def gen_bbox_transform(img_size):
transform = T.Compose([
T.NormalizeBbox(),
T.Resize(img_size),
T.DenormalizeBbox(),
])
def fun(img, bbox):
_, bbox = transform(img, bbox)
return torch.from_numpy(bbox)
return fun
def build_concept_quantization_clevr_dataset(args, configs, image_root, scenes_json):
import jactorch.transforms.bbox as T
image_transform = T.Compose([
T.NormalizeBbox(),
T.Resize(configs.data.image_size),
T.DenormalizeBbox(),
T.ToTensor(),
T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
from nscl.datasets.datasets import ConceptQuantizationDataset
dataset = ConceptQuantizationDataset(scenes_json, image_root=image_root, image_transform=image_transform)
return dataset
def build_clevr_dataset(args, configs, image_root, scenes_json, questions_json):
import jactorch.transforms.bbox as T
image_transform = T.Compose([
T.NormalizeBbox(),
T.Resize(configs.data.image_size),
T.DenormalizeBbox(),
T.ToTensor(),
T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
from nscl.datasets.datasets import NSCLDataset
dataset = NSCLDataset(
scenes_json, questions_json,
image_root=image_root, image_transform=image_transform,
vocab_json=args.data_vocab_json
)
return dataset
def __init__(self, args, phase):
self.args = args
self.phase = phase
self.loader = default_loader
self.data_dir = args.data_dir
self.label_dir = args.label_dir
self.prp_dir = args.prp_dir
self.ann_dir = args.ann_dir
self.tube_dir = args.tube_dir
self.valid_idx_lst = 'valid_idx_' + self.phase + '.txt'
self.H = 100
self.W = 150
self.bbox_size = 24
ratio = self.args.train_valid_ratio
n_train = round(self.args.n_rollout * ratio)
if phase == 'train':
self.st_idx = 0
self.n_rollout = n_train
elif phase == 'valid':
self.st_idx = n_train
self.n_rollout = self.args.n_rollout - n_train
else:
raise AssertionError("Unknown phase")
if self.args.gen_valid_idx:
if self.args.visualize_flag == 1:
self.gen_predict_input()
elif self.args.version == 'v3':
self.gen_valid_idx_from_tube_info_v3()
else:
self.gen_valid_idx_from_tube_info()
else:
self.read_valid_idx()
self.img_transform = T.Compose([
T.Resize(self.args.img_size),
T.ToTensor(),
T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
self.W = 480
self.H = 320
def build_concept_retrieval_clevrer_dataset(args, configs, program, image_root,
scenes_json):
import jactorch.transforms.bbox as T
image_transform = T.Compose([
T.NormalizeBbox(),
T.Resize(configs.data.image_size),
T.DenormalizeBbox(),
T.ToTensor(),
T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
])
from nscl.datasets.datasets import ConceptRetrievalDataset
dataset = ConceptRetrievalDataset(program,
scenes_json,
image_root=image_root,
image_transform=image_transform)
return dataset
