-
Notifications
You must be signed in to change notification settings - Fork 4
/
tailcalib.py
366 lines (288 loc) · 14.8 KB
/
tailcalib.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
from dataclasses import dataclass
from tqdm import tqdm
import numpy as np
@dataclass
class tailcalib:
base_engine: str = "numpy" # Options: NumPy, PyTorch
def generate(self, X, y, tukey_value=0.9, alpha=0.0, topk=1, extra_points=0, shuffle=True):
"""Generate new datapoints
Args:
X : Features
y : Corresponding labels
tukey_value [Hyperparameter]: Value to convert any distrubution of data into a normal distribution. Defaults to 1.0.
alpha [Hyperparameter]: Decides how spread out the generated data is. Defaults to 0.0.
topk [Hyperparameter]: Decides how many nearby classes should be taken into consideration for the mean and std of the newly generated data. Defaults to 1.
extra_points [Hyperparameter]: By default the number of datapoints to be generated is decided based on the class with the maximum datapoints. This variable decides how many more extra datapoints should be generated on top of that. Defaults to 0.
shuffle (bool, optional): Shuffles the generated and original datapoints together. Defaults to True.
Returns:
feat_all: Tukey transformed train data + generated datapoints
labs_all: Corresponding labels to feat_all
generated_points: Dict that consists of just the generated points with class label as keys.
"""
self.X = X
self.y = y
self.tukey_value = tukey_value
self.alpha = alpha
self.topk = topk
self.extra_points = extra_points
self.shuffle = shuffle
self.sanity_checks()
if self.base_engine == "numpy":
import numpy as np
return self.generate_usingNumPy()
elif self.base_engine == "pytorch":
import torch
return self.generate_usingPyTorch()
else:
raise Exception(f"Invalid base_engine choice - '{self.base_engine}' | Choose from: 'numpy', 'pytorch'.")
def sanity_checks(self,):
"""Checks whether the type of X, y inputs matches with the chosen base_engine
"""
if self.base_engine == "numpy":
import numpy as np
assert isinstance(self.X, np.ndarray),"Base Engine is set to NumPy, so the X must be a NumPy instance!"
assert isinstance(self.y, np.ndarray),"Base Engine is set to NumPy, so the y must be a NumPy instance!"
else:
import torch
assert torch.is_tensor(self.X),"Base Engine is set to PyTorch, so the X must be a PyTorch instance!"
assert torch.is_tensor(self.y),"Base Engine is set to PyTorch, so the y must be a PyTorch instance!"
def generate_usingNumPy(self,):
"""Generate new datapoints using Numpy
"""
import scipy.spatial as sp
feat = {}
labs = {}
y_unique, y_count = np.unique(self.y, return_counts=True)
assert len(y_unique) >= self.topk, "The 'topk' is greater than the number of uniquely available classes. Try a lesser value."
for i in y_unique:
feat[i] = self.X[self.y == i]
labs[i] = np.full((feat[i].shape[0],), i)
# Class statistics
base_means = []
base_covs = []
for i in feat.keys():
base_means.append(feat[i].mean(axis=0))
base_covs.append(np.expand_dims(self.get_cov(feat[i]),axis=0))
base_means = np.vstack(base_means)
base_covs = np.vstack(base_covs)
# Tukey's transform
for i in feat.keys():
feat[i] = self.tukey_transform(feat[i], self.tukey_value)
# Distribution calibration and feature sampling
sample_from_each = self.get_sample_count(y_count,feat.keys(), self.extra_points)
generated_points = {}
for i in tqdm(feat.keys()):
if np.sum(sample_from_each[i]) == 0 and self.extra_points == 0 :
continue
generated_points[i] = []
for k, x_ij in zip(sample_from_each[i], feat[i]):
if k == 0:
continue
# Getting the top k nearest classes based on l2 distance
distances = sp.distance.cdist(base_means, np.expand_dims(x_ij, axis=0)).squeeze()
topk_idx = np.argsort(-distances)[::-1][:self.topk]
# Calibrating mean and covariance
calibrated_mean, calibrated_cov = self.calibrate_distribution(base_means[topk_idx], base_covs[topk_idx], self.topk, x_ij, self.alpha)
# Trick to avoid cholesky decomposition from failing. Look at https://juanitorduz.github.io/multivariate_normal/
EPS = 1e-4
calibrated_cov += (np.eye(calibrated_cov.shape[0])*EPS)
gen = np.random.multivariate_normal(calibrated_mean, calibrated_cov,(int(k),))
generated_points[i].append(gen)
generated_points[i] = np.vstack(generated_points[i])
print("Point Generation Completed!")
print("Don't forget to use '.convert_others()' to apply tukey transformation on validation/test data before validation/testing. Use the same 'tukey_value' as the train data.\n")
feat_all = []
labs_all = []
for i in labs.keys():
feat_all.append(feat[i])
labs_all.append(labs[i])
feat_all = np.vstack(feat_all)
labs_all = np.hstack(labs_all)
for i in generated_points.keys():
feat_all = np.concatenate((feat_all, generated_points[i]))
labs_all = np.concatenate((labs_all, np.full((generated_points[i].shape[0],), int(i))))
feat_all, labs_all = self.shuffle_all(feat_all, labs_all)
return feat_all, labs_all, generated_points
def generate_usingPyTorch(self,):
"""Generate new datapoints using PyTorch
"""
import torch
device = self.X.device
feat = {}
labs = {}
y_unique, y_count = np.unique(self.y.cpu().numpy(), return_counts=True)
assert len(y_unique) >= self.topk, "The 'topk' is greater than the number of uniquely available classes. Try a lesser value."
for i in y_unique:
feat[i] = self.X[self.y == i]
labs[i] = torch.full((feat[i].size()[0],), i).to(device)
# Class statistics
base_means = []
base_covs = []
for i in feat.keys():
base_means.append(feat[i].mean(dim=0))
base_covs.append(self.get_cov(feat[i]).unsqueeze(dim=0))
base_means = torch.vstack(base_means)
base_covs = torch.vstack(base_covs)
# Tukey's transform
for i in feat.keys():
feat[i] = self.tukey_transform(feat[i], self.tukey_value)
# Distribution calibration and feature sampling
sample_from_each = self.get_sample_count(y_count,feat.keys(), self.extra_points)
generated_points = {}
for i in tqdm(feat.keys()):
if np.sum(sample_from_each[i]) == 0 and self.extra_points == 0 :
continue
generated_points[i] = []
for k, x_ij in zip(sample_from_each[i], feat[i]):
if k == 0:
continue
# Getting the top k nearest classes based on l2 distance
distances = torch.cdist(base_means, x_ij.unsqueeze(0)).squeeze()
topk_idx = torch.topk(-distances, k=self.topk)[1][:self.topk]
# Calibrating mean and covariance
calibrated_mean, calibrated_cov = self.calibrate_distribution(base_means[topk_idx], base_covs[topk_idx], self.topk, x_ij, self.alpha)
# Trick to avoid cholesky decomposition from failing. Look at https://juanitorduz.github.io/multivariate_normal/
EPS = 1e-4
calibrated_cov += (torch.eye(calibrated_cov.shape[0])*EPS).to(device)
new_dist = torch.distributions.multivariate_normal.MultivariateNormal(calibrated_mean, calibrated_cov)
gen = new_dist.sample((int(k),))
generated_points[i].append(gen)
generated_points[i] = torch.vstack(generated_points[i])
torch.cuda.empty_cache()
print("Point Generation Completed!")
print("Don't forget to use '.convert_others()' to apply tukey transformation on validation/test data before validation/testing. Use the same 'tukey_value' as the train data.\n")
feat_all = []
labs_all = []
for i in labs.keys():
feat_all.append(feat[i])
labs_all.append(labs[i])
feat_all = torch.vstack(feat_all)
labs_all = torch.hstack(labs_all).to(device)
for i in generated_points.keys():
feat_all = torch.cat((feat_all, generated_points[i].to(device)))
labs_all = torch.cat((labs_all, torch.full((generated_points[i].size()[0],), int(i)).to(device)))
feat_all, labs_all = self.shuffle_all(feat_all, labs_all)
return feat_all, labs_all, generated_points
def convert_others(self, X, tukey_value=1.0):
"""Use for applying tukey transformation on validation/test data before validation/testing!
Args:
X : Features
tukey_value [Hyperparameter]: Value to convert any distrubution of data into a normal distribution. Defaults to 1.0.
Returns:
Tukey transformed data
"""
# Sanity check
if self.base_engine == "numpy":
import numpy as np
assert isinstance(X, np.ndarray),"Base Engine is set to NumPy, so the X must be a NumPy instance!"
else:
import torch
assert torch.is_tensor(X),"Base Engine is set to PyTorch, so the X must be a PyTorch instance!"
return self.tukey_transform(X, tukey_value)
def shuffle_all(self, x, y):
"""Force shuffle data
Args:
x (float Tensor): Datapoints
y (int): Labels
Returns:
floatTensor, int: Return shuffled datapoints and corresponding labels
"""
if self.base_engine == "numpy":
index = np.random.permutation(x.shape[0])
else:
index = torch.randperm(x.size(0))
x = x[index]
y = y[index]
return x, y
def get_cov(self, X):
"""Calculate the covariance matrix for X
Args:
X (torch.tensor): Features
Returns:
[torch.tensor]: Covariance matrix of X
"""
n = X.shape[0]
if self.base_engine == "numpy":
mu = X.mean(axis=0)
else:
mu = X.mean(dim=0)
X = (X - mu)
return 1/(n-1) * (X.transpose(1, 0) @ X) # X^TX -> feat_size x num_of_samples @ num_of_samples x feat_size -> feat_size x feat_size
def get_sample_count(self, count, keys, extra_points):
"""Decides how many samples must be generated based on each existing train datapoints.
Args:
count (list): Number of samples in each class
keys (dict.keys): Class keys
extra_points (int): The number of samples to be generated is based on the class with maximum sample count. "extra_points" decide how much more samples must be generated on top of that.
Returns:
dict: dict consists that has the info as to how many samples must be generated based on each existing train datapoints.
"""
sample_count_dict = {}
for i in keys:
current = count[i]
head = max(count)
# head is the sample count that we must match after the generation. This can be offset by "config["pg"]["extra_points"]". In our experiments this is set to 0 as it worked better.
num_sample = head - current + extra_points
ratio = num_sample / current
# Makes sure each datapoint is being used atleast once
new_sample_from_each = [np.floor(ratio)] * current
# Rest of the datapoints used for generation are decided randomly
while True:
if sum(new_sample_from_each) == num_sample:
break
idx = np.random.randint(0, current)
new_sample_from_each[idx] += 1
# Sanity checks
assert sum(new_sample_from_each) == num_sample
assert len(new_sample_from_each) == current
sample_count_dict[i] = new_sample_from_each
return sample_count_dict
def calibrate_distribution(self, base_means, base_cov, k, x_ij, alpha=0.0):
"""Calibration of the distribution for generation. Check equation 7 and 8 from our paper - Feature Generation for Long-tail Classification.
Args:
base_means (torch.tensor): List of all the means that are used for calibration.
base_cov (torch.tensor): List of all the covariance matrices used for calibraton.
k (int): Number of classes used for calibration.
x_ij (torch.tensor): Datapoint chosen to be used for generation.
alpha (float, optional): Decides the spread of the generated samples. Defaults to 0.0.
Returns:
torch.tensor : Calibrated mean and covariance matrix
"""
if self.base_engine == "numpy":
calibrated_mean = (base_means.sum(axis=0) + x_ij)/(k+1)
calibrated_cov = base_cov.sum(axis=0)/k + alpha
else:
calibrated_mean = (base_means.sum(dim=0) + x_ij)/(k+1)
calibrated_cov = base_cov.sum(dim=0)/k + alpha
return calibrated_mean, calibrated_cov
def tukey_transform(self, x, lam=0.2):
"""Transforms any distribution into normal-distribution like.
Args:
x (torch.tensor): Features
lam (float, optional): Adjusts how close the transformed features will be to the origin. Defaults to 0.2.
Returns:
torch.tensor: Normal distribution like features.
"""
if lam == 0:
EPS = 1e-6
x = x + EPS
return x.log()
else :
return x**lam
def sample_test(self):
"""[For internal testing only] Uses a randomly generated sample set to check if both the engines work error free.
"""
if self.base_engine == "numpy":
import numpy as np
X = np.random.rand(200,100)
y = np.random.randint(0,10, (200,))
feat, lab, gen = self.generate(X=X, y=y)
print(f"Before: {np.unique(y, return_counts=True)}")
print(f"After: {np.unique(lab, return_counts=True)}")
else:
import torch
X = torch.rand((200,100))
y = torch.randint(0,10, (200,))
feat, lab, gen = self.generate(X=X, y=y)
print(f"Before: {torch.unique(y, return_counts=True)}")
print(f"After: {torch.unique(lab, return_counts=True)}")