def main(args):
imgaug.seed(42)
torch.random.manual_seed(42)
random.seed(42)
if os.path.isabs(args.cfg):
cfg.merge_from_file(args.cfg)
else:
cfg.merge_from_file(os.path.join(RepoPaths.configs_dir(), args.cfg))
if args.dataset == "coco":
dataset = CocoDataLoader(CocoPaths.images_dir(),
CocoPaths.ids_file(),
category_agnostic=False)
elif args.dataset == "mapillary":
dataset = MapillaryDataLoader(MapillaryPaths.images_dir(),
MapillaryPaths.ids_file())
elif args.dataset == "pascalvoc":
dataset = PascalVOCDataLoader(PascalVOCPaths.images_dir(),
PascalVOCPaths.ids_file(),
category_agnostic=False)
elif args.dataset == "ytvis":
dataset = YoutubeVISDataLoader(YoutubeVISPaths.training_base_dir(),
YoutubeVISPaths.train_vds_file(),
cfg.TRAINING.TRACKER.MAX_ITERATIONS,
category_agnostic=False,
single_instance_duplication=cfg.DATA.
YOUTUBE_VIS.SINGLE_INSTANCE_DUPLICATION)
elif args.dataset == "davis":
dataset = DavisDataLoader(
DavisUnsupervisedPaths.trainval_base_dir(),
DavisUnsupervisedPaths.train_vds_file(),
apply_augmentation=False,
samples_to_create=cfg.DATA.DAVIS.TRAINING_SUBSEQUENCES,
single_instance_duplication=cfg.DATA.DAVIS.
SINGLE_INSTANCE_DUPLICATION)
elif args.dataset == "kittimots":
dataset = MOTSDataLoader(
KITTIMOTSPaths.train_images_dir(),
KITTIMOTSPaths.train_vds_file(),
samples_to_create=cfg.TRAINING.TRACKER.MAX_ITERATIONS,
apply_augmentation=cfg.DATA.KITTI_MOTS.AUGMENTATION,
frame_gap_lower=cfg.DATA.KITTI_MOTS.FRAME_GAP_LOWER,
frame_gap_upper=cfg.DATA.KITTI_MOTS.FRAME_GAP_UPPER)
else:
raise ValueError("Invalid dataset name given")
visualize_data_loader_output(dataset, args.num_workers, args.batch_size,
args.shuffle)
def load_cfg(args):
cfg_file = os.path.join(os.path.dirname(args.model_path), 'config.yaml')
if not os.path.exists(cfg_file):
dataset_cfgs = {
"davis": "davis_2.yaml",
"ytvis": "youtube_vis.yaml",
"kittimots": "kitti_mots_2.yaml"
}
assert args.dataset in dataset_cfgs, \
"Invalid '--dataset' argument. Should be either 'davis', 'ytvis' or 'kittimots'"
cfg_file = os.path.join(RepoPaths.configs_dir(), dataset_cfgs[args.dataset])
print("Loading config from {}".format(cfg_file))
cfg.merge_from_file(cfg_file)
def main(args):
if os.path.isabs(args.cfg):
cfg_path = args.cfg
else:
cfg_path = os.path.join(RepoPaths.configs_dir(), args.cfg)
print("Restoring config from: {}".format(cfg_path))
global_cfg.merge_from_file(cfg_path)
num_gpus = torch.cuda.device_count()
if args.allow_multigpu and num_gpus > 1:
init_distributed(args, global_cfg.TRAINING, num_gpus)
else:
start(args, global_cfg.TRAINING)
def setup_cfg(args, model_dir, ignore_existing_cfg):
# if a config file has been provided, load it
if args.cfg:
print("[ INFO] Loading config from {}".format(args.cfg))
global_cfg.merge_from_file(args.cfg)
return
if ignore_existing_cfg:
return
# when restoring session, load the config file already present in that directory
if args.restore_session:
expected_config_filepath = os.path.realpath(
os.path.join(args.restore_session, os.pardir, 'config.yaml'))
print("[ INFO] Restoring config from {}".format(
expected_config_filepath))
global_cfg.merge_from_file(expected_config_filepath)
# if the output directory already exists and there is a config file present there, then load it
else:
expected_config_filepath = os.path.join(model_dir, 'config.yaml')
if os.path.exists(expected_config_filepath):
print("[ INFO] Restoring config from {}".format(
expected_config_filepath))
global_cfg.merge_from_file(expected_config_filepath)