-
Notifications
You must be signed in to change notification settings - Fork 2
/
models.py
211 lines (173 loc) · 8.14 KB
/
models.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
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable
import config as cf
from utils import stack_or_create, get_all_data
from sklearn.cluster import KMeans, MiniBatchKMeans
from copy import deepcopy
from sklearn.metrics import pairwise_distances
from sklearn.metrics.pairwise import rbf_kernel
import math
import os
import numpy as np
weights = {}
biases = {}
class HD_CNN(nn.Module):
def __init__(self, args):
super(HD_CNN, self).__init__()
self.args = args
self.fines = {}
for i in range (self.args.num_superclass):
self.fines[i]=fine(self.args)
self.croase = croase(self.args)
self.share = share(self.args)
#self.cluster = clustering(self.args)
class share(nn.Module):
def __init__(self, args):
super(share, self).__init__()
self.args = args
# Encoder layers
self.enc_conv1_1 = nn.Conv2d(3, 64, kernel_size=3, stride=1, padding=1, bias=False)
self.enc_bn1_1 = nn.BatchNorm2d(64)
self.enc_drop1 = nn.Dropout(p=self.args.drop_rate)
self.enc_conv1_2 = nn.Conv2d(64, 64, kernel_size=3, stride=1, padding=1, bias=False)
self.enc_bn1_2 = nn.BatchNorm2d(64)
self.enc_max_pool1 = nn.MaxPool2d(kernel_size=2, stride=2)
self.enc_conv2_1 = nn.Conv2d(64, 128, kernel_size=3, stride=1, padding=1, bias=False)
self.enc_bn2_1 = nn.BatchNorm2d(128)
self.enc_drop2 = nn.Dropout(p=self.args.drop_rate)
self.enc_conv2_2 = nn.Conv2d(128, 128, kernel_size=3, stride=1, padding=1, bias=False)
self.enc_bn2_2 = nn.BatchNorm2d(128)
self.enc_max_pool2 = nn.MaxPool2d(kernel_size=2, stride=2)
self.relu = nn.ReLU()
self.sigmoid = nn.Sigmoid()
def encoder(self, x, *args, **kwargs):
x = self.enc_drop1(self.relu(self.enc_bn1_1(self.enc_conv1_1(x))))
x = self.enc_max_pool1(self.relu(self.enc_bn1_2(self.enc_conv1_2(x))))
x = self.enc_drop2(self.relu(self.enc_bn2_1(self.enc_conv2_1(x))))
x = self.enc_max_pool2(self.relu(self.enc_bn2_2(self.enc_conv2_2(x))))
return x
class croase(nn.Module):
def __init__(self, args):
super(croase, self).__init__()
self.args = args
self.enc_conv3_1 = nn.Conv2d(128, 256, kernel_size=3, stride=1, padding=1, bias=False)
self.enc_bn3_1 = nn.BatchNorm2d(256)
self.enc_drop3 = nn.Dropout(p=self.args.drop_rate)
self.enc_conv3_2 = nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1, bias=False)
self.enc_bn3_2 = nn.BatchNorm2d(256)
self.enc_max_pool3 = nn.MaxPool2d(kernel_size=2, stride=2)
self.enc_conv4_1 = nn.Conv2d(256, 512, kernel_size=3, stride=1, padding=1, bias=False)
self.enc_bn4_1 = nn.BatchNorm2d(512)
self.enc_drop4 = nn.Dropout(p=self.args.drop_rate)
self.enc_conv4_2 = nn.Conv2d(512, 512, kernel_size=3, stride=1, padding=1, bias=False)
self.enc_bn4_2 = nn.BatchNorm2d(512)
self.enc_max_pool4 = nn.MaxPool2d(kernel_size=2, stride=2)
self.fc1 = nn.Linear(512*2*2, 1024 )
self.fc_drop1 = nn.Dropout(p=self.args.drop_rate)
self.fc2 = nn.Linear(1*1*1024, 1024 )
self.fc_drop2 = nn.Dropout(p=self.args.drop_rate)
self.fc3 = nn.Linear(1*1*1024, self.args.num_fine_classes)
self.relu = nn.ReLU()
self.sigmoid = nn.Sigmoid()
def independent(self, x, *args, **kwargs):
x = self.enc_drop3(self.relu(self.enc_bn3_1(self.enc_conv3_1(x))))
x = self.enc_max_pool3(self.relu(self.enc_bn3_2(self.enc_conv3_2(x))))
x = self.enc_drop4(self.relu(self.enc_bn4_1(self.enc_conv4_1(x))))
x = self.enc_max_pool4(self.relu(self.enc_bn4_2(self.enc_conv4_2(x))))
x = x.reshape(-1,512*2*2)
coarse_predict = self.fc3(self.fc_drop2(self.fc2(self.fc_drop1(self.fc1(x)))))
return F.softmax(coarse_predict, dim=1)
class fine(nn.Module):
def __init__(self, args):
super(fine, self).__init__()
self.args = args
# Independent Encoder layers
self.enc_conv3_1 = nn.Conv2d(128, 256, kernel_size=3, stride=1, padding=1, bias=False)
self.enc_bn3_1 = nn.BatchNorm2d(256)
self.enc_drop3 = nn.Dropout(p=self.args.drop_rate)
self.enc_conv3_2 = nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1, bias=False)
self.enc_bn3_2 = nn.BatchNorm2d(256)
self.enc_max_pool3 = nn.MaxPool2d(kernel_size=2, stride=2)
self.enc_conv4_1 = nn.Conv2d(256, 512, kernel_size=3, stride=1, padding=1, bias=False)
self.enc_bn4_1 = nn.BatchNorm2d(512)
self.enc_drop4 = nn.Dropout(p=self.args.drop_rate)
self.enc_conv4_2 = nn.Conv2d(512, 512, kernel_size=3, stride=1, padding=1, bias=False)
self.enc_bn4_2 = nn.BatchNorm2d(512)
self.enc_max_pool4 = nn.MaxPool2d(kernel_size=2, stride=2)
self.fc1 = nn.Linear(512*2*2, 1024 )
self.fc_drop1 = nn.Dropout(p=self.args.drop_rate)
self.fc2 = nn.Linear(1*1*1024, 1024 )
self.fc_drop2 = nn.Dropout(p=self.args.drop_rate)
self.fc3 = nn.Linear(1*1*1024, self.args.num_fine_classes)
self.relu = nn.ReLU()
self.sigmoid = nn.Sigmoid()
def independent(self, x, *args, **kwargs):
x = self.enc_drop3(self.relu(self.enc_bn3_1(self.enc_conv3_1(x))))
x = self.enc_max_pool3(self.relu(self.enc_bn3_2(self.enc_conv3_2(x))))
x = self.enc_drop4(self.relu(self.enc_bn4_1(self.enc_conv4_1(x))))
x = self.enc_max_pool4(self.relu(self.enc_bn4_2(self.enc_conv4_2(x))))
x = x.reshape(-1,512*2*2)
fine_class = self.fc3(self.fc_drop2(self.fc2(self.fc_drop1(self.fc1(x)))))
return F.softmax(fine_class, dim=1)
class clustering(nn.Module):
def __init__(self, args):
self.Kernals=10
def acc(self, ypred, y):
assert len(y) > 0
assert len(np.unique(ypred)) == len(np.unique(y))
s = np.unique(ypred)
t = np.unique(y)
N = len(np.unique(ypred))
F = np.zeros((N, N), dtype = np.int32)
for i in range(N):
for j in range(N):
idx = np.logical_and(ypred == s[i], y == t[j])
F[i][j] = np.count_nonzero(idx)
return F
# assign to the clusters (M-step)
def get_assignments(self, X, centroids):
dist = pairwise_distances(X, centroids)
assign = np.argmin(dist,axis=1)
return assign
# compute the new centroids (E-step)
def get_centroids(self, X, assignments):
centroids = []
for i in np.unique(assignments):
centroids.append(X[assignments==i].mean(axis=0))
return np.array(centroids)
# initize the centroids
def init_kmeans_plus_plus(self, X, K):
'''Choose the next centroids with a prior of distance.'''
assert K>=2, "So you want to make 1 cluster?"
compute_distance = lambda X, c: pairwise_distances(X, c).min(axis=1)
# get the first centroid
centroids = [X[np.random.choice(range(X.shape[0])),:]]
# choice next
for _ in range(K-1):
proba = compute_distance(X,centroids)**2
proba /= proba.sum()
centroids.append(X[np.random.choice(range(X.shape[0]), p=proba)])
return np.array(centroids)
def KMeans(self, X, centroids, n_iterations=5, axes=None):
if axes is not None:
axes = axes.flatten()
for i in range(n_iterations):
assignments = self.get_assignments(X, centroids)
centroids = self.get_centroids(X, assignments)
return assignments, centroids
def spectral_clustering(self, A, K=2, gamma=10):
# A = rbf_kernel(X, gamma=gamma)
# A -= np.eye(A.shape[0]) # affinity
A /=A.sum(axis=1)
A = np.multiply(A,(np.ones((10,10))-np.identity(10)))
D = A.sum(axis=1) # degree
D_inv = np.diag(D**(-.5))
L = (D_inv).dot(A).dot(D_inv) # laplacian
s, Vh = np.linalg.eig(L)
eigenvector = Vh.real[:,:K].copy()
eigenvector /= ((eigenvector**2).sum(axis=1)[:,np.newaxis]**.5)
centroids = self.init_kmeans_plus_plus(eigenvector, K)
assignments, _ = self.KMeans(eigenvector, centroids)
return assignments