STGE/scorer_train.py at main · ASCII-LAB/STGE · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
import torch
from utils import *
from transformers import RobertaTokenizer, AdamW, get_linear_schedule_with_warmup
from scorer_data_precess import get_dataloaber
from stge import extract_and_evaluate
from scorer_data_precess import save_data, get_pred_data
from config import get_args_parser
from scorer import ScorerModel


def train(args, model, train_data, optimizer):
    max_acc = 0
    for epoch_idx in range(args.epoch):
        pred_pos, pred_neg = 0, 0
        acc_num, num = 0, 0
        train_loss = 0
        for idx, batch in enumerate(train_data):
            model.train()
            pred, loss = model(batch)
            train_pred = pred.argmax(1).cpu().numpy().tolist()
            train_gold = batch[-1].cpu().numpy().tolist()
            num += len(train_gold)
            for pred_idx in range(len(train_pred)):
                if train_pred[pred_idx] == train_gold[pred_idx]:
                    acc_num += 1
                if train_pred[pred_idx] == 1:
                    pred_pos += 1
                else:
                    pred_neg += 1

            loss.backward()
            train_loss += loss.item()
            optimizer.step()
            model.zero_grad()

        print("train pred_pos, pred_neg", pred_pos, pred_neg)
        train_acc = acc_num/num
        print("epoch={}, train_acc={}, train_loss={}".format(epoch_idx, train_acc, train_loss))
        torch.save(model.state_dict(), args.model_save_path)
        print("save model")
    print("max_acc: ", max_acc)


def begin_supervise_train(args):
    if args.load_scorer_path == "":
        model = ScorerModel(args).to(args.device)
    else:
        print("load {}".format(args.load_scorer_path))
        model = ScorerModel(args)
        model.load_state_dict(torch.load(args.load_scorer_path, map_location="cpu"))
        model = model.to(args.device)
    tokenizer = RobertaTokenizer.from_pretrained(args.model_name)
    train_data = get_dataloaber(args, args.scorer_train_data_path, tokenizer)
    no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
    optimizer_grouped_parameters = [
        {
            "params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
            "weight_decay": args.weight_decay,
        },
        {
            "params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)],
            "weight_decay": 0.0
        },
    ]
    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate)
    train(args, model, train_data, optimizer)
    extract_and_evaluate(args, args.convert_path + "/test.json")


def get_iterative_data(args, epoch_idx, tokenizer):
    all_data = read_json(args.save_result_path)
    pred_data_list = get_pred_data(args, all_data)
    if len(pred_data_list) != 0:
        train_save_path = args.scorer_data_path + "{}_train_epoch_{}.tsv".format(args.model_save, epoch_idx)
        save_data(pred_data_list, train_save_path)
        train_data = get_dataloaber(args, train_save_path, tokenizer)
        return train_data
    else:
        return []


def begin_iterative_train(args):
    if args.load_scorer_path == "":
        model = ScorerModel(args).to(args.device)
    else:
        print("load {}".format(args.load_scorer_path))
        model = ScorerModel(args)
        model.load_state_dict(torch.load(args.load_scorer_path, map_location="cpu"))
        model = model.to(args.device)
    tokenizer = RobertaTokenizer.from_pretrained(args.model_name)
    no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
    optimizer_grouped_parameters = [
        {
            "params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
            "weight_decay": args.weight_decay,
        },
        {
            "params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)],
            "weight_decay": 0.0
        },
    ]
    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate)

    original_data = get_dataloaber(args, args.scorer_train_data_path, tokenizer)

    original_model_path = args.model_save_path
    args.model_save_path = args.load_scorer_path
    args.save_result_path = "output/{}/{}_{}_result_original.json".format(
        args.dataset_type, args.model_save, args.llm_name)
    with torch.no_grad():
        current_f1 = extract_and_evaluate(args, args.shot_data_path)
    print("current_f1", current_f1)
    print("begin train", "-"*30)
    for epoch_idx in range(args.epoch):
        if args.pred_max_num == 0:
            all_train_data = [iter(original_data)]
        else:
            pred_data = get_iterative_data(args, epoch_idx, tokenizer)
            all_train_data = [iter(pred_data)]
        pred_pos, pred_neg = 0, 0
        acc_num, num = 0, 0
        train_loss = 0
        for i in range(len(all_train_data)):
            iter_dataloader = all_train_data[i]
            for idx, batch in enumerate(iter_dataloader):
                model.train()
                pred, loss = model(batch)
                train_pred = pred.argmax(1).cpu().numpy().tolist()
                train_gold = batch[-1].cpu().numpy().tolist()
                num += len(train_gold)
                for pred_idx in range(len(train_pred)):
                    if train_pred[pred_idx] == train_gold[pred_idx]:
                        acc_num += 1
                    if train_pred[pred_idx] == 1:
                        pred_pos += 1
                    else:
                        pred_neg += 1

                loss.backward()
                train_loss += loss.item()
                optimizer.step()
                model.zero_grad()

        print("train pred_pos, pred_neg", pred_pos, pred_neg)
        train_acc = acc_num / num
        print("epoch={}, train_acc={}, train_loss={}".format(epoch_idx, train_acc, train_loss))
        args.model_save_path = original_model_path
        torch.save(model.state_dict(), args.model_save_path)
        print("save model", args.model_save_path)
        args.save_result_path = "output/{}/{}_{}_epoch{}.json".format(
            args.dataset_type, args.model_save, args.llm_name, epoch_idx)
        with torch.no_grad():
            current_f1 = extract_and_evaluate(args, args.shot_data_path)
        print("current_f1", current_f1)
    args.save_result_path = "output/{}/{}_{}_all.json".format(
        args.dataset_type, args.model_save, args.llm_name)
    with torch.no_grad():
        extract_and_evaluate(args, args.convert_path + "/test.json")


if __name__ == '__main__':
    args = get_args_parser()
    print(args)
    if args.train_type == "supervise":
        begin_supervise_train(args)
    elif args.train_type == "stge":
        begin_iterative_train(args)
    else:
        print("train_type error")