def get_tusimple(params):
# augmentation
flip = Flip()
translate = Translate()
rotate = Rotate()
add_noise = AddGaussianNoise()
change_intensity = ChangeIntensity()
resize = Resize(rows=256, cols=512)
norm_to_1 = NormalizeInstensity()
whc_to_cwh = TransposeNumpyArray((2, 0, 1))
train_dataset = DatasetTusimple(
root_path=params.train_root_url,
json_files=params.train_json_file,
transform=transforms.Compose([
flip, translate, rotate, add_noise, change_intensity, resize,
norm_to_1, whc_to_cwh
]),
)
val_dataset = DatasetTusimple(
params.val_root_url,
params.val_json_file,
transform=transforms.Compose([resize, norm_to_1, whc_to_cwh]),
)
return train_dataset, val_dataset
def __init__(self, settings):
super(SRTrainer, self).__init__(settings)
self.scale = settings.scale
self.criterion = ComLoss(settings.iqa_model_path,
settings.__dict__.get('weights'),
settings.__dict__.get('feat_names'),
settings.alpha, settings.iqa_patch_size,
settings.criterion)
if hasattr(self.criterion, 'iqa_loss'):
# For saving cost
self.criterion.iqa_loss.freeze()
self.model = build_model(ARCH, scale=self.scale)
self.dataset = get_dataset(DATASET)
if self.phase == 'train':
self.train_loader = torch.utils.data.DataLoader(
self.dataset(self.data_dir,
'train',
self.scale,
list_dir=self.list_dir,
transform=Compose(
MSCrop(self.scale, settings.patch_size),
Flip()),
repeats=settings.reproduce),
batch_size=self.
batch_size, #max(self.batch_size//settings.reproduce, 1),
shuffle=True,
num_workers=settings.num_workers,
pin_memory=True,
drop_last=True)
self.val_loader = self.dataset(self.data_dir,
'val',
self.scale,
subset=settings.subset,
list_dir=self.list_dir)
if not self.val_loader.lr_avai:
self.logger.warning(
"warning: the low-resolution sources are not available")
self.optimizer = torch.optim.Adam(self.model.parameters(),
betas=(0.9, 0.999),
lr=self.lr,
weight_decay=settings.weight_decay)
# self.optimizer = torch.optim.RMSprop(
# self.model.parameters(),
# lr=self.lr,
# alpha=0.9,
# weight_decay=settings.weight_decay
# )
self.logger.dump(self.model) # Log the architecture
def __init__(self, ):
print("usage examples:")
print("python -m dataset.culane.test sample")
print("python -m dataset.culane.test batch")
print("python -m dataset.culane.test batch shuffle=False")
flip = Flip(1.0)
translate = Translate(1.0)
rotate = Rotate(1.0)
add_noise = AddGaussianNoise(1.0)
change_intensity = ChangeIntensity(1.0)
resize = Resize(rows=256, cols=512)
hwc_to_chw = TransposeNumpyArray((2, 0, 1))
norm_to_1 = NormalizeInstensity()
self.train_dataset = DatasetCollections(transform=transforms.Compose([
flip, translate, rotate, add_noise, change_intensity, resize,
norm_to_1, hwc_to_chw
]), )
def __init__(self,):
flip = Flip(1.0)
translate = Translate(1.0)
rotate = Rotate(1.0)
add_noise = AddGaussianNoise(1.0)
change_intensity = ChangeIntensity(1.0)
resize = Resize(rows=256, cols=512)
hwc_to_chw = TransposeNumpyArray((2, 0, 1))
norm_to_1 = NormalizeInstensity()
json_file = ['label_data_0313.json', 'label_data_0531.json', 'label_data_0601.json']
self.train_dataset = DatasetTusimple(root_path="/media/zzhou/data-tusimple/lane_detection/train_set/",
json_files=json_file,
transform=transforms.Compose([flip,
translate,
rotate,
add_noise,
change_intensity,
resize,
norm_to_1,
hwc_to_chw]),)
def __init__(self,):
print("usage examples:")
print("python -m dataset.bdd100k.test sample")
print("python -m dataset.bdd100k.test batch")
print("python -m dataset.bdd100k.test batch shuffle=False")
flip = Flip(1.0)
translate = Translate(1.0)
rotate = Rotate(1.0)
add_noise = AddGaussianNoise(1.0)
change_intensity = ChangeIntensity(1.0)
resize = Resize(rows=256, cols=512)
hwc_to_chw = TransposeNumpyArray((2, 0, 1))
norm_to_1 = NormalizeInstensity()
self.train_dataset = DatasetBDD100K(root_path="/media/zzhou/data-BDD100K/bdd100k/",
json_files="labels/bdd100k_labels_images_train.json",
transform=transforms.Compose([flip,
translate,
rotate,
add_noise,
change_intensity,
resize,
norm_to_1,
hwc_to_chw]), )