def get_eval_loader(conf):
eval_transform = trans.Compose([
trans.ToPILImage(),
trans.RandomResizedCrop(size=tuple(conf.input_size), scale=(0.9, 1.1)),
trans.RandomHorizontalFlip(),
trans.ToTensor(),
trans.Normalize(mean=127.5, std=128)
])
root_path = '{}/{}'.format(conf.eval_root_path, conf.patch_info)
evalset = DatasetFolderFT(root_path, eval_transform, None, conf.ft_width,
conf.ft_height)
# evalset = CelebASpoofDataset(conf.val_annotation_path,
# root_prefix=conf.prefix,
# transform=eval_transform,
# ft_width=conf.ft_width,
# ft_height=conf.ft_height)
# evalset = DatasetFolderFT(root_path, eval_transform,
# None, conf.ft_width, conf.ft_height)
eval_loader = DataLoader(evalset,
batch_size=conf.batch_size,
shuffle=True,
pin_memory=True,
num_workers=16)
return eval_loader
def get_train_loader(conf):
train_transform = trans.Compose([
trans.ToPILImage(),
trans.RandomResizedCrop(size=tuple(conf.input_size), scale=(0.9, 1.1)),
trans.ColorJitter(brightness=0.4,
contrast=0.4,
saturation=0.4,
hue=0.1),
trans.RandomRotation(10),
trans.RandomHorizontalFlip(),
trans.ToTensor(),
trans.Normalize(mean=127.5, std=128),
])
root_path = '{}/{}'.format(conf.train_root_path, conf.patch_info)
trainset = DatasetFolderFT(root_path, train_transform, None, conf.ft_width,
conf.ft_height)
# trainset = CelebASpoofDataset(conf.train_annotation_path,
# root_prefix=conf.prefix,
# transform=train_transform,
# ft_width=conf.ft_width,
# ft_height=conf.ft_height)
train_loader = DataLoader(trainset,
batch_size=conf.batch_size,
shuffle=True,
pin_memory=True,
num_workers=16)
return train_loader
def predict(self, img):
test_transform = trans.Compose([
trans.ToTensor(),
])
img = test_transform(img)
img = img.unsqueeze(0).to(self.device)
self.model.eval()
with torch.no_grad():
result = self.model.forward(img)
result = F.softmax(result).cpu().numpy()
return result
def get_train_loader(conf):
train_transform = trans.Compose([
trans.ToPILImage(),
trans.RandomResizedCrop(size=tuple(conf.input_size), scale=(0.9, 1.1)),
trans.ColorJitter(brightness=0.4,
contrast=0.4,
saturation=0.4,
hue=0.1),
trans.RandomRotation(10),
trans.RandomHorizontalFlip(),
trans.ToTensor()
])
root_path = '{}/{}'.format(conf.train_root_path, conf.patch_info)
dataset = DatasetFolderFT(root_path, train_transform, None, conf.ft_width,
conf.ft_height)
# Creating data indices for training and validation splits:
shuffle_dataset = True
random_seed = 42
dataset_size = len(dataset)
indices = list(range(dataset_size))
split = int(np.floor(conf.val_size * dataset_size))
if shuffle_dataset:
np.random.seed(random_seed)
np.random.shuffle(indices)
train_indices, val_indices = indices[split:], indices[:split]
# Creating PT data samplers and loaders:
train_sampler = SubsetRandomSampler(train_indices)
valid_sampler = SubsetRandomSampler(val_indices)
train_loader = DataLoader(dataset,
batch_size=conf.batch_size,
sampler=train_sampler,
num_workers=16)
val_loader = DataLoader(dataset,
batch_size=conf.batch_size,
sampler=valid_sampler,
num_workers=16)
return train_loader, val_loader