-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmerge_predict.py
208 lines (185 loc) · 9.11 KB
/
merge_predict.py
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
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
import sys
sys.path.append('./')
from transformers import BertTokenizer
import torch
import torch.nn.functional as F
from model import Model
from train import label2vec,tokenizer
import numpy as np
from torch.utils.data import DataLoader,Dataset
from utils import ljqpy
from sklearn.metrics import *
from tqdm import tqdm
from copy import deepcopy
import matplotlib.pyplot as plt
from utils.pred_helper import transfer,check_text,Val0_Classify
from utils.sortlabel import llist
from datapreprocess import Normalize
# tokenizer_ernie = BertTokenizer.from_pretrained("nghuyong/ernie-3.0-base-zh")
def load_val(fn):
'''
有的模型是在原始文本上训练的，有的文本是在Normalize后的文本上训练的，故需加载两种文本
'''
return [(x["text_normd"],Normalize(x["text_normd"]), x["label"]) for x in ljqpy.LoadJsons(fn)]
class MergeDataset(Dataset):
def __init__(self,data, requires_index = False):
super().__init__()
self.data = []
for i,d in enumerate(data):
text1 = d[0]
text2 = d[1]
label = label2vec(d[2])
if requires_index:
self.data.append([text1,text2,label,torch.tensor([i])])
else:
self.data.append([text1,text2,label])
def __getitem__(self, index):
return self.data[index]
def __len__(self):
return len(self.data)
def merge_fn(batch):
# print(len(batch))
z1 = tokenizer([d[0] for d in batch],return_tensors='pt',truncation=True, max_length=128,padding=True)
z2 = tokenizer([d[1] for d in batch],return_tensors='pt',truncation=True, max_length=128,padding=True)
# z3 = tokenizer([d[1] for d in batch],return_tensors='pt',truncation=True, max_length=128,padding=True)
if len(batch[0]) ==3:
return (z1.input_ids,
z2.input_ids,
# z3.input_ids,
torch.stack([x[2] for x in batch], 0))
else:
return (z1.input_ids,
z2.input_ids,
# z3.input_ids,
torch.stack([x[2] for x in batch], 0),
torch.cat([x[3] for x in batch],0))
def get_rank(tensor:torch.Tensor,ind): # 要求的是某个index所在元素的排行
t= tensor.sort(descending=True)[1].tolist()
dic = {k:v+1 for v, k in enumerate(t)}
# print(ind in t)
return dic[ind]
def pick_len(data, l=10):
res = []
for d in data:
if len(d[0]) <= l:
res.append(d)
return res
# f1_l = []
zero_indices = []
class Multi_label:
def __init__(self, model,mfile_l,llist,n=4,device=torch.device('cuda')):
# self.model = model.to(device)
self.device = device
self.model_l = [deepcopy(model).to(device) for i in range(len(mfile_l))]
for i,mfile in enumerate(mfile_l):
# print(i)
self.model_l[i].load_state_dict(state_dict = torch.load(mfile, map_location=device))
self.model_l[i].eval()
self.tl =llist
self.n = n # 从n开始是去表情版
def eval_on_val(self,val_data,val_loader):
# val_loader: requires_index = True,否则取不了bad_indices
yt, yp, bad_case_indices = [], [], []
# w = torch.tensor([0.2]*5)
# self.model.eval()
pbar = tqdm(val_loader, total=len(val_loader))
Classifier = Val0_Classify(llist)
with torch.no_grad():
for xx1,xx2, yy,indices in pbar:
xx1,xx2 = xx1.to(self.device),xx2.to(self.device)
scores_l = torch.zeros(yy.shape[0],yy.shape[1],len(self.model_l))
for i,model in enumerate(self.model_l):
if i < self.n:
scores_l[...,i] = model(xx1).detach().cpu()
# elif i == len(self.model_l)-1:
# scores_l[...,i] = model(xx3).detach().cpu()
else:
scores_l[...,i] = model(xx2).detach().cpu()
# if i > 1:
# scores_l[...,i] = torch.sigmoid(scores_l[...,i])
# print(scores_l[...,0]==scores_l[...,1])
scores = scores_l.mean(-1)
# scores = torch.matmul(scores_l,w)
scores_max,_ = scores_l.max(-1)
# scores_l_vote = (scores_l > 0.5).float().cpu()
# zz1 = (scores_l_vote.mean(-1)> 0.5).float().cpu()
scores = transfer(scores)
zz = (scores > 0.5).float().cpu()
# zz = ((zz1 + zz2) > 1).float().cpu()
# true_id = yy.nonzero().squeeze(1)
# scores = transfer(scores)
# scores_max = transfer(scores_max)
for idx,z in enumerate(zz):
text = val_data[indices[idx]][0]
ind_plus = check_text(text)
for j in ind_plus:
z[j] = 1
z_vec = z
# true_id = yy[idx].nonzero().squeeze(1)
# pred_id = z.nonzero().squeeze(1)
# z_score = [(self.tl.get_token(j),scores[idx][j].item()) for j in pred_id]
if sum(z) == 0: # 没有预测出标签时，用规则的argmax方法重新预测
z_vec = Classifier.fun(scores_max[idx])
# z_vec = transfer_z(z_vec)
zz[idx] = z_vec
# if any(yy[idx] != zz[idx]):
# bad_case_indices.append(indices[idx])
# pred_score_l.append(z_score)
# true_score_l.append([(self.tl.get_token(i), scores[idx][i].item(), get_rank(scores[idx], i.item())) for i in true_id])
# if zz[idx][78]==1 or zz[idx][79]==1 or scores[idx][78] + scores[idx][79] > 0.5:
# zz[idx][78],zz[idx][79] = 1,1
# zz[idx] = transfer_z(zz[idx], scores_max[idx])
yt.append(yy[idx])
yp.append(zz[idx])
yt = torch.stack(yt,0)
yp = torch.stack(yp,0)
# for i,d in enumerate(val_data):
# text = d[0]
# ind_plus = check_text(text)
# for j in ind_plus:
# print(yp[i][j])
# yp[i][j] = 1
# print(yt.shape)
for i in range(len(yt[0])):
yt_i, yp_i = yt[:,i],yp[:,i]
if all(yp_i==0):
continue
if f1_score(yt_i,yp_i,zero_division=0) == 0:
zero_indices.append(i)
accu = accuracy_score(yt,yp)
prec = precision_score(yt,yp,average='samples',zero_division=0)
reca = recall_score(yt,yp,average='samples',zero_division=0)
f1 = f1_score(yt,yp,average='samples',zero_division=0)
print(f'Accu: {accu:.4f}, Prec: {prec:.4f}, Reca: {reca:.4f}, F1: {f1:.5f}')
# print(zero_indices)
# plt.savefig('./output/result/merge_plot')
return bad_case_indices #,true_score_l, pred_score_l
def write_bad_to_json(val_data,bad_case_indices,true_score_l, pred_score_l, out_path): # 将生成错误的例子加入到json文件中
data = []
for i in range(len(bad_case_indices)):
d = {}
d['text'] = val_data[bad_case_indices[i]][0]
d['true_score'] = true_score_l[i]
d['pred_score'] = pred_score_l[i]
data.append(d)
ljqpy.SaveJsons(data, out_path)
if __name__ == '__main__':
val_data = load_val('./dataset/val_normd2.json')
# val_data = pick_len(val_data,50)
# print(len(val_data))
val_ds = MergeDataset(val_data,requires_index=True)
val_dl = DataLoader(val_ds,collate_fn=merge_fn,batch_size=128)
model = Model(activation=True)
# mfile_l = ['./output/ljq/wb_base_retrain_lock7_diceloss_normd2_8981.pt'] + ['./output/base5/base5.ckpt', './output/base1/256base1.pt','./output/base7/128_base7.ckpt', './output/extra/128_base7_plus5.ckpt','./output/base7/wb_base_noseg_normd1_8855.pt',
# './output/ljq/wb_base_retrain_lock7_diceloss2_normd1_8846.pt','./output/ljq/wb_base_retrain2_lock10_diceloss_normd1_8871.pt','./output/ljq/wb_base_retrain2_lock7_diceloss_normd1_8858.pt',
# './output/ljq/wb_base_retrain_lock8_normd1_8844.pt','./output/ljq/wb_base_retrain_lock7_diceloss_normd1_8867.pt','./output/ljq/wb_base_retrain_lock6_normd1_8859.pt',
# './output/base7/base7noemo.ckpt','./output/base7/base7noemo_n2.ckpt','./output/ljq/wb_base2_noseg_normd2.pt']
# print(len(mfile_l))
file_l = ['wb_base_retrain_lock7_diceloss_normd2_8981.pt','wb_base2_noseg_normd2.pt'] + ['base5.ckpt', '256base1.pt','128_base7.ckpt', '128_base7_plus5.ckpt','wb_base_noseg_normd1_8855.pt',
'wb_base_retrain_lock7_diceloss2_normd1_8846.pt','wb_base_retrain2_lock10_diceloss_normd1_8871.pt','wb_base_retrain2_lock7_diceloss_normd1_8858.pt',
'wb_base_retrain_lock8_normd1_8844.pt','wb_base_retrain_lock7_diceloss_normd1_8867.pt','wb_base_retrain_lock6_normd1_8859.pt',
'base7noemo.ckpt','base7noemo_n2.ckpt']
mfile_l = ['./model_states/' + s for s in file_l]
# mfile_l = ['/home/qsm22/weibo_topic/output/base7/base7ernie.ckpt']
ml = Multi_label(model,mfile_l,llist,n=13)
ml.eval_on_val(val_data,val_dl)