/
sparsemap-monolink-v3.py
222 lines (171 loc) · 8.1 KB
/
sparsemap-monolink-v3.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
import fasttext
import numpy as np
import os
import torch as th
from torch.nn.parameter import Parameter
from torch.optim import SGD
from sklearn.metrics.pairwise import cosine_similarity
from scipy.optimize import linear_sum_assignment
from transformers import BertModel, BertTokenizer
import io
from activeset import ActiveSet
from embedding import *
from dataset import AlignDataset
from emalgos import EmAlgo
from concept import Concept
import utils
import embedding
def get_pair_hashcode(f,e):
return "%s|||%s"%(f,e)
def load_vectors(fname):
fin = io.open(fname, 'r', encoding='utf-8', newline='\n')
n, d = map(int, fin.readline().split())
data = {}
for line in fin:
tokens = line.rstrip().split(' ')
data[tokens[0]] = np.array(list(map(float, tokens[1:])))
return data
def test_alignment(align_f, concept, fs_test, es_test, out_filename, threshld=0.5):
def align_matrix_to_alignment(S,threshld=0):
res = []
for j in range(S.shape[0]):
for i in range(S.shape[1]):
if S[j,i] >= threshld:
res.append('%d-%d'%(j+1,i+1))
return ' '.join(res)
with open(out_filename,'w') as outf:
for f,e in zip(fs_test, es_test):
eta = concept.eta_U(f,e)
S = align_f(eta)
print(align_matrix_to_alignment(S,threshld), file=outf)
def test_alignment_bpe(align_f, concept, fs_test, es_test, out_filename, threshld, src_idx, trg_idx):
temp_file = 'align_temp_bpe.moses'
test_alignment(align_f, concept, fs_test, es_test, temp_file, threshld)
print('align finished')
utils.converse_BPE_to_word(temp_file, src_idx, trg_idx, out_filename)
print('convert bpe to aligndata finished')
def evaluate(align_f, concept, fs_test, es_test, true_label_file, threshld=0.5, bpe=True, src_idx=None, trg_idx=None):
prediction_file = 'evaluate_temp.moses'
if bpe:
if not src_idx or not trg_idx:
print('needs idx data')
return
test_alignment_bpe(align_f, concept, fs_test, es_test, prediction_file, threshld, src_idx, trg_idx)
else:
test_alignment(align_f, concept, fs_test, es_test, prediction_file, threshld)
return utils.compute_aer(prediction_file, true_label_file)
def train_network(dataset, network, optimizer, epochs, batch=100):
print('training ... ')
for e in range(epochs):
print(' epoch : ', e)
for i in range(dataset.sentences_num):
f = dataset.src_sentences[i]
e = dataset.trg_sentences[i]
network([f,e])
network.backward()
if (i+1) % batch == 0:
optimizer.step()
optimizer.zero_grad()
network.embedding.re_compute_sum_exp_norm()
print('OK! training finished ')
def start_by_step_e(dataset, concept, align_f, embedding_model, optimizer, batch=100):
print('start by step e : ')
for i in range(dataset.sentences_num):
f = dataset.src_sentences[i]
e = dataset.trg_sentences[i]
embedding_model((f,e))
eta = concept.eta_U(f,e)
alpha = align_f(eta)
embedding_model.backward(th.Tensor(alpha))
if (i+1) % batch == 0:
print(i)
optimizer.step()
optimizer.zero_grad()
network.embedding.re_compute_sum_exp_norm()
print('OK! start by step e finished : ')
def load_fasttext_embedding_model(f_fasttext_model_file, e_fasttext_model_file):
if os.path.exists(f_fasttext_model_file):
f_model = fasttext.load_model(f_fasttext_model_file)
if os.path.exists(e_fasttext_model_file):
e_model = fasttext.load_model(e_fasttext_model_file)
embedding_model = embedding.FastTextBPEEmbedding(f_model, e_model)
return embedding_model
def main():
# f_fasttext_vec_file = 'wiki.en.align.vec'
# e_fasttext_vec_file = 'wiki.ro.align.vec'
f_fasttext_model_file = 'en-ro-bpe-16K-en.bin'
e_fasttext_model_file = 'en-ro-bpe-16K-ro.bin'
f_fasttext_model_file = 'en_embedding_fasttext_model.bin'
e_fasttext_model_file = 'ro_embedding_fasttext_model.bin'
concept_count_v2 = 'weights/en-ro-bpe-count-p_concept-v2.weight'
# concept_viterbi_v2 = 'weights/en-ro-viterbi-p_concept-v2.weight'
# concept_activeset_v2 = 'weights/en-ro-activeset-p_concept-v2.weight'
concept_init_weight_file = concept_viterbi_v2
concept = Concept(get_pair_hashcode)
concept.load_p_concept(concept_init_weight_file)
emalgo = EmAlgo()
active_set = ActiveSet(20)
train_dataset = AlignDataset(f_train_filename, e_train_filename)
test_dataset = AlignDataset(f_test_filename, e_test_filename)
# f_pretrained_model = load_vectors(f_fasttext_vec_file)
# e_pretrained_model = load_vectors(e_fasttext_vec_file)
# embedding_model = FastTextPretrainedEmbedding(f_pretrained_model, e_pretrained_model)
# embedding_model = load_fasttext_embedding_model(f_fasttext_model_file, e_fasttext_model_file)
# embedding_model = BertEmbedding()
embedding_model = load_fasttext_embedding_model(f_fasttext_model_file, e_fasttext_model_file)
#embedding_model = FastTextEmbedding(f_model, e_model)
network = embedding.Align(embedding_model)
sgd = SGD(network.parameters(), lr=0.0005)
score_before_init = evaluate(emalgo.align_wedding,
concept,
test_dataset.src_sentences,
test_dataset.trg_sentences,
true_label_file,
threshld=0.1,
bpe=True,
src_idx=src_bpe_idx_data,
trg_idx=trg_bpe_idx_data)
print('wedding concept score before init: ', score_before_init)
score_before_init = evaluate(emalgo.align_wedding,
embedding_model,
test_dataset.src_sentences,
test_dataset.trg_sentences,
true_label_file,
threshld=0.1,
bpe=True,
src_idx=src_bpe_idx_data,
trg_idx=trg_bpe_idx_data)
print('wedding embedding score before init : ', score_before_init)
start_by_step_e(train_dataset, concept, emalgo.align_wedding, embedding_model, sgd, 100)
score_after_init = evaluate(active_set.align_sparsemap,
embedding_model,
test_dataset.src_sentences,
test_dataset.trg_sentences,
true_label_file,
threshld=0.2,
bpe=True,
src_idx=src_bpe_idx_data,
trg_idx=trg_bpe_idx_data)
print('sparsemap embedding after step e : ', score_after_init)
sgd = SGD(network.parameters(), lr=0.0001)
train_network(train_dataset, network, sgd, epoch=2, batch=100)
thresholds = [i*0.1 for i in range(1,11)]
score_after_trainings = []
for thr in thresholds:
score_after_training = evaluate(active_set.align_sparsemap,
embedding_model,
test_dataset.src_sentences,
test_dataset.trg_sentences,
true_label_file,
threshld=thr,
bpe=True,
src_idx=src_bpe_idx_data,
trg_idx=trg_bpe_idx_data)
score_after_trainings.append(score_after_training)
print('score after training : ', score_after_trainings)
print('best score after training : ', np.min(score_after_trainings))
out_parameter_file = 'embedding_weight.txt'
np.savetxt('out_parameter_file' , embedding_model.weight.detach().numpy())
print('embedding weight saved to : ', out_parameter_file)
if __name__ == '__main__':
main()