import cv2 import yaml import argparse import os from torch.utils.data import DataLoader from datasets.gl3d_dataset import GL3DDataset from trainer import Trainer from trainer_single_norel import SingleTrainerNoRel from trainer_single import SingleTrainer if __name__ == '__main__': # add argument parser parser = argparse.ArgumentParser() parser.add_argument('--config', type=str, default='./configs/config.yaml') parser.add_argument('--dataset_dir', type=str, default='/mnt/nvme2n1/hyz/data/GL3D') parser.add_argument('--data_split', type=str, default='comb') parser.add_argument('--is_training', type=bool, default=True) parser.add_argument('--job_name', type=str, default='') parser.add_argument('--gpu', type=str, default='0') parser.add_argument('--start_cnt', type=int, default=0) parser.add_argument('--stage', type=int, default=1) args = parser.parse_args() # load global config with open(args.config, 'r') as f: config = yaml.load(f, Loader=yaml.FullLoader) # setup dataloader dataset = GL3DDataset(args.dataset_dir, config['network'], args.data_split, is_training=args.is_training) data_loader = DataLoader(dataset, batch_size=2, shuffle=True, num_workers=4) os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu if args.stage == 1: trainer = SingleTrainerNoRel(config, f'cuda:0', data_loader, args.job_name, args.start_cnt) elif args.stage == 2: trainer = SingleTrainer(config, f'cuda:0', data_loader, args.job_name, args.start_cnt) elif args.stage == 3: trainer = Trainer(config, f'cuda:0', data_loader, args.job_name, args.start_cnt) else: raise NotImplementedError() trainer.train()