Skip to content

Commit b9351aa

Browse files
author
Suwon Shon
committed
update baseline
1 parent 190f9d3 commit b9351aa

File tree

1 file changed

+291
-0
lines changed

1 file changed

+291
-0
lines changed

‎train_ivector_lda.ipynb

+291
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,291 @@
1+
{
2+
"cells": [
3+
{
4+
"cell_type": "code",
5+
"execution_count": 1,
6+
"metadata": {
7+
"collapsed": true
8+
},
9+
"outputs": [],
10+
"source": [
11+
"from tensorflow.contrib.learn.python.learn.datasets import base\n",
12+
"\n",
13+
"import tensorflow as tf\n",
14+
"import numpy as np\n",
15+
"import os,sys\n",
16+
"sys.path.insert(0, './scripts')\n",
17+
"dataDir ='./data'\n",
18+
"\n",
19+
"\n",
20+
"import py_compile\n",
21+
"py_compile.compile('scripts/ivector_dataset.py')\n",
22+
"py_compile.compile('scripts/ivector_tools.py')\n",
23+
"py_compile.compile('scripts/siamese_model.py')\n",
24+
"import ivector_dataset\n",
25+
"import siamese_model\n",
26+
"import ivector_tools as it"
27+
]
28+
},
29+
{
30+
"cell_type": "code",
31+
"execution_count": 2,
32+
"metadata": {
33+
"collapsed": true
34+
},
35+
"outputs": [],
36+
"source": [
37+
"# write prototxt for siamese network\n",
38+
"\n",
39+
"languages = ['EGY','GLF','LAV','MSA','NOR']\n",
40+
"trn_labels = []\n",
41+
"trn_names = []\n",
42+
"trn_ivectors = np.empty((0,400))\n",
43+
"dev_labels = []\n",
44+
"dev_names = []\n",
45+
"dev_ivectors = np.empty((0,400))\n",
46+
"\n",
47+
"\n",
48+
"for i,lang in enumerate(languages):\n",
49+
" #load train.vardial2017\n",
50+
" filename = dataDir+'/train.vardial2017/%s.ivec' % lang\n",
51+
" name = np.loadtxt(filename,usecols=[0],dtype='string')\n",
52+
" ivector = np.loadtxt(filename,usecols=range(1,401),dtype='float32')\n",
53+
" trn_labels = np.append(trn_labels, np.ones(np.size(name))*(i+1))\n",
54+
" trn_names=np.append(trn_names,name)\n",
55+
" trn_ivectors = np.append(trn_ivectors, ivector,axis=0)\n",
56+
"\n",
57+
" #load dev.vardial2017\n",
58+
" filename = dataDir+'/dev.vardial2017/%s.ivec' % lang\n",
59+
" name = np.loadtxt(filename,usecols=[0],dtype='string')\n",
60+
" ivector = np.loadtxt(filename,usecols=range(1,401),dtype='float32')\n",
61+
" dev_names=np.append(dev_names,name)\n",
62+
" dev_ivectors = np.append(dev_ivectors, ivector,axis=0)\n",
63+
" dev_labels = np.append(dev_labels, np.ones(np.size(name))*(i+1))\n",
64+
" \n",
65+
"# load test.MGB3\n",
66+
"filename = dataDir+'/test.MGB3/ivec_features'\n",
67+
"tst_names = np.loadtxt(filename,usecols=[0],dtype='string')\n",
68+
"tst_ivectors = np.loadtxt(filename,usecols=range(1,401),dtype='float32')\n",
69+
"\n",
70+
"# merge trn+dev\n",
71+
"trndev_ivectors = np.append(trn_ivectors, dev_ivectors,axis=0)\n",
72+
"trndev_labels = np.append(trn_labels,dev_labels)\n",
73+
"trndev_name = np.append(trn_names,dev_names)\n",
74+
"\n",
75+
"\n",
76+
"# load tst.MGB3 labels\n",
77+
"filename = 'data/test.MGB3/reference'\n",
78+
"tst_ref_names = np.loadtxt(filename,usecols=[0],dtype='string')\n",
79+
"tst_ref_labels = np.loadtxt(filename,usecols=[1],dtype='int')\n",
80+
"\n",
81+
"tst_ref_labels_index = []\n",
82+
"for i, name_ref in enumerate(tst_ref_names):\n",
83+
" for j, name in enumerate(tst_names):\n",
84+
" if name == name_ref:\n",
85+
" tst_ref_labels_index = np.append(tst_ref_labels_index,int(j))\n",
86+
"\n",
87+
"tst_labels = tst_ref_labels\n",
88+
"tst_ivectors = tst_ivectors[ map(int,tst_ref_labels_index),:]"
89+
]
90+
},
91+
{
92+
"cell_type": "code",
93+
"execution_count": 3,
94+
"metadata": {},
95+
"outputs": [
96+
{
97+
"name": "stdout",
98+
"output_type": "stream",
99+
"text": [
100+
"((13825, 400), (1524, 400), (5, 400), (1492, 400))\n",
101+
"Final accurary on test dataset : 0.603\n",
102+
"Confusion matrix\n",
103+
"[[ 192. 14. 40. 10. 46.]\n",
104+
" [ 15. 118. 34. 8. 20.]\n",
105+
" [ 65. 83. 221. 16. 102.]\n",
106+
" [ 23. 28. 24. 225. 32.]\n",
107+
" [ 7. 7. 15. 3. 144.]]\n",
108+
"Precision\n",
109+
"[ 0.63576159 0.60512821 0.45379877 0.67771084 0.81818182]\n",
110+
"Recall\n",
111+
"[ 0.63576159 0.472 0.66167665 0.85877863 0.41860465]\n",
112+
"\n",
113+
"\n",
114+
"<Performance evaluation on Test dataset : CDS (baseline) >\n",
115+
"Accurary : 0.603\n",
116+
"Precision : 0.638\n",
117+
"Recall : 0.609\n"
118+
]
119+
}
120+
],
121+
"source": [
122+
"#center and length norm.\n",
123+
"m=np.mean(trn_ivectors,axis=0)\n",
124+
"A = np.cov(trn_ivectors.transpose())\n",
125+
"[a,D,V] = np.linalg.svd(A)\n",
126+
"V= V.transpose()\n",
127+
"W= np.dot(V, np.diag(1./( np.sqrt(D) + 0.0000000001 )))\n",
128+
"\n",
129+
"trn_ivectors = np.dot( np.subtract( trn_ivectors, m), W)\n",
130+
"trndev_ivectors = np.dot( np.subtract( trndev_ivectors, m), W)\n",
131+
"dev_ivectors = np.dot( np.subtract( dev_ivectors, m), W)\n",
132+
"tst_ivectors = np.dot( np.subtract( tst_ivectors, m), W)\n",
133+
"\n",
134+
"trn_ivectors = it.length_norm(trn_ivectors)\n",
135+
"trndev_ivectors = it.length_norm(trndev_ivectors)\n",
136+
"dev_ivectors = it.length_norm(dev_ivectors)\n",
137+
"tst_ivectors = it.length_norm(tst_ivectors)\n",
138+
"\n",
139+
"#language modeling\n",
140+
"lang_mean=[]\n",
141+
"for i, lang in enumerate(languages):\n",
142+
" lang_mean.append(np.mean(np.append(trn_ivectors[np.nonzero(trn_labels == i+1)] ,dev_ivectors[np.nonzero(dev_labels == i+1)],axis=0),axis=0))\n",
143+
"# lang_mean.append(np.mean(trn_ivectors[np.nonzero(trn_labels == i+1)],axis=0))\n",
144+
"\n",
145+
"lang_mean = np.array(lang_mean)\n",
146+
"lang_mean = it.length_norm(lang_mean)\n",
147+
"\n",
148+
"print( np.shape(trn_ivectors), np.shape(dev_ivectors), np.shape(lang_mean),np.shape(tst_ivectors) )\n",
149+
"\n",
150+
"\n",
151+
"tst_scores = lang_mean.dot(tst_ivectors.transpose() )\n",
152+
"# print(tst_scores.shape)\n",
153+
"hypo_lang = np.argmax(tst_scores,axis = 0)\n",
154+
"temp = ((tst_labels-1) - hypo_lang)\n",
155+
"acc =1- np.size(np.nonzero(temp)) / float(np.size(tst_labels))\n",
156+
"print 'Final accurary on test dataset : %0.3f' %(acc)\n",
157+
"\n",
158+
"confusionmat = np.zeros((5,5))\n",
159+
"for i,lang in enumerate(languages):\n",
160+
" hypo_bylang = hypo_lang[ tst_labels == i+1]\n",
161+
" hist_bylang = np.histogram(hypo_bylang,5)\n",
162+
" confusionmat[:,i] = hist_bylang[0]\n",
163+
"\n",
164+
"precision = np.diag(confusionmat) / np.sum(confusionmat,axis=1) #precision\n",
165+
"recall = np.diag(confusionmat) / np.sum(confusionmat,axis=0) # recall\n",
166+
" \n",
167+
"print 'Confusion matrix'\n",
168+
"print confusionmat\n",
169+
"print 'Precision'\n",
170+
"print precision\n",
171+
"print 'Recall'\n",
172+
"print recall\n",
173+
"\n",
174+
"print '\\n\\n<Performance evaluation on Test dataset : CDS (baseline) >'\n",
175+
"print 'Accurary : %0.3f' %(acc)\n",
176+
"print 'Precision : %0.3f' %(np.mean(precision))\n",
177+
"print 'Recall : %0.3f' %(np.mean(recall))"
178+
]
179+
},
180+
{
181+
"cell_type": "code",
182+
"execution_count": 4,
183+
"metadata": {},
184+
"outputs": [
185+
{
186+
"name": "stdout",
187+
"output_type": "stream",
188+
"text": [
189+
"((13825, 4), (1524, 4), (5, 4), (1492, 4))\n",
190+
"Final accurary on test dataset : 0.628\n",
191+
"Confusion matrix\n",
192+
"[[ 200. 22. 46. 13. 40.]\n",
193+
" [ 17. 145. 62. 10. 27.]\n",
194+
" [ 47. 49. 172. 9. 54.]\n",
195+
" [ 22. 23. 26. 224. 27.]\n",
196+
" [ 16. 11. 28. 6. 196.]]\n",
197+
"Precision\n",
198+
"[ 0.62305296 0.55555556 0.51963746 0.69565217 0.76264591]\n",
199+
"Recall\n",
200+
"[ 0.66225166 0.58 0.51497006 0.85496183 0.56976744]\n",
201+
"\n",
202+
"\n",
203+
"<Performance evaluation on Test dataset : LDA+CDS>\n",
204+
"Accurary : 0.628\n",
205+
"Precision : 0.631\n",
206+
"Recall : 0.636\n"
207+
]
208+
}
209+
],
210+
"source": [
211+
"#LDA\n",
212+
"[languages,train_languages_num] = np.unique(trndev_labels,return_inverse=True)\n",
213+
"V = it.lda2(trndev_ivectors,train_languages_num)\n",
214+
"V = np.real(V[:,0:4])\n",
215+
"trn_ivectors = np.matmul(trn_ivectors,V)\n",
216+
"dev_ivectors = np.matmul(dev_ivectors,V)\n",
217+
"tst_ivectors = np.matmul(tst_ivectors,V)\n",
218+
"trndev_ivectors = np.matmul(trndev_ivectors,V)\n",
219+
"\n",
220+
"\n",
221+
"\n",
222+
"trn_ivectors = it.length_norm(trn_ivectors)\n",
223+
"trndev_ivectors = it.length_norm(trndev_ivectors)\n",
224+
"dev_ivectors = it.length_norm(dev_ivectors)\n",
225+
"tst_ivectors = it.length_norm(tst_ivectors)\n",
226+
"\n",
227+
"\n",
228+
"#language modeling\n",
229+
"lang_mean=[]\n",
230+
"for i, lang in enumerate(languages):\n",
231+
" lang_mean.append(np.mean(np.append(trn_ivectors[np.nonzero(trn_labels == i+1)] ,dev_ivectors[np.nonzero(dev_labels == i+1)],axis=0),axis=0))\n",
232+
"# lang_mean.append(np.mean(trn_ivectors[np.nonzero(trn_labels == i+1)],axis=0))\n",
233+
"\n",
234+
"lang_mean = np.array(lang_mean)\n",
235+
"lang_mean = it.length_norm(lang_mean)\n",
236+
"\n",
237+
"print( np.shape(trn_ivectors), np.shape(dev_ivectors), np.shape(lang_mean),np.shape(tst_ivectors) )\n",
238+
"\n",
239+
"\n",
240+
"tst_scores = lang_mean.dot(tst_ivectors.transpose() )\n",
241+
"# print(tst_scores.shape)\n",
242+
"hypo_lang = np.argmax(tst_scores,axis = 0)\n",
243+
"temp = ((tst_labels-1) - hypo_lang)\n",
244+
"acc =1- np.size(np.nonzero(temp)) / float(np.size(tst_labels))\n",
245+
"print 'Final accurary on test dataset : %0.3f' %(acc)\n",
246+
"\n",
247+
"confusionmat = np.zeros((5,5))\n",
248+
"for i,lang in enumerate(languages):\n",
249+
" hypo_bylang = hypo_lang[ tst_labels == i+1]\n",
250+
" hist_bylang = np.histogram(hypo_bylang,5)\n",
251+
" confusionmat[:,i] = hist_bylang[0]\n",
252+
"\n",
253+
"precision = np.diag(confusionmat) / np.sum(confusionmat,axis=1) #precision\n",
254+
"recall = np.diag(confusionmat) / np.sum(confusionmat,axis=0) # recall\n",
255+
" \n",
256+
"print 'Confusion matrix'\n",
257+
"print confusionmat\n",
258+
"print 'Precision'\n",
259+
"print precision\n",
260+
"print 'Recall'\n",
261+
"print recall\n",
262+
"\n",
263+
"print '\\n\\n<Performance evaluation on Test dataset : LDA+CDS>'\n",
264+
"print 'Accurary : %0.3f' %(acc)\n",
265+
"print 'Precision : %0.3f' %(np.mean(precision))\n",
266+
"print 'Recall : %0.3f' %(np.mean(recall))"
267+
]
268+
}
269+
],
270+
"metadata": {
271+
"kernelspec": {
272+
"display_name": "Python 2",
273+
"language": "python",
274+
"name": "python2"
275+
},
276+
"language_info": {
277+
"codemirror_mode": {
278+
"name": "ipython",
279+
"version": 2
280+
},
281+
"file_extension": ".py",
282+
"mimetype": "text/x-python",
283+
"name": "python",
284+
"nbconvert_exporter": "python",
285+
"pygments_lexer": "ipython2",
286+
"version": "2.7.6"
287+
}
288+
},
289+
"nbformat": 4,
290+
"nbformat_minor": 2
291+
}

0 commit comments

Comments
 (0)