/
main.py
197 lines (169 loc) · 7.39 KB
/
main.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
import torch
from torch import nn
from torch.nn import functional as F
import sklearn, scipy
import numpy as np
from tqdm import trange, tqdm
from sklearn.cluster import KMeans
import random
from math import floor
class FSC(nn.Module):
"""FSC the main class of fuzzy semantic cells
"""
def __init__(self, M:int) -> None:
"""
M: the number of classes
"""
super().__init__()
self.M = M
def setP(self, mode:int, data:torch.Tensor, labels:torch.Tensor):
return {
1: self.setP1,
2: self.setP2,
3: self.setP3,
4: self.setP4,
}[mode](data, labels)
@torch.no_grad()
def setP1(self, data:torch.Tensor, labels:torch.Tensor) -> None:
"""setP1 the first methond to initialized prototypes.
Kmeans each category.
Args:
data (torch.Tensor): data
labels (torch.Tensor): labels
"""
num_clusters = 10
kmeans = KMeans(n_clusters=num_clusters, random_state=0)
result = []
rearrage_labels = []
for l in labels.unique():
kmeans.fit(data[labels==l, :].numpy())
result.append( torch.from_numpy(kmeans.cluster_centers_) )
rearrage_labels += [l,] * num_clusters
self.P = nn.Parameter( torch.cat(result, dim=0) )
self.label_belonging = torch.stack( rearrage_labels, dim=0 )
@torch.no_grad()
def setP2(self, data: torch.Tensor, labels: torch.Tensor):
"""setP2 the second methond to initialized prototypes.
Kmeans the whole dataset
Args:
data (torch.Tensor): data
labels (torch.Tensor): labels
"""
n = 10
num_clusters = n * self.M
kmeans = sklearn.cluster.KMeans(n_clusters=num_clusters, random_state=0)
kmeans.fit(data.numpy())
result = []
rearrage_labels = []
centers = kmeans.cluster_centers_
for l in labels.unique():
d = torch.cdist( torch.from_numpy(centers)[None, :, :], data[None, labels==l, :])[0]
idx = d.min(dim=1)[0].topk(k=n, largest=False)[1]
result.append( torch.from_numpy(centers[idx, :]) )
rearrage_labels += [l,]*n
centers[idx, :] = np.Inf
self.P = nn.Parameter( torch.cat(result, dim=0) )
self.label_belonging = torch.stack( rearrage_labels, dim=0 )
@torch.no_grad()
def setP3(self, data: torch.Tensor, labels: torch.Tensor):
"""setP3 the third methond to initialized prototypes.
Randomly select prototypes
Args:
data (torch.Tensor): data
labels (torch.Tensor): labels
"""
num_clusters = 10
result = []
rearrage_labels = []
for l in labels.unique():
result.append( data[labels==l, :][:num_clusters] )
rearrage_labels += [l,] * num_clusters
self.P = nn.Parameter( torch.cat(result, dim=0) )
self.label_belonging = torch.stack( rearrage_labels, dim=0 )
@torch.no_grad()
def setP4(self, data: torch.Tensor, labels: torch.Tensor):
"""setP4 the forth methond to initialized prototypes.
the mass center of each category
Args:
data (torch.Tensor): data
labels (torch.Tensor): labels
"""
result = []
rearrage_labels = []
for l in labels.unique():
result.append( data[labels==l, :].mean(dim=0, keepdim=True) )
rearrage_labels += [l,]
self.P = nn.Parameter( torch.cat(result, dim=0) )
self.label_belonging = torch.stack( rearrage_labels, dim=0 )
@torch.no_grad()
def setSigma(self, data: torch.Tensor, labels: torch.Tensor):
"""setSigma set $\sigma$ as 1/3 of the average distance
Args:
data (torch.Tensor): data
labels (torch.Tensor): labels
"""
result = []
for p, l in zip(self.P, self.label_belonging):
result.append( torch.dist(p, data[labels==l, :]).mean()/3 )
self.sigma = nn.Parameter( torch.stack(result, dim=0) )
def forward(self, data: torch.Tensor, labels: torch.Tensor) -> torch.tensor:
total_loss = torch.tensor(0.)
for l in labels.unique():
d = torch.cdist(self.P[None, :, :], data[None, labels==l, :])[0].pow(2) # |P| num_samples
mu = torch.exp( d.div( -self.sigma[:, None].pow(2).mul(2) ) ) # |P| num_samples
mu_i = mu.sum(dim=1) # |P| 1
mapping = F.one_hot(self.label_belonging, num_classes=max(self.label_belonging)+1) # |P| |C|
result = torch.mv(mapping.transpose(0, 1).float(), mu_i) # |C|
total_loss += F.cross_entropy(result[None, :], l[None]) # the conditional entropy
return total_loss
@torch.no_grad()
def test(self, data:torch.Tensor) -> torch.Tensor:
d = torch.cdist(self.P[None, :, :], data[None, :, :])[0].pow(2) # |P| num_samples
mu = d.div(-self.sigma[:, None].pow(2).mul(2)).exp() # |P| num_samples
idx = mu.max( dim=0 )[1]
return self.label_belonging[idx]
if __name__ == "__main__":
device = torch.device('cpu')
train_data = torch.from_numpy(np.loadtxt('./pendigits_sta4_train.csv', delimiter=',')).float()
train_labels = torch.from_numpy(np.loadtxt('./pendigits_label_train.csv')).long()
test_data = torch.from_numpy(np.loadtxt('./pendigits_sta4_test.csv', delimiter=',')).float()
test_labels = torch.from_numpy(np.loadtxt('./pendigits_label_test.csv')).long()
data = torch.cat([train_data, test_data], dim=0)
labels = torch.cat([train_labels, test_labels], dim=0)
for P in [4,1,2,3]:
for ratio in [0.1, 0.3, 0.5, 0.7,] * 10:
idx = list(range(len(data)))
random.shuffle(idx)
num = floor(len(data) * ratio)
# training set
train_data = data[idx[:num], :]
train_labels = labels[idx[:num]]
# testing set
test_data = data[idx[num:], :]
test_labels = labels[idx[num:]]
# initialize P and \sigma
fsc = FSC(len(train_labels.unique()), train_data.size(1))
fsc.setP(P, train_data, train_labels)
fsc.setSigma(train_data, train_labels)
# prepape to device
fsc = fsc.to(device)
train_data = train_data.to(device)
train_label = train_labels.to(device)
test_data = test_data.to(device)
test_labels = test_labels.to(device)
optimizer = torch.optim.Adam(fsc.parameters())
# training
loss = torch.tensor(0.)
max_acc = 0.
for ii in range(2000):
# update
loss = fsc(train_data, train_label)
optimizer.zero_grad()
loss.backward()
optimizer.step()
# print loss
loss = loss.item()
train_acc = fsc.test(train_data).eq(train_labels).sum().div(len(train_label)).item()
test_acc = fsc.test(test_data.float()).cpu().eq(test_labels).sum().div(len(test_labels)).item()
print(f"\r {ii} {loss:.4} {train_acc:.2f} {test_acc:.2f} ", end="")
print(f"P:{P}, ratio:{ratio}, train_acc:{train_acc}, test_acc:{test_acc},")