config.py

import argparse


def set_args():
    parser = argparse.ArgumentParser()
    parser = basic_args(parser)
    parser = tokenizer_args(parser)
    parser = train_args(parser)
    parser = eval_args(parser)
    parser = encoder_args(parser)
    parser = top_args(parser)
    parser = target_args(parser)

    args = parser.parse_args()
    return args


def basic_args(parser):
    """
        CLUE:
        clue.csl clue.afqmc clue.cluewsc2020 clue.cmnli clue.iflytek clue.tnews

        JD:
        jd21.褪黑素 jd21.维生素 jd21.无线耳机 jd21.蛋白粉 jd21.游戏机 jd21.电视 jd21.MacBook
        jd21.洗面奶 jd21.智能手表 jd21.吹风机 jd21.小米手机 jd21.红米手机 jd21.护肤品 jd21.电动牙刷
        jd21.iPhone jd21.海鲜 jd21.酒 jd21.平板电脑 jd21.修复霜 jd21.运动鞋 jd21.智能手环

        STOCK:
        stock.airline stock.car stock.communication stock.energy stock.finance stock.manufacture
        stock.medical stock.Real_estate stock.tech stock.traffic stock.wine

        AMZ:
        amz20.Sandal amz20.Magazine_Subscriptions amz20.RiceCooker amz20.Flashlight amz20.Jewelry
        amz20.CableModem amz20.GraphicsCard amz20.GPS amz20.Projector amz20.Keyboard amz20.Video_Games
        amz20.AlarmClock amz20.HomeTheaterSystem amz20.Vacuum amz20.Gloves amz20.Baby amz20.Bag
        amz20.Movies_TV amz20.Dumbbell amz20.Headphone

        SNAP:
        snap10k.Automotive_5 snap10k.Electronics_5 snap10k.Industrial_and_Scientific_5 snap10k.Kindle_Store_5
        snap10k.Cell_Phones_and_Accessories_5 snap10k.Musical_Instruments_5 snap10k.Office_Products_5
        snap10k.Patio_Lawn_and_Garden_5 snap10k.Sports_and_Outdoors_5 snap10k.Luxury_Beauty_5
        snap10k.Grocery_and_Gourmet_Food_5 snap10k.Digital_Music_5 snap10k.Tools_and_Home_Improvement_5
        snap10k.Pet_Supplies_5 snap10k.Prime_Pantry_5 snap10k.Toys_and_Games_5 snap10k.Movies_and_TV_5
        snap10k.Home_and_Kitchen_5 snap10k.Arts_Crafts_and_Sewing_5 snap10k.Video_Games_5 snap10k.CDs_and_Vinyl_5
    """

    parser.add_argument("--task_list", default='tasklist/jd21.json', type=str, help='input task list,txt, or json')
    parser.add_argument("--output_dir", default='output/', type=str, help='文件保存目录')
    parser.add_argument('--seed', type=int, default=511, help="random seed for initialization")
    parser.add_argument('--approach', type=str, default='textcnn', help='网络名称')
    parser.add_argument('--comment', type=str, default='test', help='备注')
    parser.add_argument('--f1_macro', action='store_true')
    parser.add_argument('--mutli_task', action='store_true')
    parser.add_argument('--few_shot', action='store_true')
    return parser


def tokenizer_args(parser):
    parser.add_argument("--max_seq_length", default=256, type=int,
                        help="The maximum total input sequence length after WordPiece tokenization. \n"
                             "Sequences longer than this will be truncated, and sequences shorter \n"
                             "than this will be padded.")
    return parser


def train_args(parser):
    parser.add_argument("--train_batch_size", default=128, type=int, help="Total batch size for training.")
    parser.add_argument("--eval_batch_size", default=128, type=int, help="Total batch size for eval.")
    parser.add_argument('--epochs', type=int, default=25)
    parser.add_argument('--optimizer', type=str, default='sgd')
    parser.add_argument('--sgd_momentum', action='store_true')
    parser.add_argument('--lr', type=float, default=0.08)
    parser.add_argument('--lr_min', default=1e-4, type=float, required=False, help='(default=%(default)f)')
    parser.add_argument('--lr_factor', default=3, type=float, required=False, help='(default=%(default)f)')
    parser.add_argument('--lr_patience', default=3, type=int, required=False, help='(default=%(default)f)')
    parser.add_argument('--clipgrad', default=10000, type=float, required=False, help='(default=%(default)f)')
    return parser


def eval_args(parser):
    return parser


def encoder_args(parser):
    # 注意！！！ 请确定模型对应的输出维度是不是768。
    parser.add_argument("--bert_name", default='uer/roberta-base-finetuned-jd-binary-chinese', type=str, help='bert模型存放的目录')
    parser.add_argument("--bert_hidden_size", default=768, type=int,
                        help="在target网络中很多也用到这个参数设置网络大小,可能是384,768,1024或者其他.")
    parser.add_argument('--train_bert', action='store_true')
    parser.add_argument("--hidden_dropout_prob", default=0.1, type=float, help="Dropout rate for hidden states.")
    return parser


def top_args(parser):
    parser.add_argument("--top_name", default='textcnn', type=str, help="在encoder上加一个网络")
    return parser


def target_args(parser):
    parser.add_argument("--target_dropout_prob", default=0.5, type=float, help="Dropout rate for hidden states.")
    parser.add_argument("--target", default='linear', type=str, help="从向量到输出结果之间的网络.")
    return parser