-
Notifications
You must be signed in to change notification settings - Fork 3
/
vgg_vid.R
237 lines (189 loc) · 6.84 KB
/
vgg_vid.R
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
library(dplyr)
library(plotly)
library(tidyverse) # data manipulation
library(cluster) # clustering algorithms
library(factoextra) # clustering visualization
library(dendextend) # for comparing two dendrograms
library(glmnet)
setwd("/Users/jiayun/Documents/coding/vgg")
# load data
rm(list = ls())
####################################
# compute total loss of a logistic predictor
# input: a dataset (X,y) and a logistic model beta
# output: total loss
totalloss = function(beta, X, y, lambda){
sum(log(1 + exp(-y * (beta[1] + X %*% beta[-1])))) + lambda * sum(abs(beta))
}
modelreliance = function(index, beta, X, y){
perm = sample(1:nrow(X), nrow(X), replace = FALSE) #permutation
X0 = X
loss_vector = rep(0, M)
for(iter in 1:M){
X0[,index] = X[perm, index]
loss_vector[iter] = totalloss(beta, X0, y, lambda)
}
return(mean(loss_vector)/totalloss(beta, X, y, lambda))
}
#
# # model reliance of a single model on all variables
# # input:
# # beta - the model to analyze
# # X, y - the dataset
# # output: model reliance of a single model on all variables (ratio)
#
modelreliance.full = function(beta, X, y){
mr = rep(0, length(beta)-1)
for(i in 1:length(mr)){
mr[i] = modelreliance(i, beta, X, y)
}
return(mr)
}
####################################
feature = read.csv(file="features_small.csv", head = FALSE, sep=",")
outcome = read.csv(file="outcomes_small.csv", head = FALSE, sep=",")
feature = as.matrix(feature)
outcome = as.matrix(outcome)
cv.out <- cv.glmnet(feature, outcome, alpha = 1, family = "binomial", type.measure = "mse")
# best value of lambda
lambda <- cv.out$lambda.1se
# regression coefficients
beta = coef(cv.out, s = lambda)
# prediction accuracy
lasso_prob <- predict(cv.out, newx = feature, s = lambda, type = "response")
lasso_predict = ifelse(lasso_prob > 0.5, 1, 0)
sum(lasso_predict == outcome)/length(outcome)
# feature selection
id = beta@i[-1]
X = feature[,id]
beta_star = beta@x
##############################
# vid analysis
# X = feature
y = outcome
y[y==0] = -1
# parameter
epsilon = 0.05
NUM = 500 # number of points in each round of sampling
NUM_PCA = 10 # 10 for lower dim
size = 0.001 # for initial sampling
set.seed(2018)
epsScale = 3.75 # over-sample a bit # 3.75 for lower dim
M = 20
loss0 = totalloss(beta_star, X, y, lambda)
bound = loss0 * (1 + epsilon)
# initial sampling
# get NUM models in a box around beta*
points = matrix(runif(length(beta_star) * NUM, -1, 1), nrow = NUM, ncol = length(beta_star))
points = points * max(beta_star) * 0.1
points = sweep(points, 2, beta_star, "+")
# keep good models only
myLoss = rep(0, nrow(points))
for(i in 1:nrow(points)){
myLoss[i] = totalloss(points[i,], X, y, lambda)
}
points = points[myLoss < bound, ]
source("ellipsoid.R")
for(t in 1:NUM_PCA){
# pca for points from last sampling process
pca = prcomp(points, center = TRUE, scale = FALSE, rank. = 513)
cen = pca$center
len = apply(abs(pca$x), 2, max) * epsScale
rot = pca$rotation
# next round of samping
points = ellipsoid(cen, len, rot, NUM)
# eliminates those points that exceed the bound
myLoss = rep(0, nrow(points))
for(i in 1:nrow(points)){
myLoss[i] = totalloss(points[i,],X, y, lambda)
}
points = points[myLoss < bound, ]
}
rashomon = as.data.frame(points)
# model reliance for a single model - the center of the rashomon set
mr = modelreliance.full(beta_star, X, y)
# model class reliance for all models in the rashomon set
mcr = matrix(0, nrow = nrow(rashomon), ncol = length(beta_star)-1)
for(i in 1:nrow(rashomon)){
mcr[i, ] = modelreliance.full(as.double(rashomon[i, ]), X, y)
}
mcr = data.frame(mcr)
# characteristics of the rachomon set (from the mcr perspective)
cor(mcr)
apply(mcr, 2, max)
order(apply(mcr, 2, max), decreasing = TRUE)
apply(mcr, 2, min)
order(apply(mcr, 2, min), decreasing = TRUE)
apply(mcr, 2, mean)
myOrder = order(apply(mcr, 2, mean), decreasing = TRUE)
# clustering
k2 = kmeans(mcr, centers = 4, nstart = 25)
sub_grp = k2$cluster
# clustering results
cat = sub_grp
sum(cat==1)
sum(cat==2)
sum(cat==3)
sum(cat==4)
# find centers
red_center = order(rowSums(sweep(mcr, 2, k2$centers[1,])^2), decreasing = FALSE)[1]
green_center = order(rowSums(sweep(mcr, 2, k2$centers[2,])^2), decreasing = FALSE)[1]
blue_center = order(rowSums(sweep(mcr, 2, k2$centers[3,])^2), decreasing = FALSE)[1]
purple_center = order(rowSums(sweep(mcr, 2, k2$centers[4,])^2), decreasing = FALSE)[1]
model_id = c(red_center, green_center, blue_center, purple_center)
# most important variable to a cluster
mcr_red = apply(mcr[cat == 1,], 2, mean)
mcr_green = apply(mcr[cat == 2,], 2, mean)
mcr_blue = apply(mcr[cat == 3,], 2, mean)
mcr_purple = apply(mcr[cat == 4,], 2, mean)
mcr_avg = apply(mcr, 2, mean)
red = order(mcr_red - mcr_avg, decreasing = TRUE)
(mcr_red - mcr_avg)[red]
green = order(mcr_green - mcr_avg, decreasing = TRUE)
(mcr_green - mcr_avg)[green]
blue = order(mcr_blue - mcr_avg, decreasing = TRUE)
(mcr_blue - mcr_avg)[blue]
purple = order(mcr_purple - mcr_avg, decreasing = TRUE)
(mcr_purple - mcr_avg)[purple]
vlist = c(red[1], green[1], blue[1], purple[1])
mcr_small = mcr[,vlist]
vlist = id[vlist]
vname = NULL
for(i in 1:length(vlist)){
vname = c(vname, sprintf("feature_%i", vlist[i]))
}
# VID
color = c("brown2", "chartreuse3", "aquamarine2", "darkorchid1")
m = max(mcr_small) * 1.02
plot_by_cat = function(x, y , name1, name2, m, cat){
index = 1:length(x)
i = 1
plot(x[cat == i], y[cat == i], xlab = name1, ylab = name2, pch = ".", xlim = c(1,m), ylim = c(1,m), col = color[i],
cex.lab=1.25, cex.axis=1.25, cex.main=1.5, cex.sub=1)
for(i in 2:4){
points(x[cat == i], y[cat == i], pch = "*", col = color[i])
}
}
par(mfrow=c(4,4))
plot(0,type='n',axes=FALSE,ann=FALSE)
plot_by_cat(mcr_small[,2], mcr_small[,1], vname[2], vname[1], m, cat)
plot_by_cat(mcr_small[,3], mcr_small[,1], vname[3], vname[1], m, cat)
plot_by_cat(mcr_small[,4], mcr_small[,1], vname[4], vname[1], m, cat)
plot_by_cat(mcr_small[,1], mcr_small[,2], vname[1], vname[2], m, cat)
plot(0,type='n',axes=FALSE,ann=FALSE)
plot_by_cat(mcr_small[,3], mcr_small[,2], vname[3], vname[2], m, cat)
plot_by_cat(mcr_small[,4], mcr_small[,2], vname[4], vname[2], m, cat)
plot_by_cat(mcr_small[,1], mcr_small[,3], vname[1], vname[3], m, cat)
plot_by_cat(mcr_small[,2], mcr_small[,3], vname[2], vname[3], m, cat)
plot(0,type='n',axes=FALSE,ann=FALSE)
plot_by_cat(mcr_small[,4], mcr_small[,3], vname[4], vname[3], m, cat)
plot_by_cat(mcr_small[,1], mcr_small[,4], vname[1], vname[4], m, cat)
plot_by_cat(mcr_small[,2], mcr_small[,4], vname[2], vname[4], m, cat)
plot_by_cat(mcr_small[,3], mcr_small[,4], vname[3], vname[4], m, cat)
plot(0,type='n',axes=FALSE,ann=FALSE)
# visualization models
# model_id = c(36, 285, 50, 9)
model_coefficients = rashomon[model_id,]
filter_id = id
write.csv(model_coefficients, "model.csv")
write.csv(filter_id, "filter.csv")