-
Notifications
You must be signed in to change notification settings - Fork 14
/
evaluation.py
212 lines (185 loc) · 8.1 KB
/
evaluation.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
import torch
import numpy as np
import datetime
from spodernet.utils.global_config import Config
from spodernet.utils.cuda_utils import CUDATimer
from spodernet.utils.logger import Logger
from torch.autograd import Variable
from sklearn import metrics
#timer = CUDATimer()
log = Logger('evaluation{0}.py.txt'.format(datetime.datetime.now()))
def ranking_and_hits(model, dev_rank_batcher, vocab, name, epoch, test):
log.info(name)
log.info("Evaluation")
if test is True:
fil = open(name + ".txt","a")
hits_left = []
hits_right = []
hits = []
ranks = []
ranks_left = []
ranks_right = []
for i in range(10):
hits_left.append([])
hits_right.append([])
hits.append([])
for i, str2var in enumerate(dev_rank_batcher):
e1 = str2var['e1']
e2 = str2var['e2']
rel = str2var['rel']
rel_reverse = str2var['rel_eval']
e2_multi1 = str2var['e2_multi1'].float()
e2_multi2 = str2var['e2_multi2'].float()
pred1 = model.forward(e1, rel)
pred2 = model.forward(e2, rel_reverse)
pred1, pred2 = pred1.data, pred2.data
e1, e2 = e1.data, e2.data
e2_multi1, e2_multi2 = e2_multi1.data, e2_multi2.data
# print(e1[0])
for i in range(Config.batch_size):
# these filters contain ALL labels
filter1 = e2_multi1[i].long()
filter2 = e2_multi2[i].long()
num = e1[i, 0].item()
# save the prediction that is relevant
target_value1 = pred1[i,e2[i, 0].item()].item()
target_value2 = pred2[i,e1[i, 0].item()].item()
# zero all known cases (this are not interesting)
# this corresponds to the filtered setting
pred1[i][filter1] = 0.0
pred2[i][filter2] = 0.0
# write base the saved values
pred1[i][e2[i]] = target_value1
pred2[i][e1[i]] = target_value2
# print(len(pred1[0]))
# sort and rank
max_values, argsort1 = torch.sort(pred1, 1, descending=True)
max_values, argsort2 = torch.sort(pred2, 1, descending=True)
argsort1 = argsort1.cpu().numpy()
argsort2 = argsort2.cpu().numpy()
for i in range(Config.batch_size):
# find the rank of the target entities
rank1 = np.where(argsort1[i]==e2[i, 0].item())[0][0]
rank2 = np.where(argsort2[i]==e1[i, 0].item())[0][0]
# rank+1, since the lowest rank is rank 1 not rank 0
ranks.append(rank1+1)
ranks_left.append(rank1+1)
ranks.append(rank2+1)
ranks_right.append(rank2+1)
# this could be done more elegantly, but here you go
for hits_level in range(10):
if rank1 <= hits_level:
hits[hits_level].append(1.0)
hits_left[hits_level].append(1.0)
else:
hits[hits_level].append(0.0)
hits_left[hits_level].append(0.0)
if rank2 <= hits_level:
hits[hits_level].append(1.0)
hits_right[hits_level].append(1.0)
else:
hits[hits_level].append(0.0)
hits_right[hits_level].append(0.0)
dev_rank_batcher.state.loss = [0]
for i in [0,2,9]:
#log.info('Hits left @{0}: {1}'.format(i+1, np.mean(hits_left[i])))
#log.info('Hits right @{0}: {1}'.format(i+1, np.mean(hits_right[i])))
log.info('Hits @{0}: {1}'.format(i+1, np.mean(hits[i])))
if test is True: fil.write('Hits @{0}: {1}'.format(i+1, np.mean(hits[i])))
#log.info('Mean rank left: {0}', np.mean(ranks_left))
#log.info('Mean rank right: {0}', np.mean(ranks_right))
log.info('Mean rank: {0}', np.mean(ranks))
if test is True: fil.write('Mean rank: {0}'.format(np.mean(ranks)))
#log.info('Mean reciprocal rank left: {0}', np.mean(1./np.array(ranks_left)))
#log.info('Mean reciprocal rank right: {0}', np.mean(1./np.array(ranks_right)))
log.info('Mean reciprocal rank: {0}', np.mean(1./np.array(ranks)))
if test is True: fil.write('Mean reciprocal rank: {0}'.format(np.mean(1./np.array(ranks))))
if test is True: fil.close()
return ranks
def ranking_and_hits2(model, dev_rank_batcher, vocab, name, epoch, test):
log.info(name)
log.info("Evaluation")
if test is True:
fil = open(name + ".txt", "a")
hits_left = []
hits_right = []
hits = []
ranks = []
ranks_left = []
ranks_right = []
for i in range(10):
hits_left.append([])
hits_right.append([])
hits.append([])
for i, str2var in enumerate(dev_rank_batcher):
e1 = str2var['e1']
e2 = str2var['e2']
rel = str2var['rel']
rel_reverse = str2var['rel_eval']
# e2_multi1 = str2var['e2_multi1'].float()
# e2_multi2 = str2var['e2_multi2'].float()
pred1 = model.forward(e1, rel)
pred2 = model.forward(e2, rel_reverse)
pred1, pred2 = pred1.data, pred2.data
e1, e2 = e1.data, e2.data
# e2_multi1, e2_multi2 = e2_multi1.data, e2_multi2.data
for i in range(Config.batch_size):
# these filters contain ALL labels
# filter1 = e2_multi1[i].long()
# filter2 = e2_multi2[i].long()
num = e1[i, 0].item()
# save the prediction that is relevant
target_value1 = pred1[i,e2[i, 0].item()].item()
target_value2 = pred2[i,e1[i, 0].item()].item()
# zero all known cases (this are not interesting)
# this corresponds to the filtered setting
# pred1[i][filter1] = 0.0
# pred2[i][filter2] = 0.0
# write base the saved values
pred1[i][e2[i]] = target_value1
pred2[i][e1[i]] = target_value2
# print(len(pred1)[0])
# sort and rank
max_values, argsort1 = torch.sort(pred1, 1, descending=True)
max_values, argsort2 = torch.sort(pred2, 1, descending=True)
argsort1 = argsort1.cpu().numpy()
argsort2 = argsort2.cpu().numpy()
for i in range(Config.batch_size):
# find the rank of the target entities
rank1 = np.where(argsort1[i]==e2[i, 0])[0][0]
rank2 = np.where(argsort2[i]==e1[i, 0])[0][0]
# rank+1, since the lowest rank is rank 1 not rank 0
ranks.append(rank1+1)
ranks_left.append(rank1+1)
ranks.append(rank2+1)
ranks_right.append(rank2+1)
# this could be done more elegantly, but here you go
for hits_level in range(10):
if rank1 <= hits_level:
hits[hits_level].append(1.0)
hits_left[hits_level].append(1.0)
else:
hits[hits_level].append(0.0)
hits_left[hits_level].append(0.0)
if rank2 <= hits_level:
hits[hits_level].append(1.0)
hits_right[hits_level].append(1.0)
else:
hits[hits_level].append(0.0)
hits_right[hits_level].append(0.0)
dev_rank_batcher.state.loss = [0]
for i in [0, 2, 9]:
# log.info('Hits left @{0}: {1}'.format(i+1, np.mean(hits_left[i])))
# log.info('Hits right @{0}: {1}'.format(i+1, np.mean(hits_right[i])))
log.info('Hits @{0}: {1}\n'.format(i + 1, np.mean(hits[i])))
if test is True: fil.write('Hits @{0}: {1}\n'.format(i + 1, np.mean(hits[i])))
# log.info('Mean rank left: {0}', np.mean(ranks_left))
# log.info('Mean rank right: {0}', np.mean(ranks_right))
log.info('Mean rank: {0}\n', np.mean(ranks))
if test is True: fil.write('Mean rank: {0}\n'.format(np.mean(ranks)))
# log.info('Mean reciprocal rank left: {0}', np.mean(1./np.array(ranks_left)))
# log.info('Mean reciprocal rank right: {0}', np.mean(1./np.array(ranks_right)))
log.info('Mean reciprocal rank: {0}\n', np.mean(1. / np.array(ranks)))
if test is True: fil.write('Mean reciprocal rank: {0}\n'.format(np.mean(1. / np.array(ranks))))
if test is True: fil.close()
return ranks