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()