def datasets(self, params: SemiSupervisedParams):
dataset_fn = datasets[params.dataset]
test_x, test_y = dataset_fn(False)
train_x, train_y = dataset_fn(True)
indexs, un_indexs, val_indexs = splits.semi_split(
train_y,
n_percls=params.n_percls,
val_size=params.val_size,
repeat_sup=False)
self.logger.info('sup/unsup/val : {}'.format(
(len(indexs), len(un_indexs), len(val_indexs))))
mean, std = norm_val.get(params.dataset, [None, None])
toTensor = ToNormTensor(mean, std)
weak = Weak(mean, std)
strong = Strong(mean, std)
sup_set = (DatasetBuilder(
train_x, train_y).add_x(transform=weak).add_y().subset(indexs))
if len(sup_set) < params.batch_size:
sup_set.virtual_sample(params.batch_size)
unsup_set = (DatasetBuilder(train_x, train_y).toggle_id().add_x(
transform=weak).add_x(transform=strong).add_y().subset(un_indexs))
self.cl_set = unsup_set
sup_dataloader = sup_set.DataLoader(batch_size=params.batch_size,
num_workers=params.num_workers,
shuffle=True)
self.sup_dataloader = sup_dataloader
unsup_dataloader = unsup_set.DataLoader(batch_size=params.batch_size *
params.uratio,
num_workers=1,
shuffle=True)
self.unsup_dataloader = DataBundler().add(unsup_dataloader).to(
self.device)
val_dataloader = (DatasetBuilder(
train_x[val_indexs],
train_y[val_indexs]).add_x(transform=toTensor).add_y().DataLoader(
batch_size=params.batch_size,
num_workers=params.num_workers,
shuffle=True))
test_dataloader = (DatasetBuilder(
test_x, test_y).add_x(transform=toTensor).add_y().DataLoader(
batch_size=params.batch_size,
num_workers=params.num_workers,
shuffle=True))
self.regist_databundler(train=DataBundler().cycle(sup_dataloader).add(
unsup_dataloader).zip_mode(),
eval=val_dataloader,
test=test_dataloader)
self.to(self.device)
def datasets(self, params: FitEvalParams):
from thexp import DatasetBuilder
from data.constant import norm_val
from data.transforms import ToNormTensor, Weak, Strong
mean, std = norm_val.get(params.dataset, [None, None])
weak = Weak(mean, std)
dataset_fn = datasets.datasets[params.dataset]
train_x, train_y = dataset_fn(True)
train_y = np.array(train_y)
from data.noisy import symmetric_noisy
train_x, train_y = train_x[:params.train_size], train_y[:params.train_size]
part_size = params.train_size // 2
noisy_x1, noisy_true_y1 = train_x[:part_size], train_y[:part_size]
noisy_x2, noisy_true_y2 = train_x[part_size:], train_y[part_size:]
noisy_ratio = 0.9
noisy_y1 = symmetric_noisy(noisy_true_y1, noisy_ratio, n_classes=params.n_classes)
noisy_y2 = symmetric_noisy(noisy_true_y2, noisy_ratio, n_classes=params.n_classes)
self.logger.info('noisy dataset ratio: ',
(noisy_true_y1 == noisy_y1).mean(),
(noisy_true_y2 == noisy_y2).mean())
noisy_set1 = (
DatasetBuilder(noisy_x1, noisy_true_y1)
.add_labels(noisy_y2, source_name='noisy_y')
.toggle_id()
.add_x(transform=weak)
.add_y()
.add_y(source='noisy_y')
)
noisy_set2 = (
DatasetBuilder(noisy_x2, noisy_true_y2)
.add_labels(noisy_y2, source_name='noisy_y')
.toggle_id()
.add_x(transform=weak)
.add_y()
.add_y(source='noisy_y')
)
self.noisy_loader1 = noisy_set1.DataLoader(batch_size=params.batch_size,
num_workers=params.num_workers,
drop_last=True,
shuffle=True)
self.noisy_loader2 = noisy_set2.DataLoader(batch_size=params.batch_size,
num_workers=params.num_workers,
drop_last=True,
shuffle=True)
self.eval_state = 1 - int(params.eval_mode == 'clean')
self.toggle_dataset((self.eval_state % 2) == 0)
self.to(self.device)
def datasets(self, params: DivideMixParams):
from data.dataxy import datasets
dataset_fn = datasets[params.dataset]
test_x, test_y = dataset_fn(False)
train_x, train_y = dataset_fn(True)
mean, std = norm_val.get(params.dataset, [None, None])
toTensor = ToNormTensor(mean, std)
weak = Weak(mean, std)
strong = Strong(mean, std)
if params.noisy_type == 'asymmetric':
from data.noisy import asymmetric_noisy
noisy_y = asymmetric_noisy(train_y,
params.noisy_ratio,
n_classes=params.n_classes)
elif params.noisy_type == 'symmetric':
from data.noisy import symmetric_noisy
noisy_y = symmetric_noisy(train_y,
params.noisy_ratio,
n_classes=params.n_classes)
else:
assert False, params.noisy_type
self.train_set_pack = [train_x, np.array(train_y), noisy_y]
self.logger.info('noisy acc = {}'.format((train_y == noisy_y).mean()))
train_set = (DatasetBuilder(train_x, train_y).add_labels(
noisy_y, 'noisy_y').toggle_id().add_x(transform=weak).add_x(
transform=strong).add_y().add_y(source='noisy_y'))
train_dataloader = train_set.DataLoader(batch_size=params.batch_size *
2,
num_workers=params.num_workers,
shuffle=True)
from thexp import DataBundler
self.eval_train_dataloader = (DataBundler().add(
DatasetBuilder(train_x, noisy_y).toggle_id().add_x(
transform=toTensor).add_y().DataLoader(
batch_size=params.batch_size,
num_workers=params.num_workers // 2,
shuffle=False)).to(self.device))
test_dataloader = (DatasetBuilder(
test_x, test_y).add_x(transform=toTensor).add_y().DataLoader(
batch_size=params.batch_size,
num_workers=params.num_workers // 2,
shuffle=False))
self.regist_databundler(train=train_dataloader, test=test_dataloader)
self.to(self.device)
def on_train_epoch_begin(self, trainer: Trainer, func,
params: DivideMixParams, *args, **kwargs):
if params.eidx < params.warm_up:
pass
else:
if params.eidx % 2 == 0:
prob, self.all_loss[1] = self.eval_train(
self.model2, self.all_loss[1]) # type: np.ndarray, list
pred = (prob > params.p_threshold)
else:
prob, self.all_loss[0] = self.eval_train(
self.model, self.all_loss[0]) # type: np.ndarray, list
pred = (prob > params.p_threshold)
pred_idx = pred.nonzero()[0]
unpred_idx = (1 - pred).nonzero()[0]
train_x, train_y, noisy_y = self.train_set_pack
clean = (noisy_y == train_y)
acc = (pred[clean]).mean()
self.logger.info('Numer of labeled samples', pred.sum(),
'clean ratio = {}'.format(acc))
mean, std = norm_val.get(params.dataset, [None, None])
weak = Weak(mean, std)
labeled_dataloader = (DatasetBuilder(train_x, train_y).add_labels(
noisy_y, source_name='nys').add_labels(
prob, source_name='nprob').add_x(transform=weak).add_x(
transform=weak).add_y().add_y(source='nys').add_y(
source='nprob').subset(pred_idx).DataLoader(
params.batch_size,
shuffle=True,
drop_last=True,
num_workers=params.num_workers))
unlabeled_dataloader = (DatasetBuilder(
train_x, train_y).add_labels(noisy_y, source_name='nys').add_x(
transform=weak).add_x(transform=weak).add_y().add_y(
source='nys').subset(unpred_idx).DataLoader(
params.batch_size,
shuffle=True,
drop_last=True,
num_workers=params.num_workers))
bundler = DataBundler()
bundler.add(labeled_dataloader).cycle(
unlabeled_dataloader).zip_mode()
self.logger.info('new training dataset', bundler)
self.regist_databundler(train=bundler.to(self.device))
def datasets(self, params: SemiSupervisedParams):
dataset_fn = datasets[params.dataset]
test_x, test_y = dataset_fn(False)
train_x, train_y = dataset_fn(True)
indexs, un_indexs, val_indexs = splits.semi_split(
train_y, n_percls=params.n_percls, val_size=5000, repeat_sup=False)
mean, std = norm_val.get(params.dataset, [None, None])
toTensor = ToNormTensor(mean, std)
weak = Weak(mean, std)
sup_set = (DatasetBuilder(
train_x, train_y).add_x(transform=weak).add_y().subset(indexs))
params.K = params.default(2, True)
unsup_set = DatasetBuilder(train_x, train_y)
for _ in range(params.K):
unsup_set.add_x(transform=weak)
unsup_set = unsup_set.add_y().subset(un_indexs)
sup_dataloader = sup_set.DataLoader(batch_size=params.batch_size,
num_workers=params.num_workers,
shuffle=True)
unsup_dataloader = unsup_set.DataLoader(batch_size=params.batch_size,
num_workers=params.num_workers,
shuffle=True)
val_dataloader = (DatasetBuilder(
train_x[val_indexs],
train_y[val_indexs]).add_x(transform=toTensor).add_y().DataLoader(
batch_size=params.batch_size,
num_workers=params.num_workers,
shuffle=True))
test_dataloader = (DatasetBuilder(
test_x, test_y).add_x(transform=toTensor).add_y().DataLoader(
batch_size=params.batch_size,
num_workers=params.num_workers,
shuffle=True))
self.regist_databundler(train=DataBundler().cycle(sup_dataloader).add(
unsup_dataloader).zip_mode(),
eval=val_dataloader,
test=test_dataloader)
self.to(self.device)
def datasets(self, params: GlobalParams):
dataset_fn = datasets[params.dataset]
test_x, testy = dataset_fn(False)
train_x, train_y = dataset_fn(True)
train_idx, val_idx = splits.train_val_split(train_y,
val_size=params.val_size)
mean, std = norm_val.get(params.dataset, [None, None])
toTensor = ToNormTensor(mean, std)
weak = Weak(mean, std)
strong = Strong(mean, std)
test_dataloader = (DatasetBuilder(
test_x, testy).add_x(transform=toTensor).add_y().DataLoader(
batch_size=params.batch_size, num_workers=params.num_workers))
train_dataloader = (DatasetBuilder(train_x, train_y).toggle_id().add_x(
transform=weak).add_x(transform=strong).add_y().subset(train_idx))
if params.distributed:
from torch.utils.data.distributed import DistributedSampler
sampler = DistributedSampler(train_dataloader)
else:
sampler = None
self.train_size = len(train_dataloader)
train_dataloader = train_dataloader.DataLoader(
batch_size=params.batch_size,
num_workers=params.num_workers,
sampler=sampler,
shuffle=not params.distributed)
val_datalaoder = (DatasetBuilder(train_x, train_y).add_x(
transform=toTensor).add_y().subset(val_idx).DataLoader(
batch_size=params.batch_size, num_workers=params.num_workers))
self.regist_databundler(train=train_dataloader,
eval=val_datalaoder,
test=test_dataloader)
print('dataloader in rank {}'.format(self.params.local_rank))
print(self.params.local_rank, self.train_dataloader)
print(self.params.local_rank, self._databundler_dict)
print(self.params.local_rank, train_dataloader)
self.to(self.device)
def datasets(self, params: NoisyParams):
self.rnd.mark('kk')
params.noisy_type = params.default('symmetric', True)
params.noisy_ratio = params.default(0.2, True)
from data.constant import norm_val
mean, std = norm_val.get(params.dataset, [None, None])
from data.transforms import ToNormTensor
toTensor = ToNormTensor(mean, std)
from data.transforms import Weak
weak = Weak(mean, std)
from data.transforms import Strong
dataset_fn = datasets.datasets[params.dataset]
train_x, train_y = dataset_fn(True)
train_y = np.array(train_y)
from thexp import DatasetBuilder
from data.noisy import symmetric_noisy
noisy_y = symmetric_noisy(train_y,
params.noisy_ratio,
n_classes=params.n_classes)
clean_mask = (train_y == noisy_y)
noisy_mask = np.logical_not(clean_mask)
noisy_mask = np.where(noisy_mask)[0]
nmask_a = noisy_mask[:len(noisy_mask) // 2]
nmask_b = noisy_mask[len(noisy_mask) // 2:]
clean_x, clean_y = train_x[clean_mask], noisy_y[clean_mask]
clean_true_y = train_y[clean_mask]
raw_x, raw_true_y = train_x[nmask_a], train_y[nmask_a]
raw_y = noisy_y[nmask_a]
change_x, change_true_y, change_y = train_x[nmask_b], train_y[
nmask_b], noisy_y[nmask_b]
first_x, first_y, first_true_y = (
clean_x + raw_x,
np.concatenate([clean_y, raw_y]),
np.concatenate([clean_true_y, raw_true_y]),
)
second_x, second_y, second_true_y = (
clean_x + change_x,
np.concatenate([clean_y, change_y]),
np.concatenate([clean_true_y, change_true_y]),
)
first_set = (DatasetBuilder(first_x, first_true_y).add_labels(
first_y, 'noisy_y').toggle_id().add_x(
transform=weak).add_y().add_y(source='noisy_y'))
second_set = (DatasetBuilder(second_x, second_true_y).add_labels(
second_y, 'noisy_y').toggle_id().add_x(
transform=weak).add_y().add_y(source='noisy_y'))
self.first_dataloader = first_set.DataLoader(
batch_size=params.batch_size,
num_workers=params.num_workers,
drop_last=True,
shuffle=True)
self.second_dataloader = second_set.DataLoader(
batch_size=params.batch_size,
num_workers=params.num_workers,
drop_last=True,
shuffle=True)
self.second_dataloader = second_set.DataLoader(
batch_size=params.batch_size,
num_workers=params.num_workers,
drop_last=True,
shuffle=True)
self.second = False
self.regist_databundler(train=self.first_dataloader)
self.cur_set = 0
self.to(self.device)
def datasets(self, params: NoisyParams):
params.noisy_type = params.default('symmetric', True)
params.noisy_ratio = params.default(0.2, True)
import numpy as np
dataset_fn = datasets[params.dataset]
test_x, test_y = dataset_fn(False)
train_x, train_y = dataset_fn(True)
# train_ids, query_ids = splits.train_val_split(train_y, val_size=params.val_size)
query_ids, train_ids, eval_ids = splits.semi_split(
train_y,
params.query_size // params.n_classes,
val_size=params.val_size,
repeat_sup=False)
# train_ids = train_ids[:3000]
self.train_size = len(train_ids)
train_x, query_x, eval_x = train_x[train_ids], train_x[
query_ids], train_x[eval_ids]
train_y, query_y, eval_y = train_y[train_ids], train_y[
query_ids], train_y[eval_ids]
mean, std = norm_val.get(params.dataset, [None, None])
toTensor = ToNormTensor(mean, std)
weak = Weak(mean, std)
strong = Strong(mean, std)
if params.noisy_type == 'asymmetric':
from data.noisy import asymmetric_noisy
noisy_y = asymmetric_noisy(train_y,
params.noisy_ratio,
n_classes=params.n_classes)
elif params.noisy_type == 'symmetric':
from data.noisy import symmetric_noisy
noisy_y = symmetric_noisy(train_y,
params.noisy_ratio,
n_classes=params.n_classes)
else:
assert False
self.logger.info('noisy acc = {}'.format((train_y == noisy_y).mean()))
self.rnd.shuffle()
self.logger.info(len(train_y), len(train_x), len(noisy_y))
train_set = (DatasetBuilder(train_x, train_y).add_labels(
noisy_y, 'noisy_y').toggle_id().add_x(transform=strong))
params.K = params.default(0, True)
for _ in range(params.K):
train_set.add_x(transform=weak)
train_set = (train_set.add_y().add_y(source='noisy_y'))
if params.distributed:
from torch.utils.data.distributed import DistributedSampler
sampler = DistributedSampler(train_set, num_replicas=4)
self.sampler_a = sampler
else:
sampler = None
train_set = train_set.DataLoader(batch_size=params.batch_size,
num_workers=params.num_workers,
sampler=sampler,
shuffle=not params.distributed)
query_set = (DatasetBuilder(query_x,
query_y).add_x(transform=strong).add_y())
if params.distributed:
from torch.utils.data.distributed import DistributedSampler
sampler = DistributedSampler(train_set, num_replicas=4)
self.sampler_b = sampler
else:
sampler = None
query_set = query_set.DataLoader(batch_size=params.batch_size,
num_workers=params.num_workers,
sampler=sampler,
shuffle=not params.distributed)
val_dataloader = (
DatasetBuilder(eval_x, eval_y).add_x(transform=toTensor).add_y().
DataLoader(
batch_size=params.batch_size,
shuffle=
False, # do not shuffle # no shuffle for probe, so a batch is class balanced.(?)
num_workers=params.num_workers))
train_dataloader = DataBundler().add(train_set).cycle(
query_set).zip_mode()
test_dataloader = (DatasetBuilder(
test_x, test_y).add_x(transform=toTensor).add_y().DataLoader(
batch_size=params.batch_size,
num_workers=params.num_workers,
shuffle=True))
self.regist_databundler(train=train_dataloader,
eval=val_dataloader,
test=test_dataloader)
self.to(self.device)
def datasets(self, params: NoisyParams):
self.rnd.mark('fix_noisy')
params.noisy_type = params.default('symmetric', True)
params.noisy_ratio = params.default(0.2, True)
dataset_fn = datasets[params.dataset]
test_x, test_y = dataset_fn(False)
train_x, train_y = dataset_fn(True)
train_ids, val_ids = splits.train_val_split(train_y, val_size=5000)
train_x, val_x = train_x[train_ids], train_x[val_ids]
train_y, val_y = train_y[train_ids], train_y[val_ids]
mean, std = norm_val.get(params.dataset, [None, None])
toTensor = ToNormTensor(mean, std)
weak = Weak(mean, std)
strong = Strong(mean, std)
if params.noisy_type == 'asymmetric':
from data.noisy import asymmetric_noisy
noisy_y = asymmetric_noisy(train_y,
params.noisy_ratio,
n_classes=params.n_classes)
elif params.noisy_type == 'symmetric':
from data.noisy import symmetric_noisy
noisy_y = symmetric_noisy(train_y,
params.noisy_ratio,
n_classes=params.n_classes)
else:
assert False
clean_mask = train_y == noisy_y
self.logger.info('noisy acc = {}'.format((train_y == noisy_y).mean()))
self.rnd.shuffle()
train_set = (DatasetBuilder(train_x, train_y).add_labels(
noisy_y, 'noisy_y').toggle_id().add_x(transform=weak).add_x(
transform=strong).add_y().add_y(source='noisy_y'))
from thextra.noisy_sampler import NoisySampler
sampler = None
if params.order_sampler:
sampler = NoisySampler(train_set, clean_mask)
self.train_size = len(train_set)
train_dataloader = train_set.DataLoader(batch_size=params.batch_size,
num_workers=params.num_workers,
drop_last=True,
sampler=sampler,
shuffle=True)
val_dataloader = (DatasetBuilder(
val_x, val_y).add_x(transform=toTensor).add_y().DataLoader(
batch_size=params.batch_size,
num_workers=params.num_workers,
shuffle=True))
test_dataloader = (DatasetBuilder(
test_x, test_y).add_x(transform=toTensor).add_y().DataLoader(
batch_size=params.batch_size,
num_workers=params.num_workers,
shuffle=True))
self.regist_databundler(train=train_dataloader,
eval=val_dataloader,
test=test_dataloader)
self.to(self.device)
def datasets(self, params: EvalParams):
params.noisy_type = params.default('symmetric', True)
params.noisy_ratio = params.default(0.2, True)
from data.constant import norm_val
mean, std = norm_val.get(params.dataset, [None, None])
from data.transforms import ToNormTensor
toTensor = ToNormTensor(mean, std)
from data.transforms import Weak
weak = Weak(mean, std)
from data.transforms import Strong
dataset_fn = datasets.datasets[params.dataset]
train_x, train_y = dataset_fn(True)
train_y = np.array(train_y)
from thexp import DatasetBuilder
from data.noisy import symmetric_noisy
noisy_y = symmetric_noisy(train_y,
params.noisy_ratio,
n_classes=params.n_classes)
clean_mask = (train_y == noisy_y)
noisy_mask = np.logical_not(clean_mask)
if params.eval_mode in ['full', 'same_epoch', 'same_acc']:
first_x, first_y = train_x[clean_mask], noisy_y[clean_mask]
first_true_y = train_y[clean_mask]
elif params.eval_mode in ['mix', 'raw', 'direct']:
first_x, first_y = train_x, noisy_y
first_true_y = train_y
else:
assert False
second_x, second_true_y = train_x[noisy_mask], train_y[noisy_mask]
second_y = noisy_y[noisy_mask]
self.logger.info('noisy acc = {}'.format(
(first_true_y == first_y).mean()))
self.logger.info('noisy acc = {}'.format(
(second_true_y == second_y).mean()))
self.rnd.shuffle()
first_set = (DatasetBuilder(first_x, first_true_y).add_labels(
first_y, 'noisy_y').toggle_id().add_x(
transform=weak).add_y().add_y(source='noisy_y'))
noisy_set = (DatasetBuilder(second_x, second_true_y).add_labels(
second_y, 'noisy_y').toggle_id().add_x(
transform=weak).add_y().add_y(source='noisy_y'))
first_dataloader = first_set.DataLoader(batch_size=params.batch_size,
num_workers=params.num_workers,
drop_last=True,
shuffle=True)
self.second_dataloader = noisy_set.DataLoader(
batch_size=params.batch_size,
num_workers=params.num_workers,
drop_last=True,
shuffle=True)
self.second = False
self.regist_databundler(train=first_dataloader)
self.to(self.device)
if params.eval_mode == 'direct':
self.change_dataset()
def change_dataset(self):
"""
根据当前的 filter_mem，按 thresh 将其分为有监督和无监督
:return:
"""
from data.constant import norm_val
from data.transforms import ToNormTensor, Weak, Strong
train_x, train_y, noisy_y = self.train_set
filter_prob = self.filter_mem.cpu().numpy()
clean_mask = filter_prob > 0.5
self.logger.info('sup size', clean_mask.sum())
if clean_mask.all() or not np.logical_not(clean_mask).any():
return
clean_ids = np.where(clean_mask)[0]
noisy_ids = np.where(np.logical_not(clean_mask))[0]
mean, std = norm_val.get(params.dataset, [None, None])
weak = Weak(mean, std)
strong = Strong(mean, std)
supervised_dataloader = (
DatasetBuilder(train_x, train_y)
.add_labels(noisy_y, source_name='ny')
.toggle_id()
.add_x(strong)
.add_y()
.add_y(source='ny')
.subset(clean_ids)
.DataLoader(params.batch_size // 2,
shuffle=True,
num_workers=0,
drop_last=True)
)
unsupervised_dataloader = (
DatasetBuilder(train_x, train_y)
.add_labels(noisy_y, source_name='ny')
.add_labels(filter_prob, source_name='nprob')
.toggle_id()
.add_x(strong)
.add_x(strong)
.add_y()
.add_y(source='ny')
.add_y(source='nprob')
.subset(noisy_ids)
.DataLoader(params.batch_size // 2,
shuffle=True,
num_workers=0,
drop_last=True)
)
if len(supervised_dataloader) > len(unsupervised_dataloader):
train_dataloader = (
DataBundler()
.add(supervised_dataloader)
.cycle(unsupervised_dataloader)
)
else:
train_dataloader = (
DataBundler()
.cycle(supervised_dataloader)
.add(unsupervised_dataloader)
)
if len(unsupervised_dataloader) == 0 or len(supervised_dataloader) == 0:
self.ssl_dataloader = None
return
self.ssl_dataloader = train_dataloader.zip_mode().to(self.device)
self.logger.info('ssl loader size', train_dataloader)
self.ssl_loaderiter = iter(self.ssl_dataloader)