-
Notifications
You must be signed in to change notification settings - Fork 0
/
top_pair_learning.py
247 lines (188 loc) · 10.4 KB
/
top_pair_learning.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
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
#coding=utf8
import sys
import os
import json
import random
import numpy
import timeit
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
import torch.autograd as autograd
import torchvision.transforms as T
import torch.optim as optim
from torch.optim import lr_scheduler
from sklearn.metrics import accuracy_score, average_precision_score, precision_score,recall_score
from conf import *
import DataReader
import evaluation
import net as network
import performance
import cPickle
sys.setrecursionlimit(1000000)
print >> sys.stderr, "PID", os.getpid()
torch.cuda.set_device(args.gpu)
def net_copy(net,copy_from_net):
mcp = list(net.parameters())
mp = list(copy_from_net.parameters())
n = len(mcp)
for i in range(0, n):
mcp[i].data[:] = mp[i].data[:]
def main():
DIR = args.DIR
embedding_file = args.embedding_dir
best_network_file = "./model/network_model_pretrain.best"
print >> sys.stderr,"Read model from",best_network_file
best_network_model = torch.load(best_network_file)
embedding_matrix = numpy.load(embedding_file)
"Building torch model"
network_model = network.Network(nnargs["pair_feature_dimention"],nnargs["mention_feature_dimention"],nnargs["word_embedding_dimention"],nnargs["span_dimention"],1000,nnargs["embedding_size"],nnargs["embedding_dimention"],embedding_matrix).cuda()
print >> sys.stderr,"save model ..."
net_copy(network_model,best_network_model)
reduced=""
if args.reduced == 1:
reduced="_reduced"
print >> sys.stderr,"prepare data for train ..."
train_docs = DataReader.DataGnerater("train"+reduced)
print >> sys.stderr,"prepare data for dev and test ..."
dev_docs = DataReader.DataGnerater("dev"+reduced)
test_docs = DataReader.DataGnerater("test"+reduced)
l2_lambda = 1e-6
lr = nnargs["lr"]
dropout_rate = nnargs["dropout_rate"]
epoch = nnargs["epoch"]
model_save_dir = "./model/bp/"
last_cost = 0.0
all_best_results = {
'thresh': 0.0,
'accuracy': 0.0,
'precision': 0.0,
'recall': 0.0,
'f1': 0.0
}
optimizer = optim.RMSprop(network_model.parameters(), lr=lr, eps=1e-5)
scheduler = lr_scheduler.StepLR(optimizer, step_size=75, gamma=0.5)
for echo in range(epoch):
start_time = timeit.default_timer()
print "Pretrain Epoch:",echo
scheduler.step()
pair_cost_this_turn = 0.0
ana_cost_this_turn = 0.0
pair_nums = 0
ana_nums = 0
for data in train_docs.train_generater(shuffle=True):
mention_index = autograd.Variable(torch.from_numpy(data["mention_word_index"]).type(torch.cuda.LongTensor))
mention_span = autograd.Variable(torch.from_numpy(data["mention_span"]).type(torch.cuda.FloatTensor))
candi_index = autograd.Variable(torch.from_numpy(data["candi_word_index"]).type(torch.cuda.LongTensor))
candi_spans = autograd.Variable(torch.from_numpy(data["candi_span"]).type(torch.cuda.FloatTensor))
pair_feature = autograd.Variable(torch.from_numpy(data["pair_features"]).type(torch.cuda.FloatTensor))
anaphors = autograd.Variable(torch.from_numpy(data["pair_anaphors"]).type(torch.cuda.LongTensor))
antecedents = autograd.Variable(torch.from_numpy(data["pair_antecedents"]).type(torch.cuda.LongTensor))
anaphoricity_index = autograd.Variable(torch.from_numpy(data["mention_word_index"]).type(torch.cuda.LongTensor))
anaphoricity_span = autograd.Variable(torch.from_numpy(data["mention_span"]).type(torch.cuda.FloatTensor))
anaphoricity_feature = autograd.Variable(torch.from_numpy(data["anaphoricity_feature"]).type(torch.cuda.FloatTensor))
reindex = autograd.Variable(torch.from_numpy(data["top_score_index"]).type(torch.cuda.LongTensor))
start_index = autograd.Variable(torch.from_numpy(data["top_starts"]).type(torch.cuda.LongTensor))
end_index = autograd.Variable(torch.from_numpy(data["top_ends"]).type(torch.cuda.LongTensor))
top_gold = autograd.Variable(torch.from_numpy(data["top_gold"]).type(torch.cuda.FloatTensor))
anaphoricity_target = data["anaphoricity_target"]
anaphoricity_gold = anaphoricity_target.tolist()
ana_lable = autograd.Variable(torch.cuda.FloatTensor([anaphoricity_gold]))
optimizer.zero_grad()
output,output_reindex = network_model.forward_top_pair(nnargs["word_embedding_dimention"],mention_index,mention_span,candi_index,candi_spans,pair_feature,anaphors,antecedents,reindex,start_index,end_index,dropout_rate)
loss = F.binary_cross_entropy(output,top_gold,size_average=False)/train_docs.scale_factor_top
ana_output,_,_ = network_model.forward_anaphoricity(nnargs["word_embedding_dimention"], anaphoricity_index, anaphoricity_span, anaphoricity_feature, dropout_rate)
ana_loss = F.binary_cross_entropy(ana_output,ana_lable,size_average=False)/train_docs.anaphoricity_scale_factor_top
loss_all = loss + ana_loss
loss_all.backward()
pair_cost_this_turn += loss.data[0]
optimizer.step()
end_time = timeit.default_timer()
print >> sys.stderr, "PreTrain",echo,"Pair total cost:",pair_cost_this_turn
print >> sys.stderr, "PreTRAINING Use %.3f seconds"%(end_time-start_time)
print >> sys.stderr, "Learning Rate",lr
gold = []
predict = []
ana_gold = []
ana_predict = []
for data in dev_docs.train_generater(shuffle=False):
mention_index = autograd.Variable(torch.from_numpy(data["mention_word_index"]).type(torch.cuda.LongTensor))
mention_span = autograd.Variable(torch.from_numpy(data["mention_span"]).type(torch.cuda.FloatTensor))
candi_index = autograd.Variable(torch.from_numpy(data["candi_word_index"]).type(torch.cuda.LongTensor))
candi_spans = autograd.Variable(torch.from_numpy(data["candi_span"]).type(torch.cuda.FloatTensor))
pair_feature = autograd.Variable(torch.from_numpy(data["pair_features"]).type(torch.cuda.FloatTensor))
anaphors = autograd.Variable(torch.from_numpy(data["pair_anaphors"]).type(torch.cuda.LongTensor))
antecedents = autograd.Variable(torch.from_numpy(data["pair_antecedents"]).type(torch.cuda.LongTensor))
anaphoricity_index = autograd.Variable(torch.from_numpy(data["mention_word_index"]).type(torch.cuda.LongTensor))
anaphoricity_span = autograd.Variable(torch.from_numpy(data["mention_span"]).type(torch.cuda.FloatTensor))
anaphoricity_feature = autograd.Variable(torch.from_numpy(data["anaphoricity_feature"]).type(torch.cuda.FloatTensor))
reindex = autograd.Variable(torch.from_numpy(data["top_score_index"]).type(torch.cuda.LongTensor))
start_index = autograd.Variable(torch.from_numpy(data["top_starts"]).type(torch.cuda.LongTensor))
end_index = autograd.Variable(torch.from_numpy(data["top_ends"]).type(torch.cuda.LongTensor))
top_gold = autograd.Variable(torch.from_numpy(data["top_gold"]).type(torch.cuda.FloatTensor))
anaphoricity_target = data["anaphoricity_target"]
anaphoricity_gold = anaphoricity_target.tolist()
ana_lable = autograd.Variable(torch.cuda.FloatTensor([anaphoricity_gold]))
gold += data["top_gold"].tolist()
ana_gold += anaphoricity_target.tolist()
output,output_reindex = network_model.forward_top_pair(nnargs["word_embedding_dimention"],mention_index,mention_span,candi_index,candi_spans,pair_feature,anaphors,antecedents,reindex,start_index,end_index,0.0)
predict += output.data.cpu().numpy().tolist()
ana_output,_,_ = network_model.forward_anaphoricity(nnargs["word_embedding_dimention"], anaphoricity_index, anaphoricity_span, anaphoricity_feature, 0.0)
ana_predict += ana_output.data.cpu().numpy()[0].tolist()
gold = numpy.array(gold,dtype=numpy.int32)
predict = numpy.array(predict)
best_results = {
'thresh': 0.0,
'accuracy': 0.0,
'precision': 0.0,
'recall': 0.0,
'f1': 0.0
}
thresh_list = [0.3,0.35,0.4,0.45,0.5,0.55,0.6]
for thresh in thresh_list:
evaluation_results = get_metrics(gold, predict, thresh)
if evaluation_results["f1"] >= best_results["f1"]:
best_results = evaluation_results
print "Pair accuracy: %f and Fscore: %f with thresh: %f"\
%(best_results["accuracy"],best_results["f1"],best_results["thresh"])
sys.stdout.flush()
if best_results["f1"] >= all_best_results["f1"]:
all_best_results = best_results
print >> sys.stderr, "New High Result, Save Model"
torch.save(network_model, model_save_dir+"network_model_pretrain.best.top")
ana_gold = numpy.array(ana_gold,dtype=numpy.int32)
ana_predict = numpy.array(ana_predict)
best_results = {
'thresh': 0.0,
'accuracy': 0.0,
'precision': 0.0,
'recall': 0.0,
'f1': 0.0
}
for thresh in thresh_list:
evaluation_results = get_metrics(ana_gold, ana_predict, thresh)
if evaluation_results["f1"] >= best_results["f1"]:
best_results = evaluation_results
print "Anaphoricity accuracy: %f and Fscore: %f with thresh: %f"\
%(best_results["accuracy"],best_results["f1"],best_results["thresh"])
sys.stdout.flush()
if (echo+1)%10 == 0:
best_network_model = torch.load(model_save_dir+"network_model_pretrain.best.top")
print "DEV:"
performance.performance(dev_docs,best_network_model)
print "TEST:"
performance.performance(test_docs,best_network_model)
def get_metrics(gold, predict, thresh):
pred = np.clip(np.floor(predict / thresh), 0, 1)
p, r = (0, 0) if pred.sum() == 0 else \
(precision_score(gold, pred), recall_score(gold, pred))
return {
'thresh': thresh,
'accuracy': average_precision_score(gold, predict),
'precision': p,
'recall': r,
'f1': 0 if p == 0 or r == 0 else 2 * p * r / (p + r)
}
if __name__ == "__main__":
main()