def reload_obj(self, model_group_name):
if model_group_name == self.model_group_name:
return
if hasattr(self, 'obj_feas'):
del self.obj_feas
if hasattr(self, 'objects_vocab'):
del self.objects_vocab
# load object features
obj_fea_name = None
if model_group_name == 'avg_label':
obj_fea_name = 'class-fea'
elif model_group_name == 'onehot_label':
obj_fea_name = 'class'
elif model_group_name == 'prob_label':
obj_fea_name = 'class-prob'
if obj_fea_name:
self.obj_feas = []
for data_split in self.splits:
if data_split == 'vg':
continue
obj_fname = get_feature_path(data_split, obj_fea_name)
self.obj_feas.append(np.load(obj_fname))
if len(self.obj_feas) > 0:
self.obj_feas = np.vstack(self.obj_feas)
# load object labels
if model_group_name in ('onehot_label', 'prob_label'):
with open('data/objects_vocab.txt') as f:
self.objects_vocab = f.read().splitlines()
self.objects_vocab = ['__no_objects__'] + self.objects_vocab
self.model_group_name = model_group_name
def __init__(self, split, model_group_name):
self.codebook = json.load(open('{}/data.json'.format(cfg.DATA_DIR)))
data = h5py.File('{}/data.h5'.format(
cfg.DATA_DIR))['/{}'.format(split)]
self.img_pos = data['img_pos'].value
self.que = data['que'].value
self.que_id = data['que_id'].value
if 'ans' in data:
self.ans = data['ans'].value
if cfg.SOFT_LOSS:
self.ans = self.ans.astype(np.float32)
# load image features
self.splits = cfg[split.upper()].SPLITS
self.img_feas = []
for data_split in self.splits:
if data_split == 'vg':
continue
fea_fname = get_feature_path(data_split, 'feature')
if cfg.LOAD_ALL_DATA:
img_fea = np.load(fea_fname)
else:
img_fea = open_memmap(fea_fname, dtype='float32')
self.img_feas.append(img_fea)
self.img_cnts = list(map(len, self.img_feas))
self.model_group_name = None
self.reload_obj(model_group_name)
def get_img_pos(splits, data_ids):
fea_ids = []
for split in splits:
if split == 'vg':
continue
fea_ids.append(np.load(get_feature_path(split, 'id')))
if len(fea_ids) > 0:
fea_ids = np.hstack(fea_ids)
id_to_pos = {img_id: i for i, img_id in enumerate(fea_ids)}
data_poses = [id_to_pos[img_id] for img_id in data_ids]
return np.array(data_poses, dtype='int64')
def _load_box(self):
data = h5py.File('{}/data.h5'.format(cfg.DATA_DIR))['/test']
img_pos = data['img_pos'].value
box_dict_fname = get_feature_path(self._split, 'box')
return np.load(box_dict_fname)[img_pos]