def main(dataset='cifar10',
data_path='/tmp/data',
output_dir='/tmp/fixmatch',
run_id=None,
seed=1,
block_depth=4,
num_filters=32,
num_labeled=40,
sample_mode='label_dist_min1',
num_epochs=1024,
batches_per_epoch=1024,
labeled_batch_size=64,
unlabeled_batch_size=64 * 7,
unlabeled_weight=1.,
lr=0.03,
momentum=0.9,
nesterov=True,
weight_decay=5e-4,
bn_momentum=1e-3,
exp_moving_avg_decay=1e-3,
threshold=0.95,
labeled_aug='weak',
unlabeled_aug=('weak', 'strong'),
dist_alignment=False,
dist_alignment_batches=128,
dist_alignment_eps=1e-6,
checkpoint_interval=1024,
max_checkpoints=25,
num_workers=4,
mixed_precision=True,
devices=('cuda:0', )):
"""FixMatch training.
Args:
dataset: the dataset to use ('cifar10', 'cifar100', 'svhn')
data_path: dataset root directory
output_dir: directory to save logs and model checkpoints
run_id: name for training run (output will be saved under output_dir/run_id)
seed: random seed
block_depth: WideResNet block depth
num_filters: WideResNet base filter count
num_labeled: number of labeled examples
sample_mode: labeled dataset sampling mode ('equal', 'label_dist', 'label_dist_min1', 'multinomial',
'multinomial_min1')
num_epochs: number of training epochs
batches_per_epoch: number of batches per epoch
labeled_batch_size: number of labeled examples per batch
unlabeled_batch_size: number of unlabeled examples per batch (total batch size will be
labeled_batch_size + 2 * unlabeled_batch_size)
unlabeled_weight: weight of unlabeled loss term
lr: SGD initial learning rate
momentum: SGD momentum parameter
nesterov: whether to use SGD with Nesterov acceleration
weight_decay: weight decay parameter
bn_momentum: batch normalization momentum parameter
exp_moving_avg_decay: model parameter exponential moving average decay
threshold: confidence threshold
labeled_aug: data augmentation mode for labeled examples ('none', 'weak', 'strong', 'weak_noflip',
'strong_noflip'). 'strong' augmentation uses RandAugment. 'noflip' disables horizontal flip augmentation.
unlabeled_aug: pair of augmentations for unlabeled examples
dist_alignment: whether to apply distribution alignment heuristic
dist_alignment_batches: number of batches used to compute moving average of label distribution
dist_alignment_eps: smoothing parameter for estimating label distribution
checkpoint_interval: number of batches between checkpoints
max_checkpoints: maximum number of checkpoints to retain
num_workers: number of workers per data loader
mixed_precision: whether to use mixed precision training
devices: list of devices for data parallel training
"""
# initial setup
num_batches = num_epochs * batches_per_epoch
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
args = dict(locals())
logger.info(pprint.pformat(args))
run_id = datetime.datetime.now().isoformat() if run_id is None else run_id
output_dir = os.path.join(output_dir, str(run_id))
logger.info('output dir = %s' % output_dir)
if not os.path.isdir(output_dir):
os.makedirs(output_dir)
with open(os.path.join(output_dir, 'args.pkl'), 'wb') as f:
pickle.dump(args, f)
train_logger, eval_logger = TableLogger(), TableLogger()
# load datasets
if dataset == 'cifar10':
dataset_fn = get_cifar10
elif dataset == 'cifar100':
dataset_fn = get_cifar100
elif dataset == 'svhn':
dataset_fn = get_svhn
else:
raise ValueError('Invalid dataset ' + dataset)
datasets = dataset_fn(data_path,
num_labeled,
labeled_aug=labeled_aug,
unlabeled_aug=unlabeled_aug,
sample_mode=sample_mode,
whiten=True)
model = modules.WideResNet(num_classes=datasets['labeled'].num_classes,
bn_momentum=bn_momentum,
block_depth=block_depth,
channels=num_filters)
optimizer = partial(torch.optim.SGD,
lr=lr,
momentum=momentum,
nesterov=nesterov,
weight_decay=weight_decay)
scheduler = partial(utils.WarmupCosineLrScheduler,
warmup_iter=0,
max_iter=num_batches)
evaluator = ModelEvaluator(datasets['test'],
labeled_batch_size + unlabeled_batch_size,
num_workers)
param_avg_ctor = partial(modules.EMA, alpha=exp_moving_avg_decay)
def evaluate(model, avg_model, iter):
results = evaluator.evaluate(model, device=devices[0])
avg_results = evaluator.evaluate(avg_model, device=devices[0])
valid_stats = {
'valid_loss': avg_results.log_loss,
'valid_accuracy': avg_results.accuracy,
'valid_loss_noavg': results.log_loss,
'valid_accuracy_noavg': results.accuracy
}
eval_logger.write(iter=iter, **valid_stats)
eval_logger.step()
return avg_results.accuracy
def checkpoint(model,
avg_model,
optimizer,
scheduler,
iter,
fmt='ckpt-{:08d}.pt'):
path = os.path.join(output_dir, fmt.format(iter))
torch.save(
dict(iter=iter,
model=model.state_dict(),
avg_model=avg_model.state_dict(),
optimizer=optimizer.state_dict(),
scheduler=scheduler.state_dict()), path)
checkpoint_files = sorted(
list(
filter(lambda x: re.match(r'^ckpt-[0-9]+.pt$', x),
os.listdir(output_dir))))
if len(checkpoint_files) > max_checkpoints:
os.remove(os.path.join(output_dir, checkpoint_files[0]))
train_logger.to_dataframe().to_pickle(
os.path.join(output_dir, 'train.log.pkl'))
eval_logger.to_dataframe().to_pickle(
os.path.join(output_dir, 'eval.log.pkl'))
trainer = FixMatch(num_iters=num_epochs * batches_per_epoch,
num_workers=num_workers,
model_optimizer_ctor=optimizer,
lr_scheduler_ctor=scheduler,
param_avg_ctor=param_avg_ctor,
labeled_batch_size=labeled_batch_size,
unlabeled_batch_size=unlabeled_batch_size,
unlabeled_weight=unlabeled_weight,
threshold=threshold,
dist_alignment=dist_alignment,
dist_alignment_batches=dist_alignment_batches,
dist_alignment_eps=dist_alignment_eps,
mixed_precision=mixed_precision,
devices=devices)
timer = utils.Timer()
with tqdm(desc='train', total=num_batches, position=0) as train_pbar:
train_iter = utils.Generator(
trainer.train_iter(model, datasets['labeled'],
datasets['unlabeled']))
eval_acc = None
# training loop
for i, stats in enumerate(train_iter):
train_pbar.set_postfix(loss=stats.loss,
eval_acc=eval_acc,
refresh=False)
train_pbar.update()
train_logger.write(loss=stats.loss,
loss_labeled=stats.loss_labeled,
loss_unlabeled=stats.loss_unlabeled,
threshold_frac=stats.threshold_frac,
time=timer())
if (checkpoint_interval is not None and i > 0 and (i+1) % checkpoint_interval == 0) or \
(i == num_batches - 1):
checkpoint(stats.model, stats.avg_model, stats.optimizer,
stats.scheduler, i + 1)
eval_acc = evaluate(stats.model, stats.avg_model, i + 1)
logger.info('eval acc = %.4f | allocated frac = %.4f' %
(eval_acc, stats.threshold_frac))
train_logger.step()
def main(
num_workers=8,
num_filters=32,
dataset='cifar10',
data_path='/tmp/data',
output_dir='/tmp/sla',
run_id=None,
num_labeled=40,
seed=1,
num_epochs=1024,
batches_per_epoch=1024,
checkpoint_interval=1024,
snapshot_interval=None,
max_checkpoints=25,
optimizer='sgd',
lr=0.03,
momentum=0.9,
nesterov=True,
weight_decay=5e-4,
bn_momentum=1e-3,
labeled_batch_size=64,
unlabeled_batch_size=64*7,
unlabeled_weight=1.,
exp_moving_avg_decay=1e-3,
allocation_schedule=((0., 1.), (0., 1.)),
entropy_reg=100.,
update_tol=0.01,
labeled_aug='weak',
unlabeled_aug=('weak', 'strong'),
whiten=True,
sample_mode='label_dist_min1',
upper_bound_method='empirical',
upper_bound_kwargs={},
mixed_precision=True,
devices=('cuda:0',)):
# initial setup
num_batches = num_epochs * batches_per_epoch
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
args = dict(locals())
logger.info(pprint.pformat(args))
run_id = datetime.datetime.now().isoformat() if run_id is None else run_id
output_dir = os.path.join(output_dir, str(run_id))
logger.info('output dir = %s' % output_dir)
if not os.path.isdir(output_dir):
os.makedirs(output_dir)
with open(os.path.join(output_dir, 'args.pkl'), 'wb') as f:
pickle.dump(args, f)
train_logger, eval_logger = TableLogger(), TableLogger()
# load datasets
if dataset == 'cifar10':
dataset_fn = get_cifar10
elif dataset == 'cifar100':
dataset_fn = get_cifar100
elif dataset == 'svhn':
dataset_fn = get_svhn
else:
raise ValueError('Invalid dataset ' + dataset)
datasets = dataset_fn(
data_path, num_labeled, labeled_aug=labeled_aug, unlabeled_aug=unlabeled_aug,
sample_mode=sample_mode, whiten=whiten)
model = modules.WideResNet(
num_classes=datasets['labeled'].num_classes, bn_momentum=bn_momentum, channels=num_filters)
optimizer = partial(torch.optim.SGD, lr=lr, momentum=momentum, nesterov=nesterov, weight_decay=weight_decay)
scheduler = partial(utils.WarmupCosineLrScheduler, warmup_iter=0, max_iter=num_batches)
evaluator = ModelEvaluator(datasets['test'], labeled_batch_size + unlabeled_batch_size, num_workers)
def evaluate(model, avg_model, iter):
results = evaluator.evaluate(model, device=devices[0])
avg_results = evaluator.evaluate(avg_model, device=devices[0])
valid_stats = {
'valid_loss': avg_results.log_loss,
'valid_accuracy': avg_results.accuracy,
'valid_loss_noavg': results.log_loss,
'valid_accuracy_noavg': results.accuracy
}
eval_logger.write(
iter=iter,
**valid_stats)
eval_logger.step()
return avg_results.accuracy
def checkpoint(model, avg_model, optimizer, scheduler, iter, fmt='ckpt-{:08d}.pt'):
path = os.path.join(output_dir, fmt.format(iter))
torch.save(dict(
iter=iter,
model=model.state_dict(),
avg_model=avg_model.state_dict(),
optimizer=optimizer.state_dict(),
scheduler=scheduler.state_dict()), path)
checkpoint_files = sorted(list(filter(lambda x: re.match(r'^ckpt-[0-9]+.pt$', x), os.listdir(output_dir))))
if len(checkpoint_files) > max_checkpoints:
os.remove(os.path.join(output_dir, checkpoint_files[0]))
train_logger.to_dataframe().to_pickle(os.path.join(output_dir, 'train.log.pkl'))
eval_logger.to_dataframe().to_pickle(os.path.join(output_dir, 'eval.log.pkl'))
trainer = SLASelfTraining(
num_epochs=num_epochs,
batches_per_epoch=batches_per_epoch,
num_workers=num_workers,
model_optimizer_ctor=optimizer,
lr_scheduler_ctor=scheduler,
param_avg_ctor=partial(modules.EMA, alpha=exp_moving_avg_decay),
labeled_batch_size=labeled_batch_size,
unlabeled_batch_size=unlabeled_batch_size,
unlabeled_weight=unlabeled_weight,
allocation_schedule=utils.PiecewiseLinear(*allocation_schedule),
entropy_reg=entropy_reg,
update_tol=update_tol,
upper_bound_method=upper_bound_method,
upper_bound_kwargs=upper_bound_kwargs,
mixed_precision=mixed_precision,
devices=devices)
timer = utils.Timer()
with tqdm(desc='train', total=num_batches, position=0) as train_pbar:
train_iter = utils.Generator(
trainer.train_iter(model, datasets['labeled'].num_classes, datasets['labeled'], datasets['unlabeled']))
smoothed_loss = utils.ema(0.3, avg_only=True)
smoothed_loss.send(None)
smoothed_acc = utils.ema(1., avg_only=False)
smoothed_acc.send(None)
eval_stats = None, None
# training loop
for i, stats in enumerate(train_iter):
if isinstance(stats, trainer.__class__.Stats):
train_pbar.set_postfix(
loss=smoothed_loss.send(stats.loss), eval_acc=eval_stats[0], eval_v=eval_stats[1], refresh=False)
train_pbar.update()
train_logger.write(
loss=stats.loss, loss_labeled=stats.loss_labeled, loss_unlabeled=stats.loss_unlabeled,
mean_imputed_labels=stats.label_vars.data.mean(0).cpu().numpy(),
scaling_vars=stats.scaling_vars.data.cpu().numpy(),
allocation_param=stats.allocation_param,
assigned_frac=stats.label_vars.data.sum(-1).mean(),
assignment_err=stats.assgn_err, assignment_iters=stats.assgn_iters, time=timer())
if (checkpoint_interval is not None
and i > 0 and (i + 1) % checkpoint_interval == 0) or (i == num_batches - 1):
eval_acc = evaluate(stats.model, stats.avg_model, i+1)
eval_stats = smoothed_acc.send(eval_acc)
checkpoint(stats.model, stats.avg_model, stats.optimizer, stats.scheduler, i+1)
logger.info('eval acc = %.4f | allocated frac = %.4f | allocation param = %.4f' %
(eval_acc, stats.label_vars.mean(0).sum().cpu().item(), stats.allocation_param))
logger.info('assignment err = %.4e | assignment iters = %d' % (stats.assgn_err, stats.assgn_iters))
logger.info('batch assignments = {}'.format(stats.label_vars.mean(0).cpu().numpy()))
logger.info('scaling vars = {}'.format(stats.scaling_vars.cpu().numpy()))
# take snapshots that are guaranteed to be preserved
if snapshot_interval is not None and i > 0 and (i + 1) % snapshot_interval == 0:
checkpoint(stats.model, stats.avg_model, stats.optimizer,
stats.scheduler, i + 1, 'snapshot-{:08d}.pt')
train_logger.step()