def __init__(self, dataset_type, imdb_file_index, config, *args, **kwargs):
super().__init__("vqa2", dataset_type, config)
imdb_files = self.config.imdb_files
if dataset_type not in imdb_files:
raise ValueError(
"Dataset type {} is not present in "
"imdb_files of dataset config".format(dataset_type))
self.imdb_file = imdb_files[dataset_type][imdb_file_index]
self.imdb_file = self._get_absolute_path(self.imdb_file)
self.imdb = ImageDatabase(self.imdb_file)
self.kwargs = kwargs
self.image_depth_first = self.config.image_depth_first
self._should_fast_read = self.config.fast_read
self.use_ocr = self.config.use_ocr
self.use_ocr_info = self.config.use_ocr_info
self._use_features = False
if hasattr(self.config, "image_features"):
self._use_features = True
self.features_max_len = self.config.features_max_len
self._return_info = self.config.get("return_info", True)
all_image_feature_dirs = self.config.image_features[dataset_type]
curr_image_features_dir = all_image_feature_dirs[imdb_file_index]
curr_image_features_dir = curr_image_features_dir.split(",")
curr_image_features_dir = self._get_absolute_path(
curr_image_features_dir)
self.features_db = FeaturesDataset(
"coco",
directories=curr_image_features_dir,
depth_first=self.image_depth_first,
max_features=self.features_max_len,
fast_read=self._should_fast_read,
imdb=self.imdb,
return_info=self._return_info,
)
self.fast_dir = os.path.join(config.fast_dir, self._dataset_type)
self.fasted = set()
if not os.path.exists(self.fast_dir):
os.mkdir(self.fast_dir)
for sample in os.listdir(self.fast_dir):
self.fasted.add(int(sample[:-2]))
self.use_ocr = self.config.use_ocr
self.use_ocr_info = self.config.use_ocr_info
def __init__(self, dataset_type, imdb_file_index, config, *args, **kwargs):
super().__init__("vqa_introspect", dataset_type, config)
imdb_files = self.config.imdb_files
if dataset_type not in imdb_files:
raise ValueError(
"Dataset type {} is not present in "
"imdb_files of dataset config".format(dataset_type))
self.imdb_file = imdb_files[dataset_type][imdb_file_index]
self.imdb_file = self._get_absolute_path(self.imdb_file)
self.dataset = dataset_type
if dataset_type == 'train' or dataset_type == 'val':
if 'imdb_train2014.npy' in self.imdb_file or 'imdb_val2014.npy' in self.imdb_file or 'imdb_vqa2014_val_reasoning_questions' in self.imdb_file:
self.dataset = 'train_vqa'
elif 'train_introspect.npy' in self.imdb_file or 'train_binary.npy' in self.imdb_file:
self.dataset = 'train_introspect'
self.imdb = ImageDatabase(self.imdb_file)
self.kwargs = kwargs
self.image_depth_first = self.config.image_depth_first
self._should_fast_read = self.config.fast_read
self.use_ocr = self.config.use_ocr
self.use_ocr_info = self.config.use_ocr_info
self._use_features = False
if hasattr(self.config, "image_features"):
self._use_features = True
self.features_max_len = self.config.features_max_len
self._return_info = self.config.get("return_info", True)
all_image_feature_dirs = self.config.image_features[dataset_type]
curr_image_features_dir = all_image_feature_dirs[imdb_file_index]
curr_image_features_dir = curr_image_features_dir.split(",")
curr_image_features_dir = self._get_absolute_path(
curr_image_features_dir)
self.features_db = FeaturesDataset(
"coco",
directories=curr_image_features_dir,
depth_first=self.image_depth_first,
max_features=self.features_max_len,
fast_read=self._should_fast_read,
imdb=self.imdb,
return_info=self._return_info,
)
def __init__(self, dataset_type, imdb_file_index, config, *args, **kwargs):
super().__init__("vqa2", dataset_type, config)
imdb_files = self.config.imdb_files
if dataset_type not in imdb_files:
raise ValueError(
"Dataset type {} is not present in "
"imdb_files of dataset config".format(dataset_type))
self.imdb_file = imdb_files[dataset_type][imdb_file_index]
self.imdb_file = self._get_absolute_path(self.imdb_file)
self.imdb = ImageDatabase(self.imdb_file)
self.kwargs = kwargs
self.image_depth_first = self.config.image_depth_first
self._should_fast_read = self.config.fast_read
self.use_ocr = self.config.use_ocr
self.use_ocr_info = self.config.use_ocr_info
self._use_features = False
if hasattr(self.config, "image_features"):
self._use_features = True
self.features_max_len = self.config.features_max_len
all_image_feature_dirs = self.config.image_features[dataset_type]
curr_image_features_dir = all_image_feature_dirs[imdb_file_index]
curr_image_features_dir = curr_image_features_dir.split(",")
curr_image_features_dir = self._get_absolute_path(
curr_image_features_dir)
self.features_db = FeaturesDataset(
"coco",
directories=curr_image_features_dir,
depth_first=self.image_depth_first,
max_features=self.features_max_len,
fast_read=self._should_fast_read,
imdb=self.imdb,
)
def __init__(self, dataset_type, imdb_file_index, config, *args, **kwargs):
super().__init__("vqa2", dataset_type, config)
imdb_files = self.config.imdb_files
if dataset_type not in imdb_files:
raise ValueError(
"Dataset type {} is not present in "
"imdb_files of dataset config".format(dataset_type))
self.imdb_file = imdb_files[dataset_type][imdb_file_index]
if self.imdb_file.endswith(
"imdb_karpathy_test.npy"
) and self._global_config['imdb_file_content'] is not None:
if not self.imdb_file.startswith("data"):
self.imdb_file = "data/" + os.path.dirname(
self.imdb_file
) + "/imdb_" + self._global_config['imdb_file_content']
else:
self.imdb_file = os.path.dirname(
self.imdb_file
) + "/imdb_" + self._global_config['imdb_file_content']
imdb = np.zeros(2, dtype=object)
imdb[0] = {'metadata': 'coco'}
imdb[1] = {
'feature_path': self._global_config['imdb_file_content'],
'image_id': 0,
'reference_tokens': [['<s>']]
}
np.save(self.imdb_file, imdb)
print('imdb_files', imdb_files, flush=True)
if not self.imdb_file.startswith("data"):
self.imdb_file = 'data/' + self.imdb_file
#self.imdb_file = self._get_absolute_path(self.imdb_file)
self.imdb = ImageDatabase(self.imdb_file)
self.kwargs = kwargs
self.image_depth_first = self.config.image_depth_first
self._should_fast_read = self.config.fast_read
self.use_ocr = self.config.use_ocr
self.use_ocr_info = self.config.use_ocr_info
self._use_features = False
if hasattr(self.config, "image_features"):
self._use_features = True
self.features_max_len = self.config.features_max_len
all_image_feature_dirs = self.config.image_features[dataset_type]
curr_image_features_dir = all_image_feature_dirs[imdb_file_index]
curr_image_features_dir = curr_image_features_dir.split(",")
curr_image_features_dir = self._get_absolute_path(
curr_image_features_dir)
self.features_db = FeaturesDataset(
"coco",
directories=curr_image_features_dir,
depth_first=self.image_depth_first,
max_features=self.features_max_len,
fast_read=self._should_fast_read,
imdb=self.imdb,
)