BERT-Pretrain-Multi-GPU/Predictor.py at main · fredericklee602/BERT-Pretrain-Multi-GPU · 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

import os
from posixpath import sep
import time
import random
import logging
import math
import numpy as np
import pandas as pd
import torch
import torch.nn.functional as F
from apex import amp
from tqdm.auto import tqdm
from datasets import Dataset, load_dataset, load_metric
from torch.utils.data import DataLoader
from transformers import AdamW, AutoModelForSequenceClassification, get_scheduler, get_linear_schedule_with_warmup,AutoTokenizer
# from transformers import BertTokenizer, BertConfig, AutoConfig, BertForMaskedLM, DistilBertForMaskedLM, DistilBertTokenizer, AutoTokenizer
from model.BertForMaskedLM import BertForMaskedLM
from sklearn import metrics
from Config import Config


class Predictor(object):

    def __init__(self, config):
        self.config = config
        # self.test_loader = test_loader
        self.device = torch.device(self.config.device)
        # 加載模型
        self.load_tokenizer()
        self.load_model()


    def load_tokenizer(self):
        """
        讀取分詞器
        """
        print('loading tokenizer config ...')
        self.tokenizer = AutoTokenizer.from_pretrained(self.config.initial_pretrain_tokenizer)


    def load_model(self):
        """
        加載模型及初始化模型參數
        """
        print('loading model...%s' %self.config.path_model_predict)
        self.model = BertForMaskedLM.from_pretrained(self.config.path_model_predict)
        # self.model = BertForMaskedLM.from_pretrained('bert-base-chinese')
        # self.model = BertForMaskedLM.from_pretrained('ckiplab/bert-base-chinese')
        # 將模型加載到CPU/GPU
        self.model.to(self.device)
        self.model.eval()


    def predict(self, test_loader):
        """
        預測
        """
        print('predict start')
        # 初始化指標計算
        src = []
        label = []
        pred = []
        input = []
        acc_label = []
        acc_pred = []
        acc_input = []
        print("Batch Length:{0}".format(len(test_loader)))
        for i, batch in enumerate(test_loader):
            # 推斷
            batch = {k:v.to(self.config.device) for k,v in batch.items()}
            with torch.no_grad():
                outputs = self.model(**batch)
                outputs_pred = outputs.logits
            if i%100==0:
                print("Batch No.{0} completed".format(i))
            # 還原成token string
            tmp_src = batch['input_ids'].cpu().numpy()
            tmp_label = batch['labels'].cpu().numpy()
            tmp_pred = torch.max(outputs_pred, -1)[1].cpu().numpy()
            for j in range(len(tmp_label)):
                line_s = tmp_src[j]
                line_l = tmp_label[j]
                line_l_split = [ x for x in line_l if x not in [0]]
                line_p = tmp_pred[j]
                line_p_split = line_p[:len(line_l_split)]
                token_s = self.tokenizer.convert_ids_to_tokens(line_s)
                tmp_s = self.tokenizer.convert_tokens_to_string(token_s)
                tmp_s = tmp_s.replace('[PAD] ','')
                token_lab = self.tokenizer.convert_ids_to_tokens(line_l_split)
                tmp_lab = self.tokenizer.convert_tokens_to_string(token_lab)
                token_p = self.tokenizer.convert_ids_to_tokens(line_p_split)
                tmp_p = self.tokenizer.convert_tokens_to_string(token_p)
                input.append(tmp_s)
                label.append(tmp_lab)
                pred.append(tmp_p)
                acc_label.append(token_lab)
                acc_pred.append(token_p)
                acc_input.append(token_s)

        # 計算指標
        total = 0
        count = 0
        c=0
        print("Present Label and Prediction vocabulary correspondence when the vocabulary was masked.")
        for k,(l,p,s) in enumerate(zip(acc_label, acc_pred, acc_input)):
            for i in range(len(s)):
                c=c+1
                if s[i]=="[MASK]":
                    if c<50:
                        print("l:",l[i])
                        print("p:",p[i])
                    total += 1
                    if l[i]==p[i]:
                        count += 1
        acc = count/max(1, total)
        print('\nTotal Accuracy: )
        print('\nTask: count=',count)
        print('\nTask: total=',total)
        print('\nTask: acc=',acc)

        # 保存
        # Task 1
        data = {'src':label, 'pred':pred, 'mask':input}
        data = pd.DataFrame(data)
        path = os.path.join(self.config.path_datasets, 'output')
        if not os.path.exists(path):
            os.mkdir(path)
        path_output = os.path.join(path, 'pred_data.csv')
        data.to_csv(path_output, index=False)
        print('Task 1: predict result save: {}'.format(path_output))