def lr_search(train_eval_data):
train_eval_dataset = TrainEvalDataset(train_eval_data,
transform=train_transform)
train_eval_data_loader = torch.utils.data.DataLoader(
train_eval_dataset,
batch_size=config.batch_size,
drop_last=True,
shuffle=True,
num_workers=args.workers,
worker_init_fn=worker_init_fn)
min_lr = 1e-7
max_lr = 10.
gamma = (max_lr / min_lr)**(1 / len(train_eval_data_loader))
lrs = []
losses = []
lim = None
model = Model(config.model, NUM_CLASSES)
model = model.to(DEVICE)
optimizer = build_optimizer(config.opt, model.parameters())
for param_group in optimizer.param_groups:
param_group['lr'] = min_lr
scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer, gamma)
optimizer.train()
update_transforms(1.)
model.train()
optimizer.zero_grad()
for i, (images, _, labels,
_) in enumerate(tqdm(train_eval_data_loader, desc='lr search'), 1):
images, labels = images.to(DEVICE), labels.to(DEVICE)
labels = utils.one_hot(labels, NUM_CLASSES)
images, labels = mixup(images, labels)
logits = model(images, None, True)
loss = compute_loss(input=logits, target=labels)
labels = labels.argmax(1)
lrs.append(np.squeeze(scheduler.get_lr()))
losses.append(loss.data.cpu().numpy().mean())
if lim is None:
lim = losses[0] * 1.1
if lim < losses[-1]:
break
(loss.mean() / config.opt.acc_steps).backward()
if i % config.opt.acc_steps == 0:
optimizer.step()
optimizer.zero_grad()
scheduler.step()
writer = SummaryWriter(os.path.join(args.experiment_path, 'lr_search'))
with torch.no_grad():
losses = np.clip(losses, 0, lim)
minima_loss = losses[np.argmin(utils.smooth(losses))]
minima_lr = lrs[np.argmin(utils.smooth(losses))]
step = 0
for loss, loss_sm in zip(losses, utils.smooth(losses)):
writer.add_scalar('search_loss', loss, global_step=step)
writer.add_scalar('search_loss_sm', loss_sm, global_step=step)
step += config.batch_size
fig = plt.figure()
plt.plot(lrs, losses)
plt.plot(lrs, utils.smooth(losses))
plt.axvline(minima_lr)
plt.xscale('log')
plt.title('loss: {:.8f}, lr: {:.8f}'.format(minima_loss, minima_lr))
writer.add_figure('search', fig, global_step=0)
return minima_lr
def train_fold(fold, train_eval_data, unsup_data):
train_indices, eval_indices = indices_for_fold(fold, train_eval_data)
train_dataset = TrainEvalDataset(train_eval_data.iloc[train_indices],
transform=train_transform)
train_data_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=config.batch_size,
drop_last=True,
shuffle=True,
num_workers=args.workers,
worker_init_fn=worker_init_fn)
eval_dataset = TrainEvalDataset(train_eval_data.iloc[eval_indices],
transform=eval_transform)
eval_data_loader = torch.utils.data.DataLoader(
eval_dataset,
batch_size=config.batch_size,
num_workers=args.workers,
worker_init_fn=worker_init_fn)
unsup_data = TestDataset(unsup_data, transform=unsup_transform)
unsup_data_loader = torch.utils.data.DataLoader(
unsup_data,
batch_size=config.batch_size // 2,
num_workers=args.workers,
worker_init_fn=worker_init_fn)
model = Model(config.model, NUM_CLASSES)
model = model.to(DEVICE)
if args.restore_path is not None:
model.load_state_dict(
torch.load(
os.path.join(args.restore_path, 'model_{}.pth'.format(fold))))
optimizer = build_optimizer(config.opt, model.parameters())
if config.sched.type == 'onecycle':
scheduler = lr_scheduler_wrapper.StepWrapper(
OneCycleScheduler(optimizer,
lr=(config.opt.lr / 20, config.opt.lr),
beta_range=config.sched.onecycle.beta,
max_steps=len(train_data_loader) * config.epochs,
annealing=config.sched.onecycle.anneal,
peak_pos=config.sched.onecycle.peak_pos,
end_pos=config.sched.onecycle.end_pos))
elif config.sched.type == 'step':
scheduler = lr_scheduler_wrapper.EpochWrapper(
torch.optim.lr_scheduler.StepLR(
optimizer,
step_size=config.sched.step.step_size,
gamma=config.sched.step.decay))
elif config.sched.type == 'cyclic':
step_size_up = len(
train_data_loader) * config.sched.cyclic.step_size_up
step_size_down = len(
train_data_loader) * config.sched.cyclic.step_size_down
scheduler = lr_scheduler_wrapper.StepWrapper(
torch.optim.lr_scheduler.CyclicLR(
optimizer,
0.,
config.opt.lr,
step_size_up=step_size_up,
step_size_down=step_size_down,
mode='triangular2',
gamma=config.sched.cyclic.decay**(
1 / (step_size_up + step_size_down)),
cycle_momentum=True,
base_momentum=0.85,
max_momentum=0.95))
elif config.sched.type == 'cawr':
scheduler = lr_scheduler_wrapper.StepWrapper(
torch.optim.lr_scheduler.CosineAnnealingWarmRestarts(
optimizer, T_0=len(train_data_loader), T_mult=2))
elif config.sched.type == 'plateau':
scheduler = lr_scheduler_wrapper.ScoreWrapper(
torch.optim.lr_scheduler.ReduceLROnPlateau(
optimizer,
mode='max',
factor=config.sched.plateau.decay,
patience=config.sched.plateau.patience,
verbose=True))
else:
raise AssertionError('invalid sched {}'.format(config.sched.type))
best_score = 0
for epoch in range(1, config.epochs + 1):
optimizer.train()
train_epoch(model=model,
optimizer=optimizer,
scheduler=scheduler,
data_loader=train_data_loader,
unsup_data_loader=unsup_data_loader,
fold=fold,
epoch=epoch)
gc.collect()
optimizer.eval()
metric = eval_epoch(model=model,
data_loader=eval_data_loader,
fold=fold,
epoch=epoch)
gc.collect()
score = metric['accuracy@1']
scheduler.step_epoch()
scheduler.step_score(score)
if score > best_score:
best_score = score
torch.save(
model.state_dict(),
os.path.join(args.experiment_path,
'model_{}.pth'.format(fold)))
def train_fold(fold, train_eval_data):
model = Model(config.model, NUM_CLASSES)
model = model.to(DEVICE)
if args.restore_path is not None:
model.load_state_dict(
torch.load(
os.path.join(args.restore_path, 'model_{}.pth'.format(fold))))
optimizer = build_optimizer(config.opt, model.parameters())
if config.sched.type == 'onecycle':
scheduler = lr_scheduler_wrapper.EpochWrapper(
OneCycleScheduler(optimizer,
lr=(config.opt.lr / 20, config.opt.lr),
beta_range=config.sched.onecycle.beta,
max_steps=config.epochs,
annealing=config.sched.onecycle.anneal,
peak_pos=config.sched.onecycle.peak_pos,
end_pos=config.sched.onecycle.end_pos))
elif config.sched.type == 'cyclic':
step_size_up = len(
train_data_loader) * config.sched.cyclic.step_size_up
step_size_down = len(
train_data_loader) * config.sched.cyclic.step_size_down
scheduler = lr_scheduler_wrapper.StepWrapper(
torch.optim.lr_scheduler.CyclicLR(
optimizer,
0.,
config.opt.lr,
step_size_up=step_size_up,
step_size_down=step_size_down,
mode='exp_range',
gamma=config.sched.cyclic.decay**(
1 / (step_size_up + step_size_down)),
cycle_momentum=True,
base_momentum=0.75,
max_momentum=0.95))
elif config.sched.type == 'cawr':
scheduler = lr_scheduler_wrapper.StepWrapper(
torch.optim.lr_scheduler.CosineAnnealingWarmRestarts(
optimizer, T_0=len(train_data_loader), T_mult=2))
elif config.sched.type == 'plateau':
scheduler = lr_scheduler_wrapper.ScoreWrapper(
torch.optim.lr_scheduler.ReduceLROnPlateau(
optimizer,
mode='max',
factor=config.sched.plateau.decay,
patience=config.sched.plateau.patience,
verbose=True))
else:
raise AssertionError('invalid sched {}'.format(config.sched.type))
best_score = 0
for epoch in range(1, config.epochs + 1):
train_indices, eval_indices = indices_for_fold(fold, train_eval_data)
eval_pl = pd.read_csv(
'./tf_log/cells/tmp-512-progres-crop-norm-la/eval_{}.csv'.format(
fold))
eval_pl['root'] = os.path.join(args.dataset_path, 'train')
test_pl = pd.read_csv(
'./tf_log/cells/tmp-512-progres-crop-norm-la/test.csv')
test_pl['root'] = os.path.join(args.dataset_path, 'test')
pl = pd.concat([eval_pl, test_pl])
pl_size = len(pl)
pl = pl.sample(frac=np.linspace(1., 0., config.epochs)[epoch -
1].item())
print('frac: {:.4f}, lr: {:.8f}'.format(
len(pl) / pl_size, scheduler.get_lr()))
train_dataset = TrainEvalDataset(pd.concat(
[train_eval_data.iloc[train_indices], pl]),
transform=train_transform)
train_data_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=config.batch_size,
drop_last=True,
shuffle=True,
num_workers=args.workers,
worker_init_fn=worker_init_fn)
eval_dataset = TrainEvalDataset(train_eval_data.iloc[eval_indices],
transform=eval_transform)
eval_data_loader = torch.utils.data.DataLoader(
eval_dataset,
batch_size=config.batch_size,
num_workers=args.workers,
worker_init_fn=worker_init_fn)
train_epoch(model=model,
optimizer=optimizer,
scheduler=scheduler,
data_loader=train_data_loader,
fold=fold,
epoch=epoch)
gc.collect()
metric = eval_epoch(model=model,
data_loader=eval_data_loader,
fold=fold,
epoch=epoch)
gc.collect()
score = metric['accuracy@1']
scheduler.step_epoch()
scheduler.step_score(score)
if score > best_score:
best_score = score
torch.save(
model.state_dict(),
os.path.join(args.experiment_path,
'model_{}.pth'.format(fold)))