def __init__(self, name, split):
self.name = name
self.split = split
self.path = os.path.join(settings.checkpoints,
name + '-split_{}'.format(split))
self.net = RefineNet(SCSENoPoolResNextBase(se_resnext101_32x4d()),
num_features=128,
classifier=lambda c: RefineNetUpsampleClassifier(
c, scale_factor=2),
block=SCSERefineNetBlock)
self.tta = [
tta.Pipeline([tta.Pad((13, 14, 13, 14))]),
tta.Pipeline([tta.Pad((13, 14, 13, 14)),
tta.Flip()])
]
def __init__(self, name, split):
self.name = name
self.split = split
self.path = os.path.join(settings.checkpoints, name + '-split_{}'.format(split))
self.net = AuxDualHypercolumnCatRefineNet(
SCSENoPoolResNextBase(se_resnext50_32x4d()),
num_features=128,
classifier=lambda c: SmallDropoutRefineNetUpsampleClassifier(2*c, scale_factor=2, dropout=0.1),
block=SCSERefineNetBlock,
crp=[IdentityCRP, CRP, CRP, CRP]
)
self.tta = [
tta.Pipeline([tta.Pad((13, 14, 13, 14))]),
tta.Pipeline([tta.Pad((13, 14, 13, 14)), tta.Flip()])
]
def __init__(self, name, split):
self.name = name
self.split = split
self.path = os.path.join(settings.checkpoints, name + '-split_{}'.format(split))
self.net = DualHypercolumnCatRefineNet(
NoPoolDPN92Base(dpn92()),
num_features=128,
block_multiplier=1,
num_features_base=[256 + 80, 512 + 192, 1024 + 528, 2048 + 640],
classifier=lambda c: SmallDropoutRefineNetUpsampleClassifier(2*128, scale_factor=2),
)
self.tta = [
tta.Pipeline([tta.Pad((13, 14, 13, 14))]),
tta.Pipeline([tta.Pad((13, 14, 13, 14)), tta.Flip()])
]
def __init__(self, name, split):
self.name = name
self.split = split
self.path = os.path.join(settings.checkpoints,
name + '-split_{}'.format(split))
self.net = RefineNet(NoPoolDPN107Base(dpn107()),
num_features=128,
block_multiplier=1,
num_features_base=[376, 1152, 2432, 2048 + 640],
classifier=lambda c: RefineNetUpsampleClassifier(
c, scale_factor=2))
self.tta = [
tta.Pipeline([tta.Pad((13, 14, 13, 14))]),
tta.Pipeline([tta.Pad((13, 14, 13, 14)),
tta.Flip()])
]
def __init__(self, name, split):
self.name = name
self.split = split
self.path = os.path.join(settings.checkpoints, name + '-split_{}'.format(split))
self.net = DualHypercolumnCatRefineNet(
SCSENoPoolResNextBase(se_resnet152()),
num_features=128,
classifier=lambda c: RefineNetUpsampleClassifier(2*c, scale_factor=2),
block=SCSERefineNetBlock
)
self.tta = [
tta.Pipeline([tta.Pad((13, 14, 13, 14))]),
tta.Pipeline([tta.Pad((13, 14, 13, 14)), tta.Flip()])
]
self.test_predictions = utils.TestPredictions('ensemble-{}'.format(split)).load()
def __init__(self, name, split):
self.name = name
self.split = split
self.path = os.path.join(settings.checkpoints, name + '-split_{}'.format(split))
self.net = DualHypercolumnCatRefineNet(
NoPoolDPN107Base(dpn107()),
num_features=128,
block_multiplier=1,
num_features_base=[376, 1152, 2432, 2048 + 640],
classifier=lambda c: SmallDropoutRefineNetUpsampleClassifier(2 * 128, scale_factor=2),
)
self.tta = [
tta.Pipeline([tta.Pad((13, 14, 13, 14))]),
tta.Pipeline([tta.Pad((13, 14, 13, 14)), tta.Flip()])
]
self.test_predictions = utils.TestPredictions('ensemble_top_6_postprocessed-split_{}'.format(split)).load()
def __init__(self, name, split):
self.name = name
self.split = split
self.path = os.path.join(settings.checkpoints,
name + '-split_{}'.format(split))
self.net = RefineNet(NoPoolDPN98Base(dpn98()),
num_features=128,
block_multiplier=1,
num_features_base=[336, 768, 1728, 2688],
classifier=lambda c: RefineNetUpsampleClassifier(
c, scale_factor=2))
self.optimizer = Adam(self.net.parameters(),
lr=1e-4,
weight_decay=1e-4)
self.tta = [
tta.Pipeline([tta.Pad((13, 14, 13, 14))]),
tta.Pipeline([tta.Pad((13, 14, 13, 14)),
tta.Flip()])
]
self.batch_size = 16
def train(self, net, samples, optimizer, e):
alpha = 2 * max(0, ((50 - e) / 50))
criterion = losses.ELULovaszFocalWithLogitsLoss(alpha, 2 - alpha)
transforms = generator.TransformationsGenerator([
random.RandomFlipLr(),
random.RandomAffine(
image_size=101,
translation=lambda rs: (rs.randint(-20, 20), rs.randint(-20, 20)),
scale=lambda rs: (rs.uniform(0.85, 1.15), 1),
**utils.transformations_options
)
])
dataset = datasets.ImageDataset(samples, settings.train, transforms)
dataloader = DataLoader(
dataset,
num_workers=10,
batch_size=16,
shuffle=True
)
average_meter_train = meters.AverageMeter()
with tqdm(total=len(dataloader), leave=False, ascii=True) as pbar, torch.enable_grad():
net.train()
padding = tta.Pad((13, 14, 13, 14))
for images, masks_targets in dataloader:
masks_targets = masks_targets.to(gpu)
masks_predictions = padding.transform_backward(net(padding.transform_forward(images))).contiguous()
loss = criterion(masks_predictions, masks_targets)
loss.backward()
optimizer.step()
optimizer.zero_grad()
average_meter_train.add('loss', loss.item())
self.update_pbar(
torch.sigmoid(masks_predictions),
masks_targets,
pbar,
average_meter_train,
'Training epoch {}'.format(e)
)
train_stats = {'train_' + k: v for k, v in average_meter_train.get_all().items()}
return train_stats
def train(self, net, samples, optimizer, e):
alpha = 2 * max(0, ((50 - e) / 50))
criterion = losses.ELULovaszFocalWithLogitsLoss(alpha, 2 - alpha)
transforms = generator.TransformationsGenerator([
random.RandomFlipLr(),
random.RandomAffine(image_size=101,
translation=lambda rs:
(rs.randint(-20, 20), rs.randint(-20, 20)),
scale=lambda rs: (rs.uniform(0.85, 1.15), 1),
**utils.transformations_options)
])
samples_aux = list(
set(samples).intersection(set(utils.get_aux_samples())))
dataset_aux = datasets.ImageDataset(samples_aux, settings.train,
transforms)
dataset_pseudo = datasets.SemiSupervisedImageDataset(
samples_test,
settings.test,
transforms,
size=len(samples_test),
test_predictions=self.test_predictions,
momentum=0.0)
dataset = datasets.ImageDataset(samples, settings.train, transforms)
weight_train = len(dataset_pseudo) / len(dataset) * 2
weight_aux = weight_train / 2
weights = [weight_train] * len(dataset) + [weight_aux] * len(
dataset_aux) + [1] * len(dataset_pseudo)
dataloader = DataLoader(
ConcatDataset([dataset, dataset_aux, dataset_pseudo]),
num_workers=10,
batch_size=16,
sampler=WeightedRandomSampler(weights=weights, num_samples=3200))
average_meter_train = meters.AverageMeter()
with tqdm(total=len(dataloader), leave=False,
ascii=True) as pbar, torch.enable_grad():
net.train()
padding = tta.Pad((13, 14, 13, 14))
for images, masks_targets in dataloader:
masks_targets = masks_targets.to(gpu)
masks_predictions = padding.transform_backward(
net(padding.transform_forward(images))).contiguous()
loss = criterion(masks_predictions, masks_targets)
loss.backward()
optimizer.step()
optimizer.zero_grad()
average_meter_train.add('loss', loss.item())
self.update_pbar(torch.sigmoid(masks_predictions),
masks_targets, pbar, average_meter_train,
'Training epoch {}'.format(e))
train_stats = {
'train_' + k: v
for k, v in average_meter_train.get_all().items()
}
return train_stats