-
Notifications
You must be signed in to change notification settings - Fork 3
/
rsom.py
407 lines (330 loc) · 14.4 KB
/
rsom.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
"""
Rectifying Self Organazing Maps a.k.a RSOM
RSOM is a clustering and outlier detection method that is predicated with
old Self Organazing Maps.
It includes Batch and Stochastic learning rules. There are two different
implementations. One is based on Numpy and tthe other is Theano. If you have
tall and wide data matrix, we suggest to use Theano version. Otherwise
Numpy version is faster. You can also use GPU with Theano but you need to
set Theano configurations.
For more detail about RSOM refer to http://arxiv.org/abs/1312.4384
AUTHOR:
Eren Golge
www.erengolge.com
"""
"""
TO DO:
-> Try dot product distance instead of Euclidean
-> Normzalize only updated weight vectors in that epoch
-> compare code with https://github.com/JustGlowing/minisom/blob/master/minisom.py
-> print resulting objective values
-> write bookeeping for best objective value
-> learning rate is already decreasing so radius might be good to keep it constant
-> UPDATE only winners
"""
import logging
from typing import Optional, Tuple
import numpy as np
import torch
# Set up logging
logging.basicConfig(
level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s"
)
logger = logging.getLogger(__name__)
class RSOM(torch.nn.Module):
def __init__(
self,
data: torch.Tensor,
num_units: int = 10,
height: Optional[int] = None,
width: Optional[int] = None,
alpha_max: float = 0.05,
alpha_min: float = 0.001,
set_count_activations: bool = True,
set_outlier_unit_det: bool = True,
set_inunit_outlier_det: bool = True,
outlier_unit_thresh: float = 0.5,
inunit_outlier_thresh: float = 95,
dist: str = "euclidean",
log_full_data_cost: bool = False,
steps_to_full_data_cost: int = -1, # TODO: number of steps to compute full data cost
):
"""Rectifying Self Organizing Maps. RSOM is a clustering and outlier detection method that is predicated with old Self Organizing Maps.
Args:
data: Input data.
num_units: Number of units.
height: Height of the map.
width: Width of the map.
alpha_max: Maximum learning rate.
alpha_min: Minimum learning rate.
set_count_activations: Whether to count activations.
set_outlier_unit_det: Whether to detect outlier units.
set_inunit_outlier_det: Whether to detect in-unit outliers.
outlier_unit_thresh: Threshold for outlier unit detection.
inunit_outlier_thresh: Threshold for in-unit outlier detection.
dist: Distance metric. Can be "euclidean" or "cosine".
log_full_data_cost: Whether to log full data cost or use exponential moving average. Computing full data
cost is expensive and might cause OOM.
"""
super(RSOM, self).__init__()
self.X = data
self.num_units = num_units
self.height = height
self.width = width
self.alpha_max = alpha_max
self.alpha_min = alpha_min
self.set_count_activations = set_count_activations
self.set_outlier_unit_det = set_outlier_unit_det
self.set_inunit_outlier_det = set_inunit_outlier_det
self.outlier_unit_thresh = outlier_unit_thresh
self.inunit_outlier_thresh = inunit_outlier_thresh
self.log_full_data_cost = log_full_data_cost
self._estimate_map_shape()
self.data_dim = self.X.shape[1]
self.W = torch.nn.Parameter(torch.randn(self.num_units, self.data_dim))
self._normalize_weights()
self.distance_metric = dist
if self.distance_metric not in ["euclidean", "cosine"]:
raise ValueError("distance_metric must be either 'euclidean' or 'cosine'")
self.activations = torch.zeros(self.num_units)
self.unit_saliency_coeffs = torch.zeros(self.num_units)
self.unit_saliency = torch.ones(self.num_units, dtype=torch.bool)
self.inst_saliency = torch.tensor([])
self.ins_unit_assign = torch.tensor([])
self.ins_unit_dist = torch.tensor([])
self.unit_coher = torch.tensor([])
def _normalize_weights(self):
self.W.data = self.W.data / torch.norm(self.W.data, dim=1, keepdim=True)
def _estimate_map_shape(self):
if self.height is None or self.width is None:
u, s, v = torch.svd(self.X)
ratio = s[0] / s[1]
self.height = min(
self.num_units, int(np.ceil(np.sqrt(self.num_units / ratio)))
)
self.width = int(np.ceil(self.num_units / self.height))
self.num_units = self.height * self.width
logging.info(
f"Estimated map size is -> height = {self.height}, width = {self.width}"
)
def unit_cords(self, index: int) -> Tuple[int, int]:
return index % self.width, index // self.width
def _calc_distance(self, X, X2=None):
if self.distance_metric == "euclidean":
return self._euclidean_distance(X, X2)
elif self.distance_metric == "cosine":
return self._cosine_distance(X)
def _euclidean_distance(self, X, X2=None):
if X2 is None:
X2 = (X**2).sum(1)[:, None]
W2 = (self.W**2).sum(1)[:, None]
return -2 * torch.mm(self.W, X.t()) + W2 + X2.t()
def _cosine_distance(self, X):
X_norm = X / torch.norm(X, dim=1, keepdim=True)
W_norm = self.W / torch.norm(self.W, dim=1, keepdim=True)
return 1 - torch.mm(W_norm, X_norm.t())
def find_neighbors(self, unit_id: int, radius: int) -> torch.Tensor:
neighbors = torch.zeros(1, self.num_units)
unit_x, unit_y = self.unit_cords(unit_id)
min_y = max(int(unit_y - radius), 0)
max_y = min(int(unit_y + radius), self.height - 1)
min_x = max(int(unit_x - radius), 0)
max_x = min(int(unit_x + radius), self.width - 1)
for y in range(min_y, max_y + 1):
for x in range(min_x, max_x + 1):
dist = abs(y - unit_y) + abs(x - unit_x)
neighbors[0, x + (y * self.width)] = dist
return neighbors
def best_match(self, X: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
if X.dim() == 1:
X = X.unsqueeze(0)
X2 = (X**2).sum(1).unsqueeze(1)
D = -2 * torch.mm(self.W, X.t()) + (self.W**2).sum(1).unsqueeze(1) + X2.t()
BMU = (D == D.min(0)[0]).float().t()
return BMU, D
def assing_to_units(self, X=None):
if X is None:
D = self._calc_distance(self.X)
self.ins_unit_assign = D.argmin(axis=0)
self.ins_unit_dist = D[self.ins_unit_assign, torch.arange(self.X.shape[0])]
else:
D = self._calc_distance(X)
ins_unit_assign = D.argmin(axis=0)
ins_unit_dist = D[ins_unit_assign, torch.arange(X.shape[0])]
return ins_unit_assign, ins_unit_dist
def set_params(self, num_epoch: int) -> dict:
U = {"alphas": [], "H_maps": [], "radiuses": []}
dist_map = torch.zeros(self.num_units, self.num_units)
radius = np.ceil(1 + np.floor(min(self.width, self.height) - 1) / 2) - 1
for u in range(self.num_units):
dist_map[u, :] = self.find_neighbors(u, self.num_units)
for epoch in range(num_epoch):
alpha = self.alpha_max - self.alpha_min
alpha = alpha * (num_epoch - epoch) / num_epoch + self.alpha_min
radius = np.ceil(1 + np.floor(min(self.width, self.height) - 1) / 2) - 1
radius = radius * (num_epoch - epoch) / (num_epoch - 1) - 1
radius = max(radius, 0)
neigh_updt_map = alpha * (1 - dist_map / float((1 + radius)))
neigh_updt_map[dist_map > radius] = 0
U["H_maps"].append(neigh_updt_map)
U["alphas"].append(alpha)
U["radiuses"].append(radius)
return U
def train_batch(
self,
num_epoch: Optional[int] = None,
batch_size: Optional[int] = None,
verbose: bool = True,
):
"""
Args:
num_epoch: number of epochs to train
batch_size: number of samples to train in one batch
verbose: if True, print the progress of training
"""
if num_epoch is None:
num_epoch = 500 * self.num_units
if batch_size is None:
batch_size = self.X.shape[0]
logger.info("Learning...")
U = self.set_params(num_epoch)
X2 = None
if batch_size == self.X.shape[0]:
X2 = (self.X**2).sum(1).unsqueeze(1)
for epoch in range(num_epoch):
logger.info(f"Epoch --- {epoch}")
update_rate = U["H_maps"][epoch]
learn_rate = U["alphas"][epoch]
shuffle_indices = torch.randperm(self.X.shape[0])
win_counts = torch.zeros(self.num_units)
batches = torch.split(shuffle_indices, batch_size)
num_steps = len(batches)
for step, batch_indices in enumerate(batches):
batch_data = self.X[batch_indices, :]
D = self._calc_distance(batch_data, X2)
BMU = (D == D.min(0)[0][None, :]).float().t()
win_counts += BMU.sum(dim=0)
if self.set_count_activations:
self.activations += win_counts
A = torch.mm(BMU, update_rate)
S = A.sum(0)
non_zeros = S.nonzero().squeeze()
self.W.data[non_zeros] = torch.mm(A[:, non_zeros].t(), batch_data) / S[
non_zeros
].unsqueeze(1)
self._print_cost(X2, D, epoch, num_epoch, step, num_steps)
if self.set_outlier_unit_det:
self._update_unit_saliency(win_counts, update_rate, learn_rate)
if self.set_count_activations:
self.activations /= self.activations.sum()
self.assing_to_units()
if self.set_outlier_unit_det:
self._find_outlier_units()
if self.set_inunit_outlier_det:
self._find_inunit_outliers()
def _print_cost(
self,
X2: torch.Tensor,
D: torch.Tensor,
epoch: int,
num_epoch: int,
step: int,
num_steps: int,
):
batch_cost = D.min(0)[0].mean()
if self.log_full_data_cost:
# TODO: handle when data is too large
if self.distance_metric == "euclidean":
D = self._calc_distance(self.X, X2)
cost = torch.norm(D.min(0)[0], p=1) / self.X.shape[0]
elif self.distance_metric == "cosine":
D = self._calc_distance(self.X)
cost = D.min(0)[0].mean() # Average minimum cosine distance
else:
# exponential moving average cost
if not hasattr(self, "avg_cost"):
self.avg_cost = batch_cost
else:
self.avg_cost = self.avg_cost * 0.01 + batch_cost * 0.99
cost = self.avg_cost
logger.info(
f"epoch {epoch} / {num_epoch} -- step {step} / {num_steps} -- avg-cost: {cost.item():.6f} -- batch-cost: {batch_cost.item():.6f}"
)
def _update_unit_saliency(
self, win_counts: torch.Tensor, update_rate: torch.Tensor, learn_rate: float
):
excitations = (update_rate * win_counts.unsqueeze(1)).sum(dim=0) / learn_rate
excitations = excitations / excitations.sum()
single_excitations = win_counts * learn_rate
single_excitations = single_excitations / single_excitations.sum()
self.unit_saliency_coeffs += excitations + single_excitations
def _find_outlier_units(self):
self.unit_saliency_coeffs /= self.unit_saliency_coeffs.sum()
self.unit_saliency = (
self.unit_saliency_coeffs > self.outlier_unit_thresh / self.num_units
)
self.inst_saliency = torch.ones(self.X.shape[0], dtype=torch.bool)
outlier_units = torch.where(self.unit_saliency == False)[0]
for i in outlier_units:
self.inst_saliency[torch.where(self.ins_unit_assign == i)[0]] = False
def _find_inunit_outliers(self):
if self.inst_saliency.numel() == 0:
self.inst_saliency = torch.ones(self.X.shape[0], dtype=torch.bool)
for i in torch.unique(self.ins_unit_assign):
indices = torch.where(self.ins_unit_assign == i)[0]
unit_thresh = torch.quantile(
self.ins_unit_dist[indices], self.inunit_outlier_thresh / 100
)
outlier_insts = indices[self.ins_unit_dist[indices] > unit_thresh]
self.inst_saliency[outlier_insts] = False
def salient_inst_index(self) -> torch.Tensor:
return torch.where(self.inst_saliency == True)[0]
def salient_unit_index(self) -> torch.Tensor:
return torch.where(self.unit_saliency == True)[0]
def salient_insts(self) -> torch.Tensor:
return self.X[self.inst_saliency]
def salient_units(self) -> torch.Tensor:
return self.W[self.unit_saliency]
def inst_to_unit_mapping(self) -> torch.Tensor:
return torch.stack((torch.arange(self.X.shape[0]), self.ins_unit_assign))
if __name__ == "__main__":
import matplotlib.pyplot as plt
import som_plot
import torch
from sklearn.datasets import load_digits
from sklearn.preprocessing import StandardScaler
# Load the digits dataset
digits = load_digits()
X = digits.data
y = digits.target
# Preprocess the data
scaler = StandardScaler()
X_scaled = scaler.fit_transform(X)
# Convert to PyTorch tensor
X_tensor = torch.tensor(X_scaled, dtype=torch.float32)
# Initialize and train SOM
som = SOM(X_tensor, num_units=100, alpha_max=0.05, alpha_min=0.01)
som.train_batch(num_epoch=1000, batch_size=32, verbose=True)
# Get the weights and assign instances to units
W = som.W.detach().numpy()
som.assing_to_units()
# Plot scatter plot
som_plot.som_plot_scatter(W, X_scaled, som.activations.numpy())
# Plot outlier scatter plot
som_plot.som_plot_outlier_scatter(
W,
X_scaled,
som.unit_saliency.numpy(),
som.inst_saliency.numpy(),
som.activations.numpy(),
)
# Plot mapping
distance_map = (
som._euq_dist(torch.sum(X_tensor**2, dim=1).unsqueeze(1), X_tensor)
.detach()
.numpy()
)
distance_map = distance_map.reshape(som.height, som.width)
som_plot.som_plot_mapping(distance_map)
plt.show()