def __init__(self, cf, model):
self.cf = cf
self.model = model
# Compose preprocesing function for dataloaders
if self.cf.problem_type == 'detection':
self.img_preprocessing = standard_transforms.Compose([
Random_distort(self.cf),
preproces_input(self.cf),
standard_transforms.ToTensor()
])
self.test_img_preprocessing = standard_transforms.Compose(
[preproces_input(self.cf),
standard_transforms.ToTensor()])
self.train_transformation = ComposeObjDet(
[CropObjDet(self.cf),
RandomHorizontalFlipObjDet(self.cf)])
self.resize = ComposeResize([Resize(self.cf)])
else:
self.img_preprocessing = standard_transforms.Compose([
Random_distort(self.cf),
preproces_input(self.cf),
ToTensor()
])
self.test_img_preprocessing = standard_transforms.Compose(
[preproces_input(self.cf),
ToTensor()])
self.train_transformation = ComposeSemSeg(
[CropSegSem(self.cf),
RandomHorizontalFlipSegSem(self.cf)])
def build_valid(self, valid_samples, images_txt, gt_txt, resize_image, batch_size):
if self.cf.problem_type == 'segmentation':
self.loader_set = fromFileDatasetSegmentation(self.cf, images_txt, gt_txt,
valid_samples, resize_image,
preprocess=self.img_preprocessing, transform=None,
valid=True)
elif self.cf.problem_type == 'classification':
if self.cf.valid_path is not None:
self.loader_set = fromPathDatasetClassification(self.cf, self.cf.valid_path, resize_image,
preprocess=self.img_preprocessing, transform=None,
valid=True)
else:
self.loader_set = fromFileDatasetClassification(self.cf, images_txt, gt_txt,
valid_samples, resize_image,
preprocess=self.img_preprocessing, transform=None,
valid=True)
elif self.cf.problem_type == 'detection':
self.train_transformation = ComposeObjDet([Resize(self.cf)])
self.loader_set = fromFileDatasetDetection(self.cf, images_txt, gt_txt,
valid_samples, resize_image,
preprocess=self.img_preprocessing,
transform=None,
valid=True,
resize_process=self.resize)
self.loader = DataLoader(self.loader_set, batch_size=batch_size, num_workers=4)