/
mnist_global.py
227 lines (192 loc) · 9.72 KB
/
mnist_global.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
# -*- coding: utf-8 -*-
# @Author: twankim
# @Date: 2017-02-24 17:46:51
# @Last Modified by: twankim
# @Last Modified time: 2018-03-09 22:14:25
import numpy as np
import time
import sys
import os
import argparse
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
from ssac import weakSSAC
from ssac_org import SSAC
# from gen_data import genData
from utils import *
import cPickle as pickle
weak = "global"
delta = 0.99
base_dir= os.path.join('./results',weak+'_compare_mnist')
def main(args):
plotted = False
# Load MNIST 2500 subset
with open('dataset/mnist2500.pkl','rb') as fp:
dataset = pickle.load(fp)
rep = args.rep
i_plot = 0
# qs = [float(q) for q in args.qs.split(',')]
etas = [float(eta) for eta in args.etas.split(',')]
beta = args.beta
verbose = args.verbose
cs = [float(q) for q in args.cs.split(',')]
res_acc = np.zeros((rep,len(cs),len(etas))) # Accuracy of clustering
res_mean_acc = np.zeros((rep,len(cs),len(etas))) # Mean accuracy of clustering (per cluster)
# res_err = np.zeros((rep,len(qs),len(etas))) # Number of misclustered points
res_fail = np.zeros((rep,len(cs),len(etas))) # Number of Failure
res_acc_org = np.zeros((rep,len(cs),len(etas))) # Accuracy of clustering
res_mean_acc_org = np.zeros((rep,len(cs),len(etas))) # Mean accuracy of clustering (per cluster)
# res_err = np.zeros((rep,len(qs),len(etas))) # Number of misclustered points
res_fail_org = np.zeros((rep,len(cs),len(etas))) # Number of Failure
gammas = np.zeros(rep)
rhos = np.zeros((rep,len(cs)))
digits = [str(int(dataset.dict_label[label])) for label in dataset.dict_label.keys()]
# Make directories to save results
if not os.path.exists(base_dir):
os.makedirs(base_dir)
res_dir = base_dir + '/{}'.format(
','.join(digits))
if not os.path.exists(res_dir):
os.makedirs(res_dir)
if verbose:
print "MNIST2500 Subset... Digits:{}".format(','.join(digits))
for i_rep in xrange(rep):
if verbose:
print "({}/{})... Testing Algorithm".format(i_rep+1,rep)
X = dataset.X
y_true = dataset.y
ris = dataset.ris
gamma = dataset.gamma
gammas[i_rep] = gamma
n = dataset.n
m = dataset.m
k = dataset.k
algo = weakSSAC(X,y_true,k,wtype=weak,ris=ris)
algo_org = SSAC(X,y_true,k,wtype=weak,ris=ris)
# Test SSAC algorithm for different c's and eta's (fix beta in this case)
for i_c,c_dist in enumerate(cs):
assert (c_dist>0.5) & (c_dist<=1.0), "c_dist must be in (0.5,1]"
rhos[i_rep,i_c] = c_dist
# Calculate proper eta and beta based on parameters including delta
if verbose:
print " - Proper eta={}, beta={} (delta={})".format(
dataset.calc_eta(delta,weak=weak,rho=rhos[i_rep,i_c]),
dataset.calc_beta(delta,weak=weak,rho=rhos[i_rep,i_c]),
delta)
for i_eta,eta in enumerate(etas):
if verbose:
print " <Test: c_dist={}, eta={}, beta={}>".format(c_dist,eta,beta)
algo.set_params(eta,beta,rho=rhos[i_rep,i_c])
algo_org.set_params(eta,beta,rho=rhos[i_rep,i_c])
if not algo.fit():
# Algorithm has failed
res_fail[i_rep,i_c,i_eta] = 1
if not plotted:
i_plot = np.random.randint(i_rep+1,rep) # Index of experiment to plot the figure
if not algo_org.fit():
# Algorithm has failed
res_fail_org[i_rep,i_c,i_eta] = 1
# i_plot = np.random.randint(i_rep+1,rep) # Index of experiment to plot the figure
y_pred = algo.y
mpps = algo.mpps # Estimated cluster centers
# print " ... Clustering is done. Number of binary search steps = {}\n".format(algo.bs_num)
y_pred_org = algo_org.y
mpps_org = algo_org.mpps # Estimated cluster centers
# For evaluation & plotting, find best permutation of cluster assignment
y_pred_perm = find_permutation(dataset,algo)
y_pred_perm_org = find_permutation(dataset,algo_org)
# Calculate accuracy and mean accuracy
res_acc[i_rep,i_c,i_eta] = accuracy(y_true,y_pred_perm)
res_mean_acc[i_rep,i_c,i_eta] = mean_accuracy(y_true,y_pred_perm)
res_acc_org[i_rep,i_c,i_eta] = accuracy(y_true,y_pred_perm_org)
res_mean_acc_org[i_rep,i_c,i_eta] = mean_accuracy(y_true,y_pred_perm_org)
# # Calculate number of errors
# res_err[i_rep,i_c,i_eta] = error(y_true,y_pred_perm)
if (i_rep == i_plot) and (m<=2) and (not plotted):
if (i_eta==len(etas)-1) and (i_c==len(cs)-1):
plotted = True
list_classes = ['Not assigned']
for i in xrange(k):
list_classes.append('Digit {}'.format(i))
title = r"SSAC with {} weak oracle ($\eta={}, \beta={}, \rho={:.2f}$)".format(
weak,eta,beta,rhos[i_rep,i_c])
f_name = res_dir+'/fig_n{}_m{}_k{}_c{:03d}_e{:d}.png'.format(n,m,k,int(100*c_dist),int(eta))
plot_cluster(X,y_true,y_pred_perm,k,mpps,gamma,
title,f_name,verbose,
classes=list_classes)
# title_org = r"SSAC(original) with {} weak oracle ($\eta={}, \beta={}, \rho={:.2f}$)".format(
# weak,eta,beta,rhos[i_rep,i_c])
# f_name_org = res_dir+'/fig_org_n{}_m{}_k{}_c{:03d}_e{:d}.png'.format(n,m,k,int(100*c_dist),int(eta))
# plot_cluster(X,y_true,y_pred_perm_org,k,mpps_org,gamma,
# title_org,f_name_org,verbose)
# Write result as table
print_eval("Accuracy(%)",res_acc,etas,
res_dir+'/res_{}_n{}_m{}_k{}.csv'.format("acc",n,m,k),weak=weak,params=cs)
print_eval("Mean Accuracy(%)",res_mean_acc,etas,
res_dir+'/res_{}_n{}_m{}_k{}.csv'.format("meanacc",n,m,k),weak=weak,params=cs)
# print_eval("# Error(%)",res_err,qs,etas,
# res_dir+'/res_{}_n{}_m{}_k{}.csv'.format("err",n,m,k))
print_eval("# Failures",res_fail,etas,
res_dir+'/res_{}_n{}_m{}_k{}.csv'.format("fail",n,m,k),
is_sum=True,weak=weak,params=cs)
print_eval("Accuracy(%)",res_acc_org,etas,
res_dir+'/res_org_{}_n{}_m{}_k{}.csv'.format("acc",n,m,k),weak=weak,params=cs)
print_eval("Mean Accuracy(%)",res_mean_acc_org,etas,
res_dir+'/res_org_{}_n{}_m{}_k{}.csv'.format("meanacc",n,m,k),weak=weak,params=cs)
# print_eval("# Error(%)",res_err,qs,etas,
# res_dir+'/res_{}_n{}_m{}_k{}.csv'.format("err",n,m,k))
print_eval("# Failures",res_fail_org,etas,
res_dir+'/res_org_{}_n{}_m{}_k{}.csv'.format("fail",n,m,k),
is_sum=True,weak=weak,params=cs)
# if args.isplot:
# Plot Accuracy vs. eta
fig_name = res_dir+'/fig_{}_n{}_m{}_k{}.pdf'.format("acc",n,m,k)
plot_eval("Accuracy(%)",res_acc,etas,fig_name,weak=weak,params=cs,res_org=res_acc_org)
# Plot Mean Accuracy vs. eta
fig_name = res_dir+'/fig_{}_n{}_m{}_k{}.pdf'.format("meanacc",n,m,k)
plot_eval("Mean Accuracy(%)",res_mean_acc,etas,fig_name,weak=weak,params=cs,res_org=res_mean_acc_org)
# Plot Failure vs. eta
fig_name = res_dir+'/fig_{}_n{}_m{}_k{}.pdf'.format("fail",n,m,k)
plot_eval("# Failure",res_fail,etas,fig_name,is_sum=True,weak=weak,params=cs,res_org=res_fail_org)
# # Plot histogram of gammas
# fig_name = res_dir+'/fig_gamma_hist.pdf'
# plot_hist(gammas,args.min_gamma,args.max_gamma,fig_name)
if args.isplot:
plt.show()
def parse_args():
def str2bool(v):
return v.lower() in ('true', '1')
parser = argparse.ArgumentParser(description=
'Test Semi-Supervised Active Clustering with Weak Oracles: Random-weak model')
# parser.add_argument('-m', dest='m',
# help='Dimension of data points in synthetic data',
# default = 2, type = int)
parser.add_argument('-rep', dest='rep',
help='Number of experiments to repeat',
default = 1000, type = int)
parser.add_argument('-qs', dest='qs',
help='Probabilities q (not-sure with 1-q) ex) 0.7,0.85,1',
default = '0.7,0.85,1', type = str)
parser.add_argument('-etas', dest='etas',
help='etas: parameter for sampling (phase 1) ex) 10,50',
default = '2,5,10,20,30', type = str)
parser.add_argument('-beta', dest='beta',
help='beta: parameter for sampling (phase 2)',
default = 1, type = int)
parser.add_argument('-cs', dest='cs',
help='Fractions to set distance-weak parameters (0.5,1] ex) 0.7,0.85,1',
default = '0.6,0.8,1', type = str)
parser.add_argument('-isplot', dest='isplot',
help='plot the result: True/False',
default = False, type = str2bool)
parser.add_argument('-verbose', dest='verbose',
help='verbose: True/False',
default = False, type = str2bool)
args = parser.parse_args()
return args
if __name__ == '__main__':
args = parse_args()
print "Called with args:"
print args
sys.exit(main(args))