def loading_data():
mean_std = cfg_data.MEAN_STD
log_para = cfg_data.LOG_PARA
factor = cfg_data.LABEL_FACTOR
train_main_transform = own_transforms.Compose(
[own_transforms.RandomHorizontallyFlip()])
img_transform = standard_transforms.Compose([
standard_transforms.ToTensor(),
standard_transforms.Normalize(*mean_std)
])
gt_transform = standard_transforms.Compose([
own_transforms.GTScaleDown(factor),
own_transforms.LabelNormalize(log_para)
])
restore_transform = standard_transforms.Compose([
own_transforms.DeNormalize(*mean_std),
standard_transforms.ToPILImage()
])
train_set = SHHA(cfg_data.DATA_PATH + '/train_data',
mode='train',
preload=True,
main_transform=train_main_transform,
img_transform=img_transform,
gt_transform=gt_transform)
train_loader = None
if cfg_data.TRAIN_BATCH_SIZE == 1:
train_loader = DataLoader(train_set,
batch_size=1,
num_workers=0,
collate_fn=SHHA_raw_collate,
shuffle=True,
drop_last=True)
elif cfg_data.TRAIN_BATCH_SIZE > 1:
train_loader = DataLoader(train_set,
batch_size=cfg_data.TRAIN_BATCH_SIZE,
num_workers=8,
collate_fn=SHHA_crop_collate,
shuffle=True,
drop_last=True)
val_set = SHHA(cfg_data.DATA_PATH + '/test_data',
mode='test',
preload=True,
main_transform=None,
img_transform=img_transform,
gt_transform=gt_transform)
val_loader = DataLoader(val_set,
batch_size=cfg_data.VAL_BATCH_SIZE,
num_workers=0,
collate_fn=SHHA_raw_collate,
shuffle=True,
drop_last=False)
return train_loader, val_loader, restore_transform
def data_transforms(cfg_data):
mean_std = cfg_data.MEAN_STD
log_para = cfg_data.LOG_PARA
# train and val main data transformations
if cfg_data.DATASET == 'City':
train_main_transform = own_transforms.Compose([
own_transforms.RandomHorizontallyFlip()
])
val_main_transform = None
elif cfg_data.DATASET in ['FDST', 'PETS']:
train_main_transform = standard_transforms.Compose([
own_transforms.FreeScale(cfg_data.TRAIN_SIZE),
])
val_main_transform = standard_transforms.Compose([
own_transforms.FreeScale(cfg_data.TRAIN_SIZE),
])
else:
train_main_transform = own_transforms.Compose([
own_transforms.RandomCrop(cfg_data.TRAIN_SIZE),
own_transforms.RandomHorizontallyFlip()
])
val_main_transform = None
# image and gt transformations
if cfg_data.DATASET == 'FDST':
gt_transform = standard_transforms.Compose([
own_transforms.GTScaleDown(cfg_data.TRAIN_DOWNRATE),
own_transforms.LabelNormalize(log_para)
])
else:
gt_transform = standard_transforms.Compose([
own_transforms.LabelNormalize(log_para)
])
img_transform = standard_transforms.Compose([
standard_transforms.ToTensor(),
standard_transforms.Normalize(*mean_std)
])
restore_transform = standard_transforms.Compose([
own_transforms.DeNormalize(*mean_std),
standard_transforms.ToPILImage()
])
return train_main_transform, val_main_transform, img_transform, gt_transform, restore_transform
def loading_data():
# shanghai Tech A
mean_std = cfg.DATA.MEAN_STD
log_para = cfg.DATA.LOG_PARA
factor = cfg.DATA.LABEL_FACTOR
train_main_transform = own_transforms.Compose([
own_transforms.RandomCrop(cfg.TRAIN.INPUT_SIZE),
own_transforms.RandomHorizontallyFlip()
])
val_main_transform = None
img_transform = standard_transforms.Compose([
standard_transforms.ToTensor(),
standard_transforms.Normalize(*mean_std)
])
gt_transform = standard_transforms.Compose([
own_transforms.GTScaleDown(factor),
own_transforms.LabelNormalize(log_para)
])
restore_transform = standard_transforms.Compose([
own_transforms.DeNormalize(*mean_std),
standard_transforms.ToPILImage()
])
train_set = UCF_QNRF(cfg.DATA.DATA_PATH + '/train',
'train',
main_transform=train_main_transform,
img_transform=img_transform,
gt_transform=gt_transform)
train_loader = DataLoader(train_set,
batch_size=cfg.TRAIN.BATCH_SIZE,
num_workers=8,
shuffle=True,
drop_last=True)
val_set = UCF_QNRF(cfg.DATA.DATA_PATH + '/test',
'test',
main_transform=val_main_transform,
img_transform=img_transform,
gt_transform=gt_transform)
val_loader = DataLoader(val_set,
batch_size=cfg.VAL.BATCH_SIZE,
num_workers=8,
shuffle=True,
drop_last=False)
return train_set, train_loader, val_set, val_loader, restore_transform
def loading_data():
mean_std = cfg_data.MEAN_STD
log_para = cfg_data.LOG_PARA
factor = cfg_data.LABEL_FACTOR
train_main_transform = own_transforms.Compose([
own_transforms.RandomHorizontallyFlip()
])
img_transform = standard_transforms.Compose([
standard_transforms.ToTensor(),
standard_transforms.Normalize(*mean_std)
])
gt_transform = standard_transforms.Compose([
own_transforms.GTScaleDown(factor),
own_transforms.LabelNormalize(log_para)
])
restore_transform = standard_transforms.Compose([
own_transforms.DeNormalize(*mean_std),
standard_transforms.ToPILImage()
])
if cfg_data.IS_CROSS_SCENE:
train_set = AC(img_path=cfg_data.IMAGE_PATH, den_path=cfg_data.DENSITY_PATH + '/cross_scene_train',
aud_path=cfg_data.AUDIO_PATH,
mode='train', main_transform=train_main_transform, img_transform=img_transform,
gt_transform=gt_transform, is_noise=cfg_data.IS_NOISE, brightness_decay=cfg_data.BRIGHTNESS,
noise_sigma=cfg_data.NOISE_SIGMA, longest_side=cfg_data.LONGEST_SIDE
)
else:
train_set = AC(img_path=cfg_data.IMAGE_PATH, den_path=cfg_data.DENSITY_PATH + '/train',
aud_path=cfg_data.AUDIO_PATH,
mode='train', main_transform=train_main_transform, img_transform=img_transform,
gt_transform=gt_transform, is_noise=cfg_data.IS_NOISE, brightness_decay=cfg_data.BRIGHTNESS,
noise_sigma=cfg_data.NOISE_SIGMA, longest_side=cfg_data.LONGEST_SIDE,
black_area_ratio=cfg_data.BLACK_AREA_RATIO, is_random=cfg_data.IS_RANDOM, is_denoise=cfg_data.IS_DENOISE
)
train_loader = None
if cfg_data.TRAIN_BATCH_SIZE == 1:
train_loader = DataLoader(train_set, batch_size=1, num_workers=8, shuffle=True, drop_last=True)
elif cfg_data.TRAIN_BATCH_SIZE > 1:
train_loader = DataLoader(train_set, batch_size=cfg_data.TRAIN_BATCH_SIZE, num_workers=8,
collate_fn=AC_collate, shuffle=True, drop_last=True)
if cfg_data.IS_CROSS_SCENE:
val_set = AC(img_path=cfg_data.IMAGE_PATH, den_path=cfg_data.DENSITY_PATH + '/cross_scene_val',
aud_path=cfg_data.AUDIO_PATH,
mode='val', main_transform=None, img_transform=img_transform, gt_transform=gt_transform,
is_noise=cfg_data.IS_NOISE, brightness_decay=cfg_data.BRIGHTNESS,
noise_sigma=cfg_data.NOISE_SIGMA, longest_side=cfg_data.LONGEST_SIDE
)
else:
val_set = AC(img_path=cfg_data.IMAGE_PATH, den_path=cfg_data.DENSITY_PATH + '/val',
aud_path=cfg_data.AUDIO_PATH,
mode='val', main_transform=None, img_transform=img_transform, gt_transform=gt_transform,
is_noise=cfg_data.IS_NOISE, brightness_decay=cfg_data.BRIGHTNESS,
noise_sigma=cfg_data.NOISE_SIGMA, longest_side=cfg_data.LONGEST_SIDE,
black_area_ratio=cfg_data.BLACK_AREA_RATIO, is_random=cfg_data.IS_RANDOM, is_denoise=cfg_data.IS_DENOISE
)
val_loader = DataLoader(val_set, batch_size=cfg_data.VAL_BATCH_SIZE, num_workers=1, shuffle=False, drop_last=False)
if cfg_data.IS_CROSS_SCENE:
test_set = AC(img_path=cfg_data.IMAGE_PATH, den_path=cfg_data.DENSITY_PATH + '/cross_scene_test',
aud_path=cfg_data.AUDIO_PATH,
mode='test', main_transform=None, img_transform=img_transform, gt_transform=gt_transform,
is_noise=cfg_data.IS_NOISE, brightness_decay=cfg_data.BRIGHTNESS,
noise_sigma=cfg_data.NOISE_SIGMA, longest_side=cfg_data.LONGEST_SIDE
)
else:
test_set = AC(img_path=cfg_data.IMAGE_PATH, den_path=cfg_data.DENSITY_PATH + '/test',
aud_path=cfg_data.AUDIO_PATH,
mode='test', main_transform=None, img_transform=img_transform, gt_transform=gt_transform,
is_noise=cfg_data.IS_NOISE, brightness_decay=cfg_data.BRIGHTNESS,
noise_sigma=cfg_data.NOISE_SIGMA, longest_side=cfg_data.LONGEST_SIDE,
black_area_ratio=cfg_data.BLACK_AREA_RATIO, is_random=cfg_data.IS_RANDOM, is_denoise=cfg_data.IS_DENOISE
)
test_loader = DataLoader(test_set, batch_size=cfg_data.VAL_BATCH_SIZE, num_workers=1, shuffle=False, drop_last=False)
return train_loader, val_loader, test_loader, restore_transform