Vincentqyw
fix: roma
c74a070
raw
history blame
1.81 kB
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()