-
Notifications
You must be signed in to change notification settings - Fork 0
/
scorer.py
212 lines (177 loc) · 6.71 KB
/
scorer.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
212
#!/usr/bin/python
# -*- coding: utf-8 -*-
"""
Author: Gözde Gül Şahin
Evaluate predicted labels by comparing to gold labels
"""
import torch
from itertools import chain
import sys
use_cuda = torch.cuda.is_available()
dtype = torch.cuda.FloatTensor if use_cuda else torch.FloatTensor
def buildi2s(s2i):
i2s = {}
for key in s2i:
i2s[s2i[key]]=key
return i2s
def evalConll(pred_labels, gold_labels, dummyrole, stoprole, role_to_ix, mode, type):
"""
Run one epoch (either in training or evaluation mode)
:param pred_labels: predicted labels
:param gold_labels: ground truth
:param dummyrole: "_", don't score it
:param stoprole: <STOPTAG> if using CRF sequence model don't score it
:param mode: if "eval", save the predicted and gold labels into a list
:param type: if "seq", remove the start and stop tag
:return: plst: predicted label list
:return: glst: gold label list
:return: num_corr,num_found,num_gold : correctly found, total found argument count, total gold
"""
num_corr = 0.
num_gold = 0.
num_found = 0.
# local lists
plst = []
glst = []
index2role = buildi2s(role_to_ix)
index2role[0] = u"_"
for i in range(pred_labels.size(0)):
found = pred_labels[i,:].contiguous().view(1,-1)
if found.type() != 'torch.cuda.FloatTensor':
found = found.float()
if mode=="eval":
fstr=""
for ind in found[0]:
fstr+=index2role[int(ind)]
fstr+=" "
plst.append(fstr[:-1])
if (type=="seq"):
# remove the start and the stop tag
gold = gold_labels[i,1:-1].contiguous().view(1,-1)
else:
gold = gold_labels[i,:].contiguous().view(1, -1)
if mode=="eval":
gstr=""
for ind in gold[0]:
gstr+=index2role[int(ind)]
gstr+=" "
glst.append(gstr[:-1])
label_mask = gold.ne(dummyrole)&gold.ne(stoprole)
is_corr = (found==gold)*label_mask
num_corr += is_corr.ne(0).sum()
num_gold += label_mask.sum()
num_found += (found.ne(dummyrole)&found.ne(stoprole)).sum()
return plst,glst,num_corr,num_found,num_gold
def testRoleLabels(model, data, role_to_ix, mode="eval", type="simple"):
"""
Basic Inference Code
Given the trained model and gold labels, calculate model's best semantic role predictions and compare to gold
:param model: SRL model
:param data: test data
:param role_to_ix: semantic role dictionary
:param mode:
:param type: can be ignored
:return:
"""
total_corr = 0.
total_found = sys.float_info.epsilon
total_gold = 0.
predictionslst = []
goldlst = []
# get index of non-roles
dummyrole = role_to_ix['_']
if mode == "eval":
model.eval()
for i in range(len(data)):
batch = data[i]
modscore = model(batch[0])
_, mod_tag_seq = torch.max(modscore, 1)
mod_tag_seq = mod_tag_seq.data
mod_tag_seq = mod_tag_seq.view(model.batch_size, -1)
gold_lab = batch[1]
plst, glst, num_corr, num_found, num_gold = evalConll(mod_tag_seq, gold_lab.data.type(dtype), dummyrole,
dummyrole, role_to_ix, mode, type)
predictionslst.append(plst)
goldlst.append(glst)
total_corr += num_corr
total_gold += num_gold
total_found += num_found
predictionslst = list(chain.from_iterable(predictionslst))
goldlst = list(chain.from_iterable(goldlst))
return predictionslst, goldlst, total_corr, total_found, total_gold
def testRoleLabelsEnsemble(models, datas, role_to_ix, mode="eval", type="simple"):
"""
Given multiple models, calculate predictions via average voting and compare to gold
"""
total_corr = 0.
total_found = 0.
total_gold = 0.
predictionslst = []
goldlst = []
# get index of non-roles
dummyrole = role_to_ix['_']
if mode=="eval":
for model in models:
model.eval()
for i in range(len(datas[0])):
log_probs = torch.zeros(datas[0][i][1].size(1), len(role_to_ix)).cuda()
# averaging - voting
for k, model in enumerate(models):
batch = datas[k][i]
lp = model(batch[0]).data
log_probs += lp
avg_log_probs = torch.autograd.Variable(torch.div(log_probs, len(models)), volatile=True)
_, mod_tag_seq = torch.max(avg_log_probs, 1)
mod_tag_seq = mod_tag_seq.data
mod_tag_seq = mod_tag_seq.view(1, -1)
# gold labels
plst, glst, num_corr, num_found, num_gold = evalConll(mod_tag_seq, datas[0][i][1].data.type(dtype), dummyrole,
dummyrole, role_to_ix, mode, type)
predictionslst.append(plst)
goldlst.append(glst)
total_corr += num_corr
total_gold += num_gold
total_found += num_found
predictionslst = list(chain.from_iterable(predictionslst))
goldlst = list(chain.from_iterable(goldlst))
return predictionslst, goldlst, total_corr, total_found, total_gold
def testRoleLabelsEnsembleLearner(models, ensmodel, datas, role_to_ix, mode="eval", type="simple"):
"""
Stack Generalizer (ensmodel) learns to weigh the predictions of other SRL models
Calculate SG's prediction given other models' predictions
:param models: Pretrained SRL models
:param ensmodel: Stack Generalizer model
"""
total_corr = 0.
total_found = 0.
total_gold = 0.
predictionslst = []
goldlst = []
# get index of non-roles
dummyrole = role_to_ix['_']
if mode=="eval":
for model in models:
model.eval()
ensmodel.eval()
for i in range(len(datas[0])):
log_probs = []
for z, model in enumerate(models):
batch = datas[z][i]
lp = model(batch[0])
log_probs += [lp]
input = torch.stack(log_probs)
final_log_probs = ensmodel(input)
_, mod_tag_seq = torch.max(final_log_probs, 1)
mod_tag_seq = mod_tag_seq.data
mod_tag_seq = mod_tag_seq.view(1, -1)
# gold labels
plst, glst, num_corr, num_found, num_gold = evalConll(mod_tag_seq, datas[0][i][1].data.type(dtype), dummyrole,
dummyrole, role_to_ix, mode, type)
predictionslst.append(plst)
goldlst.append(glst)
total_corr += num_corr
total_gold += num_gold
total_found += num_found
predictionslst = list(chain.from_iterable(predictionslst))
goldlst = list(chain.from_iterable(goldlst))
return predictionslst, goldlst, total_corr, total_found, total_gold