MemGen-main/main.py

import argparse
from datetime import datetime
import os
import random

import numpy as np
import torch

from common.config import Config
from common.logger import setup_logger
from data import get_data_builder
from memgen.model import MemGenModel
from memgen.runner import MemGenRunner

def set_seed(random_seed: int, use_gpu: bool):

    random.seed(random_seed)
    os.environ['PYTHONHASHSEED'] = str(random_seed)
    np.random.seed(random_seed)
    torch.manual_seed(random_seed)
    torch.cuda.manual_seed(random_seed)
    if use_gpu:
        torch.cuda.manual_seed_all(random_seed)

    torch.backends.cudnn.deterministic = True   
    torch.backends.cudnn.benchmark = False      

    print(f"set seed: {random_seed}")

def parse_args():
    parser = argparse.ArgumentParser(description="Memory Generator")

    parser.add_argument("--cfg-path", required=True, help="path to configuration file.")
    parser.add_argument(
        "--options",
        nargs="+",
        help="override some settings in the used config, the key-value pair "
        "in xxx=yyy format will be merged into config file (deprecate), "
        "change to --cfg-options instead.",
    )

    args = parser.parse_args()

    return args

def build_working_dir(config: Config) -> str:
    
    # parent dir: <train/evaluate>/<dataset_name>/<reasoner_model_name>
    mode = config.run_cfg.mode
    dataset_name = config.dataset_cfg.name
    model_name = config.model_cfg.model_name.split("/")[1]
    parent_dir = os.path.join(".cache", mode, dataset_name, model_name)

    # name: <prompt_aug_num>_<prompt_latents_len>_<inference_aug_num>_<inference_latents_len>_<timestamp>
    max_prompt_aug_num = config.model_cfg.max_prompt_aug_num
    prompt_latents_len = config.model_cfg.weaver.prompt_latents_len
    max_inference_aug_num = config.model_cfg.max_inference_aug_num
    inference_latents_len = config.model_cfg.weaver.inference_latents_len
    time = datetime.now().strftime("%Y%m%d-%H%M%S")
    working_dir = f"pn={max_prompt_aug_num}_pl={prompt_latents_len}_in={max_inference_aug_num}_il={inference_latents_len}_{time}" 

    return os.path.join(parent_dir, working_dir)

def main():

    args = parse_args()
    config = Config(args)

    set_seed(config.run_cfg.seed, use_gpu=True)
    
    # set up working directory
    working_dir = build_working_dir(config)
    
    # set up logger
    config.run_cfg.log_dir = os.path.join(working_dir, "logs")
    setup_logger(output_dir=config.run_cfg.log_dir)

    config.pretty_print()

    # build components
    config_dict = config.to_dict()
    data_builder = get_data_builder(config_dict.get("dataset"))
    model = MemGenModel.from_config(config_dict.get("model"))
    
    runner = MemGenRunner(
        model=model,
        data_builder=data_builder,
        config=config_dict,
        working_dir=working_dir
    )

    # train or evaluate 
    if config.run_cfg.mode == "train":
        runner.train()
    
    elif config.run_cfg.mode == "evaluate":
        runner.evaluate()

if __name__ == "__main__":
    main()