train_demo.py

from fewshot_re_kit.data_loader import get_loader, get_loader_pair, get_loader_unsupervised
from fewshot_re_kit.framework import FewShotREFramework
from fewshot_re_kit.sentence_encoder import CNNSentenceEncoder, BERTSentenceEncoder, BERTPAIRSentenceEncoder, RobertaSentenceEncoder, RobertaPAIRSentenceEncoder
import models
from models.proto import Proto
from models.gnn import GNN
from models.snail import SNAIL
from models.metanet import MetaNet
from models.siamese import Siamese
from models.pair import Pair
from models.d import Discriminator
from models.mtb import Mtb
import sys
import torch
from torch import optim, nn
import numpy as np
import json
import argparse
import os

def main():
    parser = argparse.ArgumentParser()
    parser.add_argument('--train', default='train_wiki',
            help='train file')
    parser.add_argument('--val', default='val_wiki',
            help='val file')
    parser.add_argument('--test', default='test_wiki',
            help='test file')
    parser.add_argument('--adv', default=None,
            help='adv file')
    parser.add_argument('--trainN', default=10, type=int,
            help='N in train')
    parser.add_argument('--N', default=5, type=int,
            help='N way')
    parser.add_argument('--K', default=5, type=int,
            help='K shot')
    parser.add_argument('--Q', default=5, type=int,
            help='Num of query per class')
    parser.add_argument('--batch_size', default=4, type=int,
            help='batch size')
    parser.add_argument('--train_iter', default=30000, type=int,
            help='num of iters in training')
    parser.add_argument('--val_iter', default=1000, type=int,
            help='num of iters in validation')
    parser.add_argument('--test_iter', default=10000, type=int,
            help='num of iters in testing')
    parser.add_argument('--val_step', default=2000, type=int,
           help='val after training how many iters')
    parser.add_argument('--model', default='proto',
            help='model name')
    parser.add_argument('--encoder', default='cnn',
            help='encoder: cnn or bert or roberta')
    parser.add_argument('--max_length', default=128, type=int,
           help='max length')
    parser.add_argument('--lr', default=-1, type=float,
           help='learning rate')
    parser.add_argument('--weight_decay', default=1e-5, type=float,
           help='weight decay')
    parser.add_argument('--dropout', default=0.0, type=float,
           help='dropout rate')
    parser.add_argument('--na_rate', default=0, type=int,
           help='NA rate (NA = Q * na_rate)')
    parser.add_argument('--grad_iter', default=1, type=int,
           help='accumulate gradient every x iterations')
    parser.add_argument('--optim', default='sgd',
           help='sgd / adam / adamw')
    parser.add_argument('--hidden_size', default=230, type=int,
           help='hidden size')
    parser.add_argument('--load_ckpt', default=None,
           help='load ckpt')
    parser.add_argument('--save_ckpt', default=None,
           help='save ckpt')
    parser.add_argument('--fp16', action='store_true',
           help='use nvidia apex fp16')
    parser.add_argument('--only_test', action='store_true',
           help='only test')
    parser.add_argument('--ckpt_name', type=str, default='',
           help='checkpoint name.')


    # only for bert / roberta
    parser.add_argument('--pair', action='store_true',
           help='use pair model')
    parser.add_argument('--pretrain_ckpt', default=None,
           help='bert / roberta pre-trained checkpoint')
    parser.add_argument('--cat_entity_rep', action='store_true',
           help='concatenate entity representation as sentence rep')

    # only for prototypical networks
    parser.add_argument('--dot', action='store_true', 
           help='use dot instead of L2 distance for proto')

    # only for mtb
    parser.add_argument('--no_dropout', action='store_true',
           help='do not use dropout after BERT (still has dropout in BERT).')
    
    # experiment
    parser.add_argument('--mask_entity', action='store_true',
           help='mask entity names')
    parser.add_argument('--use_sgd_for_bert', action='store_true',
           help='use SGD instead of AdamW for BERT.')

    opt = parser.parse_args()
    trainN = opt.trainN
    N = opt.N
    K = opt.K
    Q = opt.Q
    batch_size = opt.batch_size
    model_name = opt.model
    encoder_name = opt.encoder
    max_length = opt.max_length
    
    print("{}-way-{}-shot Few-Shot Relation Classification".format(N, K))
    print("model: {}".format(model_name))
    print("encoder: {}".format(encoder_name))
    print("max_length: {}".format(max_length))
    
    if encoder_name == 'cnn':
        try:
            glove_mat = np.load('./pretrain/glove/glove_mat.npy')
            glove_word2id = json.load(open('./pretrain/glove/glove_word2id.json'))
        except:
            raise Exception("Cannot find glove files. Run glove/download_glove.sh to download glove files.")
        sentence_encoder = CNNSentenceEncoder(
                glove_mat,
                glove_word2id,
                max_length)
    elif encoder_name == 'bert':
        pretrain_ckpt = opt.pretrain_ckpt or 'bert-base-uncased'
        if opt.pair:
            sentence_encoder = BERTPAIRSentenceEncoder(
                    pretrain_ckpt,
                    max_length)
        else:
            sentence_encoder = BERTSentenceEncoder(
                    pretrain_ckpt,
                    max_length,
                    cat_entity_rep=opt.cat_entity_rep,
                    mask_entity=opt.mask_entity)
    elif encoder_name == 'roberta':
        pretrain_ckpt = opt.pretrain_ckpt or 'roberta-base'
        if opt.pair:
            sentence_encoder = RobertaPAIRSentenceEncoder(
                    pretrain_ckpt,
                    max_length)
        else:
            sentence_encoder = RobertaSentenceEncoder(
                    pretrain_ckpt,
                    max_length,
                    cat_entity_rep=opt.cat_entity_rep)
    else:
        raise NotImplementedError
    
    if opt.pair:
        train_data_loader = get_loader_pair(opt.train, sentence_encoder,
                N=trainN, K=K, Q=Q, na_rate=opt.na_rate, batch_size=batch_size, encoder_name=encoder_name)
        val_data_loader = get_loader_pair(opt.val, sentence_encoder,
                N=N, K=K, Q=Q, na_rate=opt.na_rate, batch_size=batch_size, encoder_name=encoder_name)
        test_data_loader = get_loader_pair(opt.test, sentence_encoder,
                N=N, K=K, Q=Q, na_rate=opt.na_rate, batch_size=batch_size, encoder_name=encoder_name)
    else:
        train_data_loader = get_loader(opt.train, sentence_encoder,
                N=trainN, K=K, Q=Q, na_rate=opt.na_rate, batch_size=batch_size)
        val_data_loader = get_loader(opt.val, sentence_encoder,
                N=N, K=K, Q=Q, na_rate=opt.na_rate, batch_size=batch_size)
        test_data_loader = get_loader(opt.test, sentence_encoder,
                N=N, K=K, Q=Q, na_rate=opt.na_rate, batch_size=batch_size)
        if opt.adv:
            adv_data_loader = get_loader_unsupervised(opt.adv, sentence_encoder,
                N=trainN, K=K, Q=Q, na_rate=opt.na_rate, batch_size=batch_size)
   
    if opt.optim == 'sgd':
        pytorch_optim = optim.SGD
    elif opt.optim == 'adam':
        pytorch_optim = optim.Adam
    elif opt.optim == 'adamw':
        from transformers import AdamW
        pytorch_optim = AdamW
    else:
        raise NotImplementedError
    if opt.adv:
        d = Discriminator(opt.hidden_size)
        framework = FewShotREFramework(train_data_loader, val_data_loader, test_data_loader, adv_data_loader, adv=opt.adv, d=d)
    else:
        framework = FewShotREFramework(train_data_loader, val_data_loader, test_data_loader)
        
    prefix = '-'.join([model_name, encoder_name, opt.train, opt.val, str(N), str(K)])
    if opt.adv is not None:
        prefix += '-adv_' + opt.adv
    if opt.na_rate != 0:
        prefix += '-na{}'.format(opt.na_rate)
    if opt.dot:
        prefix += '-dot'
    if opt.cat_entity_rep:
        prefix += '-catentity'
    if len(opt.ckpt_name) > 0:
        prefix += '-' + opt.ckpt_name
    
    if model_name == 'proto':
        model = Proto(sentence_encoder, dot=opt.dot)
    elif model_name == 'gnn':
        model = GNN(sentence_encoder, N, hidden_size=opt.hidden_size)
    elif model_name == 'snail':
        model = SNAIL(sentence_encoder, N, K, hidden_size=opt.hidden_size)
    elif model_name == 'metanet':
        model = MetaNet(N, K, sentence_encoder.embedding, max_length)
    elif model_name == 'siamese':
        model = Siamese(sentence_encoder, hidden_size=opt.hidden_size, dropout=opt.dropout)
    elif model_name == 'pair':
        model = Pair(sentence_encoder, hidden_size=opt.hidden_size)
    elif model_name == 'mtb':
        model = Mtb(sentence_encoder, use_dropout=not opt.no_dropout)
    else:
        raise NotImplementedError
    
    if not os.path.exists('checkpoint'):
        os.mkdir('checkpoint')
    ckpt = 'checkpoint/{}.pth.tar'.format(prefix)
    if opt.save_ckpt:
        ckpt = opt.save_ckpt

    if torch.cuda.is_available():
        model.cuda()

    if not opt.only_test:
        if encoder_name in ['bert', 'roberta']:
            bert_optim = True
        else:
            bert_optim = False

        if opt.lr == -1:
            if bert_optim:
                opt.lr = 2e-5
            else:
                opt.lr = 1e-1
        
        opt.train_iter = opt.train_iter * opt.grad_iter
        framework.train(model, prefix, batch_size, trainN, N, K, Q,
                pytorch_optim=pytorch_optim, load_ckpt=opt.load_ckpt, save_ckpt=ckpt,
                na_rate=opt.na_rate, val_step=opt.val_step, fp16=opt.fp16, pair=opt.pair, 
                train_iter=opt.train_iter, val_iter=opt.val_iter, bert_optim=bert_optim, 
                learning_rate=opt.lr, use_sgd_for_bert=opt.use_sgd_for_bert, grad_iter=opt.grad_iter)
    else:
        ckpt = opt.load_ckpt
        if ckpt is None:
            print("Warning: --load_ckpt is not specified. Will load Hugginface pre-trained checkpoint.")
            ckpt = 'none'

    acc = framework.eval(model, batch_size, N, K, Q, opt.test_iter, na_rate=opt.na_rate, ckpt=ckpt, pair=opt.pair)
    print("RESULT: %.2f" % (acc * 100))

if __name__ == "__main__":
    main()