-
Notifications
You must be signed in to change notification settings - Fork 364
/
Copy pathtest_aug_multimodel.py
211 lines (188 loc) · 9.29 KB
/
test_aug_multimodel.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
209
210
211
#encoding:utf-8
'''
测试时对数据进行增强，分别元数据 shuffle drop三种
情况下结果，通过hyper search进行融合
'''
from utils import get_score
from torch.utils import data
import torch as t
import numpy as np
from config import opt
import models
import json
import os
import sys
import fire
from glob import glob
import csv
import tqdm
import pickle
from torch.autograd import Variable
def load_data(type_='char'):
with open(opt.labels_path) as f:
labels_ = json.load(f)
question_d = np.load(opt.test_data_path)
print "test_path: ",opt.test_data_path
index2qid = question_d['index2qid'].item()
index=np.arange(len(question_d['title_char']))
np.random.shuffle(index)
char_data=(question_d['title_char'],question_d['content_char'])
word_data=(question_d['title_word'],question_d['content_word'])
return char_data,word_data,index2qid,labels_,index
def write_csv(result,index2qid,labels):
f=open(opt.result_path, "wa")
csv_writer = csv.writer(f, dialect="excel")
rows=[0 for _ in range(result.size(0))]
for i in range(result.size(0)):
tmp=result[i].sort(dim=0,descending=True)
tmp=tmp[1][:5]
row=[index2qid[i]]+[labels[str(int(i_))] for i_ in tmp]
rows[i]=row
csv_writer.writerows(rows)
def dotest(model,title,content):
title,content = (Variable(t.from_numpy(title[0]).long().cuda(),volatile=True),Variable(t.from_numpy(title[1]).long().cuda(),volatile=True)),(Variable(t.from_numpy(content[0]).long().cuda(),volatile=True),Variable(t.from_numpy(content[1]).long().cuda(),volatile=True))
score = model(title,content)
probs=t.nn.functional.sigmoid(score)
return probs.data.cpu().numpy()
def dropout(d,p=0.5):
len_ = d.shape[1]
batch_=d.shape[0]
for ii in range(batch_):
index = np.random.choice(len_,int(len_*p))
d[ii,index]=0
return d
def test_val():
#####
#####测试增强前的验证集结果
#####
data_path='/data_ssd/zhihu/result/tmp/'
result_path=['inception0.41254_aug_word_val.pth','LSTMText0.41368_aug_word_val.pth','DeepText0.38738_aug_char_val.pth',
' RCNN0.39854_aug_char_val.pth','RCNN0.41344_aug_word_val.pth']#,'LSTMText0.4120_aug_word_val.pth']
result=0
for i in range(len(result_path)):
path=data_path+result_path[i]
print "loading",path
result +=t.load(path.replace(' ','')).float()
test_data_path='/home/a/code/pytorch/zhihu/ddd/val.npz'
index2qid = np.load(test_data_path)['index2qid'].item()
label_path="/home/a/code/pytorch/zhihu/ddd/labels.json"
with open(label_path) as f:
labels_info = json.load(f)
qid2label = labels_info['d']
true_labels = [qid2label[index2qid[2999967-200000+ii]] for ii in range(200000)]
result_ = result.topk(5,1)[1]
predict_label_and_marked_label_list = [[_1,_2] for _1,_2 in zip(result_,true_labels)]
score,_,_,_ = get_score(predict_label_and_marked_label_list)
print (score,_)
def test_val_aug():
#####
#####测试增强后的验证集结果
#####
data_path="/data_ssd/zhihu/result/test_aug/"
result_path=glob(data_path+"*.npz")
result_path.sort()
result=0
for i in range(len(result_path)):
print "loading",result_path[i]
result +=np.load(result_path[i])['result_prob']
test_data_path='/home/a/code/pytorch/zhihu/ddd/val.npz'
index2qid = np.load(test_data_path)['index2qid'].item()
label_path="/home/a/code/pytorch/zhihu/ddd/labels.json"
with open(label_path) as f:
labels_info = json.load(f)
qid2label = labels_info['d']
true_labels = [qid2label[index2qid[2999967-200000+ii]] for ii in range(200000)]
result_ = t.from_numpy(result).topk(5,1)[1]
predict_label_and_marked_label_list = [[_1,_2] for _1,_2 in zip(result_,true_labels)]
score,_,_,_ = get_score(predict_label_and_marked_label_list)
print (score,_)
def main(**kwargs):
opt.parse(kwargs)
####################### MultiModelAll_word_0.417185977233#################
opt.model_names=['MultiCNNTextBNDeep','LSTMText','CNNText_inception','RCNN']
opt.model_paths = ['checkpoints/MultiCNNTextBNDeep_0.37125473788','checkpoints/LSTMText_word_0.381833388089','checkpoints/CNNText_tmp_0.376364647145','checkpoints/RCNN_char_0.3456599248']
#########################################################################
model = getattr(models,opt.model)(opt).cuda().eval()
if opt.model_path is not None:
model.load(opt.model_path)
opt.parse(kwargs)
model = model.eval()
test_data_title,test_data_content,index2qid,labels,index=load_data(type_=opt.type_)
Num=len(test_data_title[0])
print "Num: ",Num
result1=np.zeros((Num,1999))
result2=np.zeros((Num,1999))
result3=np.zeros((Num,1999))
for i in tqdm.tqdm(range(Num)):
if i%opt.batch_size==0 and i>0:
# import ipdb;ipdb.set_trace()
title1=np.array(test_data_title[0][i-opt.batch_size:i]),np.array(test_data_title[1][i-opt.batch_size:i])
content1=np.array(test_data_content[0][i-opt.batch_size:i]),np.array(test_data_content[1][i-opt.batch_size:i])
result1[i-opt.batch_size:i,:]=dotest(model,title1,content1)
title3=dropout(title1[0],0.3),dropout(title1[1],0.3)
content3=dropout(content1[0],0.7),dropout(content1[1],0.7)
result3[i-opt.batch_size:i,:]=dotest(model,title3,content3)
title2=np.array(test_data_title[0][index[i-opt.batch_size:i]]),np.array(test_data_title[1][index[i-opt.batch_size:i]])
content2=np.array(test_data_content[0][index[i-opt.batch_size:i]]),np.array(test_data_content[1][index[i-opt.batch_size:i]])
result2[index[i-opt.batch_size:i]]=dotest(model,title2,content2)
if Num%opt.batch_size!=0:
title1=np.array(test_data_title[0][opt.batch_size*(Num/opt.batch_size):]),np.array(test_data_title[1][opt.batch_size*(Num/opt.batch_size):])
content1=np.array(test_data_content[0][opt.batch_size*(Num/opt.batch_size):]),np.array(test_data_content[1][opt.batch_size*(Num/opt.batch_size):])
result1[opt.batch_size*(Num/opt.batch_size):,:]=dotest(model,title1,content1)
title3=dropout(title1[0],0.3),dropout(title1[1],0.3)
content3=dropout(content1[0],0.7),dropout(content1[1],0.7)
result3[opt.batch_size*(Num/opt.batch_size):]=dotest(model,title3,content3)
title2=np.array(test_data_title[0][index[opt.batch_size*(Num/opt.batch_size):]]),np.array(test_data_title[1][index[opt.batch_size*(Num/opt.batch_size):]])
content2=np.array(test_data_content[0][index[opt.batch_size*(Num/opt.batch_size):]]),np.array(test_data_content[1][index[opt.batch_size*(Num/opt.batch_size):]])
result2[index[opt.batch_size*(Num/opt.batch_size):]]=dotest(model,title2,content2)
#######
#hyper search
######
if opt.val:
probs=[result1,result2,result3]
def target(args):
r=0
for r_,k_ in zip(args,probs):
r=r+r_*k_
result = t.from_numpy(r).topk(5,1)[1]
predict_label_and_marked_label_list = [[_1,_2] for _1,_2 in zip(result,true_labels)]
score,_,_,_ = get_score(predict_label_and_marked_label_list)
print (args,score,_)#list_space = [hp.uniform('a',0,1),hp.uniform('b',0,1)]
return -score
true_labels = [labels['d'][index2qid[2999967-200000+ii]] for ii in range(200000)]
from hyperopt import hp, fmin, rand, tpe, space_eval
list_space = [hp.normal('a0',1,0.8),hp.normal('a1',0.5,0.3),hp.normal('a2',0.2,0.2)]
from hyperopt import Trials
trials_to_keep=Trials()
best = fmin(target,list_space,algo=tpe.suggest,max_evals=100, trials = trials_to_keep)
best_params=[best['a0'],best['a1'],best['a2']]
sums=best['a0']+best['a1']+best['a2']
best_score=trials_to_keep.best_trial['result']['loss']*(-1)
result=(result1*best['a0']+result2*best['a1']+result3*best['a2'])/sums
output = open('trials_to_keep_'+opt.model+"_"+str(best_score)+'.pkl', 'wb')
pickle.dump(trials_to_keep, output)
npz_file='/data_ssd/zhihu/result/test_aug/'+opt.model+"_"+str(best_score)+'.npz'
np.savez_compressed(npz_file, weights=np.array(best_params), result_prob=result)
else:
a1=1.08507012#1.1419278470112755
a2=-1.04290239#-1.0272661867201016
a3=1.16344534#1.1450416449670469
result=(a1*result1+a2*result2+a3*result3)/(a1+a2+a3)
labels__=labels['id2label']
#result_top=t.from_numpy(result).topk(5,1)[1]
write_csv(t.from_numpy(result),index2qid,labels__)
#t.save(t.from_numpy(result1).float(),"test_aug_1.pth")
#t.save(t.from_numpy(result2).float(),"test_aug_2.pth")
#t.save(t.from_numpy(result3).float(),"test_aug_3.pth")
#if opt.val:
# true_labels = [labels[index2qid[2999967-200000+ii]] for ii in range(len(test_data_title))]
# result_ = t.from_numpy(result).topk(5,1)[1]
# predict_label_and_marked_label_list = [[_1,_2] for _1,_2 in zip(result_,true_labels)]
# score,_,_,_ = get_score(predict_label_and_marked_label_list)
# print score
del result1
del result2
del result3
del result
if __name__=='__main__':
fire.Fire()