 * cfg.batch_size * get_num_nodes() / 256,
momentum=0.9,
weight_decay=1e-4)
scheduler = lr_scheduler.MultiStepLR([30, 60, 80])
tq = reporters.TQDMReporter(range(cfg.epochs))
c = [
callbacks.AccuracyCallback(),
callbacks.AccuracyCallback(k=5),
callbacks.LossCallback(), tq,
reporters.TensorboardReporter("."),
reporters.IOReporter(".")
]
_train_loader, _test_loader = imagenet_loaders(
cfg.root,
cfg.batch_size,
distributed=cfg.distributed.enable,
num_train_samples=cfg.batch_size * 10 if cfg.debug else None,
num_test_samples=cfg.batch_size * 10 if cfg.debug else None)
use_multi_gpus = not cfg.distributed.enable and torch.cuda.device_count(
) > 1
with SupervisedTrainer(model,
optimizer,
F.cross_entropy,
callbacks=c,
scheduler=scheduler,
data_parallel=use_multi_gpus,
use_horovod=cfg.distributed.use_horovod) as trainer:
for epoch in tq:
if cfg.use_prefetcher:
train_loader = prefetcher.DataPrefetcher(_train_loader)
test_loader = prefetcher.DataPrefetcher(_test_loader)
else:
train_loader, test_loader = _train_loader, _test_loader
# following apex's training scheme
trainer.train(train_loader)
trainer.test(test_loader)
def train_and_eval(cfg: BaseConfig):
if cfg.path is None:
print('cfg.path is None, so FasterAutoAugment is not used')
policy = None
else:
path = Path(hydra.utils.get_original_cwd()) / cfg.path
assert path.exists()
policy_weight = torch.load(path, map_location='cpu')
policy = Policy.faster_auto_augment_policy(
num_chunks=cfg.model.num_chunks, **policy_weight['policy_kwargs'])
policy.load_state_dict(policy_weight['policy'])
train_loader, test_loader, num_classes = DATASET_REGISTRY(cfg.data.name)(
batch_size=cfg.data.batch_size,
drop_last=True,
download=cfg.data.download,
return_num_classes=True,
norm=[
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
],
num_workers=4)
model = MODEL_REGISTRY(cfg.model.name)(num_classes)
optimizer = optim.SGD(cfg.optim.lr,
momentum=cfg.optim.momentum,
weight_decay=cfg.optim.weight_decay,
nesterov=cfg.optim.nesterov)
scheduler = lr_scheduler.CosineAnnealingWithWarmup(
cfg.optim.epochs, cfg.optim.scheduler.mul, cfg.optim.scheduler.warmup)
tqdm = callbacks.TQDMReporter(range(cfg.optim.epochs))
c = [callbacks.LossCallback(), callbacks.AccuracyCallback(), tqdm]
with EvalTrainer(model,
optimizer,
F.cross_entropy,
callbacks=c,
scheduler=scheduler,
policy=policy,
cfg=cfg.model,
use_cuda_nonblocking=True) as trainer:
for _ in tqdm:
trainer.train(train_loader)
trainer.test(test_loader)
print(f"Min. Error Rate: {1 - max(c[1].history['test']):.3f}")
def main():
Trainer = trainers.SupervisedTrainer if args.baseline else MixupTrainer
model = MODELS[args.model](num_classes=NUMCLASSES[args.dataset])
train_loader, test_loader = DATASETS[args.dataset](args.batch_size)
optimizer = optim.SGD(lr=1e-1, momentum=0.9, weight_decay=1e-4)
scheduler = lr_scheduler.MultiStepLR(args.steps, gamma=0.1)
c = [callbacks.AccuracyCallback(), callbacks.LossCallback()]
with reporters.TQDMReporter(
range(args.epochs),
callbacks=c) as tq, reporters.TensorboardReporter(c) as tb:
trainer = Trainer(model,
optimizer,
naive_cross_entropy_loss,
callbacks=[tq, tb],
scheduler=scheduler,
alpha=args.alpha,
num_classes=NUMCLASSES[args.dataset])
for _ in tq:
trainer.train(train_loader)
trainer.test(test_loader)
def main():
model = MODELS[args.teacher_model](num_classes=10)
train_loader, test_loader = cifar10_loaders(args.batch_size)
weight_decay = 1e-4 if "resnet" in args.teacher_model else 5e-4
lr_decay = 0.1 if "resnet" in args.teacher_model else 0.2
optimizer = optim.SGD(lr=1e-1, momentum=0.9, weight_decay=weight_decay)
scheduler = lr_scheduler.MultiStepLR([50, 80], gamma=lr_decay)
trainer = trainers.SupervisedTrainer(model,
optimizer,
F.cross_entropy,
scheduler=scheduler)
trainer.logger.info("Train the teacher model!")
for _ in trange(args.teacher_epochs, ncols=80):
trainer.train(train_loader)
trainer.test(test_loader)
teacher_model = model.eval()
weight_decay = 1e-4 if "resnet" in args.student_model else 5e-4
lr_decay = 0.1 if "resnet" in args.student_model else 0.2
optimizer = optim.SGD(lr=1e-1, momentum=0.9, weight_decay=weight_decay)
scheduler = lr_scheduler.MultiStepLR([50, 80], gamma=lr_decay)
model = MODELS[args.student_model](num_classes=10)
c = [callbacks.AccuracyCallback(), callbacks.LossCallback(), kl_loss]
with reporters.TQDMReporter(
range(args.student_epochs),
callbacks=c) as tq, reporters.TensorboardReporter(c) as tb:
trainer = DistillationTrainer(model,
optimizer,
F.cross_entropy,
callbacks=[tq, tb],
scheduler=scheduler,
teacher_model=teacher_model,
temperature=args.temperature)
trainer.logger.info("Train the student model!")
for _ in tq:
trainer.train(train_loader)
trainer.test(test_loader)
def search(cfg: BaseConfig):
train_loader, _, num_classes = DATASET_REGISTRY(cfg.data.name)(
batch_size=cfg.data.batch_size,
train_size=cfg.data.train_size,
drop_last=True,
download=cfg.data.download,
return_num_classes=True,
num_workers=4)
model = {
'main':
Discriminator(MODEL_REGISTRY('wrn40_2')(num_classes)),
'policy':
Policy.faster_auto_augment_policy(cfg.model.num_sub_policies,
cfg.model.temperature,
cfg.model.operation_count,
cfg.model.num_chunks)
}
optimizer = {
'main': optim.Adam(lr=cfg.optim.main_lr, betas=(0, 0.999)),
'policy': optim.Adam(lr=cfg.optim.policy_lr, betas=(0, 0.999))
}
tqdm = callbacks.TQDMReporter(range(cfg.optim.epochs))
c = [
callbacks.LossCallback(), # classification loss
callbacks.metric_callback_by_name('d_loss'), # discriminator loss
callbacks.metric_callback_by_name('a_loss'), # augmentation loss
tqdm
]
with AdvTrainer(model,
optimizer,
F.cross_entropy,
callbacks=c,
cfg=cfg.model,
use_cuda_nonblocking=True) as trainer:
for _ in tqdm:
trainer.train(train_loader)
trainer.save(
pathlib.Path(hydra.utils.get_original_cwd()) / 'policy_weights' /
cfg.data.name)