-
Notifications
You must be signed in to change notification settings - Fork 6
/
util.py
413 lines (328 loc) · 12.5 KB
/
util.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
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
import numpy as np
import scipy.io
from scipy.sparse import csr_matrix, coo_matrix
from gensim.matutils import Sparse2Corpus, corpus2csc
from gensim.models import TfidfModel
from sklearn.metrics import pairwise_distances
from sklearn.preprocessing import normalize
from sklearn.model_selection import train_test_split
def compute_tfidf(X_train, X_test):
"""
Compute TF-IDF vectors
It uses only training samples to compute IDF weights
Parameters
----------
X_train : numpy.array
BOW vectors of training samples
Shape: (n, d), where n is the number of training documents, d is the size of the vocabulary
X[i, j] is the number of occurences of word j in document i
X_test : numpy.array
BOW vectors of test samples
Shape: (m, d), where m is the number of test documents, d is the size of the vocabulary
X[i, j] is the number of occurences of word j in document i
Returns
-------
X_train : numpy.array
TF-TDF vectors of training samples
Shape: (n, d), where n is the number of training documents, d is the size of the vocabulary
X_test : numpy.array
BOW vectors of test samples
Shape: (m, d), where m is the number of test documents, d is the size of the vocabulary
"""
corpus = Sparse2Corpus(X_train, documents_columns=False)
model = TfidfModel(corpus, normalize=False)
X_train = csr_matrix(corpus2csc(model[corpus], num_terms=X_train.shape[1]).T)
corpus = Sparse2Corpus(X_test, documents_columns=False)
X_test = csr_matrix(corpus2csc(model[corpus], num_terms=X_train.shape[1]).T)
return X_train, X_test
#####################
# #
# our re-evaluation #
# #
#####################
def knn_evaluation(y_train, y_test, D, k):
"""
Compute kNN accuracy
Parameters
----------
y_train : numpy.array
Labels of training samples
Shape: (n,), where n is the number of training documents
y[i] is the label of document i
y_test : numpy.array
Labels of test samples
Shape: (m,), where m is the number of test documents
y[i] is the label of document i
D : numpy.array
Distance matrix of training and test samples
Shape: (n, m), where n is the number of training documents, m is the number of test documents
D[i, j] is the distance between training document i and test document j
k : int
Size of neighborhood in kNN classification
Returns
-------
acc : float
Accuracy
"""
acc = 0
for i in range(y_test.shape[0]):
rank = np.argsort(D[i])
if np.bincount(y_train[rank[:k]]).argmax() == y_test[i]:
acc += 1
acc = acc / y_test.shape[0]
return acc
def select_k(y_train, D_train):
"""
Select the hyperparameter k using validation data
Parameters
----------
y_train : numpy.array
Labels of training samples
Shape: (n,), where n is the number of training documents
y[i] is the label of document i
D_train : numpy.array
Distance matrix of training samples
Shape: (n, n), where n is the number of training documents
D[i, j] is the distance between training documents i and j
Returns
-------
best_k : int
Chosen hyperparamter k
"""
train, validation = train_test_split(np.arange(len(y_train)), test_size=0.2, random_state=0)
best_score = None
best_k = None
for k in range(1, 20):
score = knn_evaluation(y_train[train], y_train[validation], D_train[validation][:, train], k)
if best_score is None or score > best_score:
best_score = score
best_k = k
return best_k
def evaluate_D(y_train, y_test, D, D_train):
"""
Evaluation using distance metrices
Parameters
----------
y_train : numpy.array
Labels of training samples
Shape: (n,), where n is the number of training documents
y[i] is the label of document i
y_test : numpy.array
Labels of test samples
Shape: (m,), where m is the number of test documents
y[i] is the label of document i
D : numpy.array
Distance matrix of training and test samples
Shape: (n, m), where n is the number of training documents, m is the number of test documents
D[i, j] is the distance between training document i and test document j
D_train : numpy.array
Distance matrix of training samples
Shape: (n, n), where n is the number of training documents
D[i, j] is the distance between training documents i and j
Returns
-------
acc : float
Accuracy
"""
k = select_k(y_train, D_train)
return knn_evaluation(y_train, y_test, D, k)
def evaluate_onehot(X_train, y_train, X_test, y_test, tfidf=False, norm='l1', metric='l1'):
"""
Evaluation using onhot vectors
Parameters
----------
X_train : numpy.array
BOW vectors of training samples
Shape: (n, d), where n is the number of training documents, d is the size of the vocabulary
X[i, j] is the number of occurences of word j in document i
y_train : numpy.array
Labels of training samples
Shape: (n,), where n is the number of training documents
y[i] is the label of document i
X_test : numpy.array
BOW vectors of test samples
Shape: (m, d), where m is the number of test documents, d is the size of the vocabulary
X[i, j] is the number of occurences of word j in document i
y_test : numpy.array
Labels of test samples
Shape: (m,), where m is the number of test documents
y[i] is the label of document i
tfidf : bool
TF-IDF (True) or BOW (False)
norm : {None, 'l1', 'l2'}
Norm to normalize vectors
If norm is None, vectors are not normalized.
Otherwise, this argument is passed to `norm` argument of `sklearn.preprocessing.normalize`.
metric : {'l1', 'l2'}
Norm to compare vectors
This argument is passes to `metric` argument of `sklearn.metrics.pairwise_distances`.
Returns
-------
acc : float
Accuracy
"""
if tfidf:
X_train, X_test = compute_tfidf(X_train, X_test)
if norm:
X_train = normalize(X_train, axis=1, norm=norm)
X_test = normalize(X_test, axis=1, norm=norm)
D = pairwise_distances(X_test, X_train, metric=metric)
D_train = pairwise_distances(X_train, metric=metric)
return evaluate_D(y_train, y_test, D, D_train)
############################
# #
# weighted k-NN evaluation #
# #
############################
def knn_evaluation_smooth(y_train, y_test, D, gamma, k=19):
"""
Compute wkNN accuracy
Parameters
----------
y_train : numpy.array
Labels of training samples
Shape: (n,), where n is the number of training documents
y[i] is the label of document i
y_test : numpy.array
Labels of test samples
Shape: (m,), where m is the number of test documents
y[i] is the label of document i
D : numpy.array
Distance matrix of training and test samples
Shape: (n, m), where n is the number of training documents, m is the number of test documents
D[i, j] is the distance between training document i and test document j
gamma : float
Smoothness
k : int
Size of neighborhood in kNN classification
Returns
-------
acc : float
Accuracy
"""
acc = 0
for i in range(y_test.shape[0]):
rank = np.argsort(D[i])
if np.bincount(y_train[rank[:k]], np.exp(-D[i]/gamma)[rank[:k]]).argmax() == y_test[i]:
acc += 1
acc = acc / y_test.shape[0]
return acc
def select_gamma(y_train, D_train):
"""
Select the hyperparameter gamma in wkNN using validation data
Parameters
----------
y_train : numpy.array
Labels of training samples
Shape: (n,), where n is the number of training documents
y[i] is the label of document i
D_train : numpy.array
Distance matrix of training samples
Shape: (n, n), where n is the number of training documents
D[i, j] is the distance between training documents i and j
Returns
-------
best_gamma : float
Chosen hyperparamter gamma
"""
train, validation = train_test_split(np.arange(len(y_train)), test_size=0.3, random_state=0)
best_score = None
best_gamma = None
for gamma in [(i+1)/200 for i in range(20)]:
score = knn_evaluation_smooth(y_train[train], y_train[validation], D_train[validation][:, train], gamma)
if best_score is None or score > best_score:
best_score = score
best_gamma = gamma
return best_gamma
def evaluate_D_smooth(y_train, y_test, D, D_train):
"""
Evaluation using wkNN and distance metrices
Parameters
----------
y_train : numpy.array
Labels of training samples
Shape: (n,), where n is the number of training documents
y[i] is the label of document i
y_test : numpy.array
Labels of test samples
Shape: (m,), where m is the number of test documents
y[i] is the label of document i
D : numpy.array
Distance matrix of training and test samples
Shape: (n, m), where n is the number of training documents, m is the number of test documents
D[i, j] is the distance between training document i and test document j
D_train : numpy.array
Distance matrix of training samples
Shape: (n, n), where n is the number of training documents
D[i, j] is the distance between training documents i and j
Returns
-------
acc : float
Accuracy
"""
gamma = select_gamma(y_train, D_train)
return knn_evaluation_smooth(y_train, y_test, D, gamma)
def evaluate_onehot_smooth(X_train, y_train, X_test, y_test, tfidf=False):
"""
Evaluation using wkNN and onehot vectors
Parameters
----------
X_train : numpy.array
BOW vectors of training samples
Shape: (n, d), where n is the number of training documents, d is the size of the vocabulary
X[i, j] is the number of occurences of word j in document i
y_train : numpy.array
Labels of training samples
Shape: (n,), where n is the number of training documents
y[i] is the label of document i
X_test : numpy.array
BOW vectors of test samples
Shape: (m, d), where m is the number of test documents, d is the size of the vocabulary
X[i, j] is the number of occurences of word j in document i
y_test : numpy.array
Labels of test samples
Shape: (m,), where m is the number of test documents
y[i] is the label of document i
tfidf : bool
TF-IDF (True) or BOW (False)
Returns
-------
acc : float
Accuracy
"""
if tfidf:
X_train, X_test = compute_tfidf(X_train, X_test)
X_train = normalize(X_train, axis=1, norm='l1')
X_test = normalize(X_test, axis=1, norm='l1')
D = pairwise_distances(X_test, X_train, metric='manhattan')
D_train = pairwise_distances(X_train, metric='manhattan')
return evaluate_D_smooth(y_train, y_test, D, D_train)
###########
# #
# Loading #
# #
###########
def load(filename):
data = scipy.io.loadmat(filename)
X = np.vstack([x.T for x in data['X'][0]])
_, inverse = np.unique(X, axis=0, return_inverse=True)
docid = [[i for w in enumerate(x.T)] for i, x in enumerate(data['X'][0])]
docid = sum(docid, [])
freq = np.hstack([x[0] for x in data['BOW_X'][0]])
X = csr_matrix(coo_matrix((freq, (docid, inverse))))
y = data['Y'][0]
return data, X, y
def load_one(filename):
data = scipy.io.loadmat(filename)
n_train = len(data['xtr'][0])
X = np.vstack([x.T for x in data['xtr'][0]] + [x.T for x in data['xte'][0] if len(x.T) > 0])
_, inverse = np.unique(X, axis=0, return_inverse=True)
docid = [[i for w in enumerate(x.T)] for i, x in enumerate(data['xtr'][0])] + [[n_train + i for w in enumerate(x.T)] for i, x in enumerate(data['xte'][0])]
docid = sum(docid, [])
freq = np.hstack([x[0] for x in data['BOW_xtr'][0]] + [x[0] for x in data['BOW_xte'][0] if len(x.T) > 0])
X = csr_matrix(coo_matrix((freq, (docid, inverse))))
X_train = X[:n_train]
y_train = data['ytr'][0].astype(int)
X_test = X[n_train:]
y_test = data['yte'][0].astype(int)
return X_train, y_train, X_test, y_test