main.py

"""
Implementation of ECCV 2018 paper "Graph R-CNN for Scene Graph Generation".
Author: Jianwei Yang, Jiasen Lu, Stefan Lee, Dhruv Batra, Devi Parikh
Contact: jw2yang@gatech.edu
"""

import os
import pprint
import argparse
import numpy as np
import torch
import datetime

from lib.config import cfg
from lib.model import build_model
from lib.scene_parser.rcnn.utils.miscellaneous import mkdir, save_config, get_timestamp
from lib.scene_parser.rcnn.utils.comm import synchronize, get_rank
from lib.scene_parser.rcnn.utils.logger import setup_logger

def train(cfg, args):
    """
    train scene graph generation model
    """
    arguments = {}
    arguments["iteration"] = 0
    model = build_model(cfg, arguments, args.local_rank, args.distributed)
    model.train()
    return model

def test(cfg, args, model=None):
    """
    test scene graph generation model
    """
    if model is None:
        arguments = {}
        arguments["iteration"] = 0
        model = build_model(cfg, arguments, args.local_rank, args.distributed)
    model.test(visualize=args.visualize)

def main():
    ''' parse config file '''
    parser = argparse.ArgumentParser(description="Scene Graph Generation")
    parser.add_argument("--config-file", default="configs/baseline_res101.yaml")
    parser.add_argument("--local_rank", type=int, default=0)
    parser.add_argument("--session", type=int, default=0)
    parser.add_argument("--resume", type=int, default=0)
    parser.add_argument("--batchsize", type=int, default=0)
    parser.add_argument("--inference", action='store_true')
    parser.add_argument("--instance", type=int, default=-1)
    parser.add_argument("--use_freq_prior", action='store_true')
    parser.add_argument("--visualize", action='store_true')
    parser.add_argument("--algorithm", type=str, default='sg_baseline')
    args = parser.parse_args()

    num_gpus = int(os.environ["WORLD_SIZE"]) if "WORLD_SIZE" in os.environ else 1
    args.distributed = num_gpus > 1
    if args.distributed:
        torch.cuda.set_device(args.local_rank)
        torch.distributed.init_process_group(
            backend="nccl", init_method="env://"
        )
        synchronize()

    cfg.merge_from_file(args.config_file)
    cfg.resume = args.resume
    cfg.instance = args.instance
    cfg.inference = args.inference
    cfg.MODEL.USE_FREQ_PRIOR = args.use_freq_prior
    cfg.MODEL.ALGORITHM = args.algorithm
    if args.batchsize > 0:
        cfg.DATASET.TRAIN_BATCH_SIZE = args.batchsize
    if args.session > 0:
        cfg.MODEL.SESSION = str(args.session)
    # cfg.freeze()

    if not os.path.exists("logs") and get_rank() == 0:
        os.mkdir("logs")
    logger = setup_logger("scene_graph_generation", "logs", get_rank(),
        filename="{}_{}.txt".format(args.algorithm, get_timestamp()))
    logger.info(args)
    logger.info("Loaded configuration file {}".format(args.config_file))
    output_config_path = os.path.join("logs", 'config.yml')
    logger.info("Saving config into: {}".format(output_config_path))
    save_config(cfg, output_config_path)

    if not args.inference:
        model = train(cfg, args)
    else:
        test(cfg, args)

if __name__ == "__main__":
    main()