def test_basic_trainer():
model = nn.Linear(10, 10)
optimizer = optim.SGD()
scheduler = lr_scheduler.StepLR(9)
trainer = trainers.SupervisedTrainer(model,
optimizer,
F.cross_entropy,
scheduler=scheduler,
update_scheduler_by_epoch=False)
loader = [(torch.randn(2, 10), torch.zeros(2, dtype=torch.long))
for _ in range(10)]
for _ in trainer.epoch_range(1):
trainer.train(loader)
assert pytest.approx(trainer.optimizer.param_groups[0]["lr"], 0.01)
optimizer = torch.optim.SGD(model.parameters(), lr=1e-1)
trainer = trainers.SupervisedTrainer(model,
optimizer,
F.cross_entropy,
scheduler=scheduler,
update_scheduler_by_epoch=False)
for _ in trainer.epoch_range(1):
trainer.train(loader)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer, 9)
trainer = trainers.SupervisedTrainer(model,
optimizer,
F.cross_entropy,
scheduler=scheduler,
update_scheduler_by_epoch=False)
trainer.run(loader, loader, 15, 11)
assert trainer.step == 11 - 1
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 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(cfg):
if cfg.use_accimage:
enable_accimage()
model = MODEL_REGISTRY(cfg.model.name)(num_classes=10)
train_loader, test_loader = DATASET_REGISTRY("fast_cifar10" if cfg.use_fast_collate else "cifar10"
)(cfg.data.batch_size, num_workers=4,
use_prefetcher=cfg.use_prefetcher)
optimizer = None if cfg.bn_no_wd else optim.SGD(lr=1e-1, momentum=0.9, weight_decay=cfg.optim.weight_decay)
scheduler = lr_scheduler.MultiStepLR([100, 150], gamma=cfg.optim.lr_decay)
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": cfg.optim.weight_decay},
]
trainer.optimizer = torch.optim.SGD(optim_params, lr=1e-1, momentum=0.9)
trainers.SupervisedTrainer.set_optimizer = set_optimizer
if cfg.use_zerograd_none:
import types
def set_optimizer(trainer):
# see Apex for details
def zero_grad(self):
for group in self.param_groups:
for p in group['params']:
p.grad = None
trainer.optimizer = trainer.optimizer(trainer.model.parameters())
trainer.optimizer.zero_grad = types.MethodType(zero_grad, trainer.optimizer)
trainers.SupervisedTrainer.set_optimizer = set_optimizer
with trainers.SupervisedTrainer(model,
optimizer,
F.cross_entropy,
reporters=[reporters.TensorboardReporter('.')],
scheduler=scheduler,
use_amp=cfg.use_amp,
debug=cfg.debug
) as trainer:
for _ in trainer.epoch_range(cfg.optim.epochs):
trainer.train(train_loader)
trainer.test(test_loader)
print(f"Max Test Accuracy={max(trainer.reporter.history('accuracy/test')):.3f}")
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.use_accimage:
enable_accimage()
model = MODEL_REGISTRY(cfg.name)(num_classes=10)
train_loader, test_loader = DATASET_REGISTRY(
"fast_cifar10" if cfg.use_fast_collate else "cifar10")(
cfg.batch_size, num_workers=4, use_prefetcher=cfg.use_prefetcher)
optimizer = None if cfg.bn_no_wd else optim.SGD(
lr=cfg.lr, momentum=0.9, weight_decay=cfg.weight_decay)
scheduler = lr_scheduler.CosineAnnealingWithWarmup(cfg.epochs, 4, 5)
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": cfg.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,
reporters=[reporters.TensorboardReporter('.')],
scheduler=scheduler,
use_amp=cfg.use_amp,
debug=cfg.debug) as trainer:
for _ in trainer.epoch_range(cfg.epochs):
trainer.train(train_loader)
trainer.test(test_loader)
trainer.scheduler.step()
print(
f"Max Test Accuracy={max(trainer.reporter.history('accuracy/test')):.3f}"
)
def test(tmp_path, rep, save_freq):
temp_dir = tmp_path / "test"
@callbacks.metric_callback_decorator
def ca(data):
output, target = data["output"], data["data"][1]
return {
i: v
for i, v in enumerate(metrics.classwise_accuracy(output, target))
}
model = nn.Linear(10, 10)
optimizer = optim.SGD(lr=0.1)
c = callbacks.CallbackList(
callbacks.AccuracyCallback(), ca,
callbacks.WeightSave(save_path=temp_dir, save_freq=save_freq))
epoch = range(1)
loader = [(torch.randn(2, 10), torch.zeros(2, dtype=torch.long))
for _ in range(10)]
with {
"tqdm": lambda: reporters.TQDMReporter(epoch, c, temp_dir),
"logger": lambda: reporters.LoggerReporter(c, temp_dir),
"tensorboard": lambda: reporters.TensorboardReporter(c, temp_dir)
}[rep]() as _rep:
tr = trainers.SupervisedTrainer(model,
optimizer,
F.cross_entropy,
callbacks=_rep,
verb=False)
if rep == "tqdm":
epoch = _rep
for _ in epoch:
tr.train(loader)
tr.test(loader)
tr.exit()
try:
# .../test/**/0.pkl
save_file = list(Path(temp_dir).glob("*/*.pkl"))[0]
except IndexError as e:
print(list(Path(temp_dir).glob("*/*")))
raise e
tr.resume(save_file)
c = callbacks.AccuracyCallback()
with {
"tqdm": lambda: reporters.TQDMReporter(epoch, c, temp_dir),
"logger": lambda: reporters.LoggerReporter(c, temp_dir),
"tensorboard": lambda: reporters.TensorboardReporter(c, temp_dir)
}[rep]() as _rep:
inferencer = Inferencer(model, _rep)
inferencer.load(save_file)
inferencer.run(loader)
def test(rep):
tmpdir = str(gettempdir())
if rep == "tensorboard" and not is_tensorboardX_available:
pytest.skip("tensorboardX is not available")
@callbacks.metric_callback_decorator
def ca(data):
output, target = data["output"], data["data"][1]
return {
i: v
for i, v in enumerate(metrics.classwise_accuracy(output, target))
}
model = nn.Linear(10, 10)
optimizer = optim.SGD(lr=0.1)
c = callbacks.CallbackList(callbacks.AccuracyCallback(), ca,
callbacks.WeightSave(tmpdir))
epoch = range(1)
loader = [(torch.randn(2, 10), torch.zeros(2, dtype=torch.long))
for _ in range(10)]
with {
"tqdm": lambda: reporters.TQDMReporter(epoch, c, tmpdir),
"logger": lambda: reporters.LoggerReporter(c, tmpdir),
"tensorboard": lambda: reporters.TensorboardReporter(c, tmpdir)
}[rep]() as _rep:
tr = trainers.SupervisedTrainer(model,
optimizer,
F.cross_entropy,
callbacks=_rep,
verb=False)
if rep == "tqdm":
epoch = _rep
for _ in epoch:
tr.train(loader)
tr.test(loader)
save_file = list(Path(tmpdir).glob("*/*.pkl"))[0]
tr.resume(save_file)
c = callbacks.AccuracyCallback()
with {
"tqdm": lambda: reporters.TQDMReporter(epoch, c, tmpdir),
"logger": lambda: reporters.LoggerReporter(c, tmpdir),
"tensorboard": lambda: reporters.TensorboardReporter(c, tmpdir)
}[rep]() as _rep:
inferencer = Inferencer(model, _rep)
inferencer.load(save_file)
inferencer.run(loader)
def test_update_scheduler():
model = nn.Linear(10, 10)
optimizer = optim.SGD(lr=0.1)
trainer = trainers.SupervisedTrainer(model, optimizer, F.cross_entropy)
trainer.update_scheduler(lr_scheduler.LambdaLR(lambda step: 0.1 ** step),
update_scheduler_by_epoch=False)
loader = [(torch.randn(2, 10), torch.zeros(2, dtype=torch.long)) for _ in range(2)]
trainer.train(loader)
# lambda calculates the factor!
assert list(trainer.optimizer.param_groups)[0]['lr'] == 0.1 ** 2
trainer.update_scheduler(lr_scheduler.LambdaLR(lambda epoch: 0.1 ** epoch, last_epoch=1),
update_scheduler_by_epoch=True)
trainer.train(loader)
assert list(trainer.optimizer.param_groups)[0]['lr'] == 0.1 ** 3
def main(batch_size):
layers = ["layer1.0.conv1", "layer2.0.conv1", "layer3.0.conv1", "fc"]
train_loader, test_loader = cifar10_loaders(128)
weight_save = callbacks.WeightSave("checkpoints")
model = resnet20(num_classes=10)
model2 = deepcopy(model)
optimizer = torch.optim.SGD(params=model.parameters(),
lr=0.1,
momentum=0.9,
weight_decay=1e-4)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer, 50)
trainer = trainers.SupervisedTrainer(model,
optimizer,
F.cross_entropy,
scheduler=scheduler,
callbacks=weight_save,
verb=False)
for ep in trange(100, ncols=80):
trainer.train(train_loader)
hooks1 = [CCAHook(model, name, svd_device=args.device) for name in layers]
hooks2 = [CCAHook(model2, name, svd_device=args.device) for name in layers]
device = next(model.parameters()).device
model2.to(device)
input = hooks1[0].data(train_loader.dataset,
batch_size=batch_size).to(device)
history = []
def distance():
model.eval()
model2.eval()
with torch.no_grad():
model(input)
model2(input)
return [h1.distance(h2) for h1, h2 in zip(hooks1, hooks2)]
# 0 and 99
history.append(distance())
# 29 and 99, ...
for ep in (29, 49, 99):
saved = torch.load(weight_save.save_path / f"{ep}.pkl")
model2.load_state_dict(saved["model"])
history.append(distance())
plot(history, layers)
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 main(cfg):
if cfg.use_accimage:
enable_accimage()
data = DATASET_REGISTRY(cfg.data).setup(
cfg.batch_size,
num_workers=4,
download=cfg.download,
prefetch_factor=cfg.prefetch_factor,
persistent_workers=cfg.persistent_workers)
model = MODEL_REGISTRY(cfg.model)(num_classes=data.num_classes)
optimizer = None if cfg.bn_no_wd else optim.SGD(
lr=cfg.lr,
momentum=0.9,
weight_decay=cfg.weight_decay,
multi_tensor=cfg.use_multi_tensor)
scheduler = lr_scheduler.CosineAnnealingWithWarmup(cfg.epochs, 4, 5)
if cfg.bn_no_wd:
def set_optimizer(trainer):
bn_params = []
non_bn_parameters = []
for name, p in trainer.model.named_parameters():
if "norm" 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": cfg.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,
reporters=[reporters.TensorboardReporter('.')],
scheduler=scheduler,
use_amp=cfg.use_amp,
use_channel_last=cfg.use_channel_last,
debug=cfg.debug) as trainer:
for _ in trainer.epoch_range(cfg.epochs):
trainer.train(data.train_loader)
trainer.test(data.test_loader)
trainer.scheduler.step()
print(
f"Max Test Accuracy={max(trainer.reporter.history('accuracy/test')):.3f}"
)