/
mnist.py
177 lines (146 loc) · 5.53 KB
/
mnist.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
import matplotlib
matplotlib.use('Agg')
from matplotlib import pyplot
import numpy as np
from os.path import dirname, join as joinpath
from scipy.sparse.csgraph import connected_components
from sklearn.metrics import confusion_matrix
from sklearn.metrics.pairwise import pairwise_distances
from sklearn.semi_supervised import LabelPropagation
from neighborhood import neighbor_graph
from b_matching import hacky_b_matching
from msg import manifold_spanning_graph
from util import Timer, savefig
FIGURE_DIR = joinpath(dirname(__file__), 'figures')
MNIST_PATH = joinpath(dirname(__file__), 'data/')
def mnist_experiment(digits, opts):
X, GT, num_ccs = load_data(digits)
if opts.cache:
# TODO: include the digits used in the cache filename
D = np.load(joinpath(opts.cache, 'mnist_D.npy'))
Wknn = np.load(joinpath(opts.cache, 'mnist_Wknn.npy'))
Wbma = np.load(joinpath(opts.cache, 'mnist_Wbma.npy'))
Wmsg = np.load(joinpath(opts.cache, 'mnist_Wmsg.npy'))
else:
D, Wknn, Wbma, Wmsg = compute_Ws(X, num_ccs)
bad_edges = GT != GT[:,None]
edge_error(Wknn, bad_edges, 'knn')
edge_error(Wbma, bad_edges, 'b-matching')
edge_error(Wmsg, bad_edges, 'MSG')
if opts.sparsity:
plot_sparsity(D, Wknn, Wbma, Wmsg)
classify(Wknn, Wbma, Wmsg, GT, opts.classify)
def edge_error(W, bad_edges, name):
wrong = np.count_nonzero(W[bad_edges])
total = np.count_nonzero(W)
print '%s: err = %d/%d = %f' % (name, wrong, total, float(wrong)/total)
def load_data(digits=None):
X = np.load(joinpath(MNIST_PATH, 'test_data.npy'))
GT = np.load(joinpath(MNIST_PATH, 'test_labels.npy'))
if digits is not None:
assert all(0 <= d < 10 for d in digits)
mask = np.logical_or.reduce(GT==np.array(digits)[:,None])
num_ccs = len(digits)
GT = GT[mask]
X = X[mask].reshape((len(GT), -1))
else:
num_ccs = 10
X = X.reshape((len(GT), -1))
order = np.argsort(GT)
X = X[order]
GT = GT[order]
return X, GT, num_ccs
def compute_Ws(X, num_ccs):
with Timer('Calculating pairwise distances...'):
D = pairwise_distances(X, metric='sqeuclidean')
np.save('mnist_D.npy', D)
# k-nn
with Timer('Calculating knn graph...'):
for k in xrange(1,10):
Wknn = neighbor_graph(D, precomputed=True, k=k, symmetrize=True)
n = connected_components(Wknn, directed=False, return_labels=False)
if n <= num_ccs:
break
else:
assert False, 'k too low'
np.save('mnist_Wknn.npy', Wknn)
print 'knn (k=%d)' % k
# b-matching
with Timer('Calculating b-matching graph...'):
# using 8 decimal places kills the disk
Wbma = hacky_b_matching(D, k, fmt='%.1f')
np.save('mnist_Wbma.npy', Wbma)
# msg
with Timer('Calculating MSG graph...'):
Wmsg = manifold_spanning_graph(X, 2, num_ccs=num_ccs)
np.save('mnist_Wmsg.npy', Wmsg)
return D, Wknn, Wbma, Wmsg
def plot_sparsity(D, Wknn, Wbma, Wmsg):
# plot distances and sparsity patterns
pyplot.imshow(D, interpolation='nearest')
savefig('mnist_l2_dist.png', FIGURE_DIR)
pyplot.spy(Wknn, markersize=1)
savefig('mnist_knn_edges.png', FIGURE_DIR)
pyplot.spy(Wbma, markersize=1)
savefig('mnist_bma_edges.png', FIGURE_DIR)
pyplot.spy(Wmsg, markersize=1)
savefig('mnist_msg_edges.png', FIGURE_DIR)
def classify(Wknn, Wbma, Wmsg, GT, num_labeled):
# Note: in GT and labels, -1 means missing label
while True:
label_idx = np.random.choice(len(GT), size=num_labeled, replace=False)
labels = np.zeros(GT.shape, dtype=int) - 1
labels[label_idx] = GT[label_idx]
n = len(np.unique(labels))
if n == 11: # all labels represented
break
knn_res = edge_propagation(Wknn, labels)
bma_res = edge_propagation(Wbma, labels)
msg_res = edge_propagation(Wmsg, labels)
confusion(knn_res, 'knn', GT)
confusion(bma_res, 'bma', GT)
confusion(msg_res, 'msg', GT)
def confusion(result, name, GT):
header = '%s classifier: %d' % (name, np.count_nonzero(result == GT))
print header
cm = confusion_matrix(GT, result)
fname = joinpath(FIGURE_DIR, name+'_cm.tex')
with open(fname, 'w') as fh:
print >>fh, '%', header
print >>fh, '\\begin{tabular}{c|*{10}{r}}'
print >>fh, ' & 0 & 1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 \\\\'
print >>fh, '\\hline'
for i, row in enumerate(cm):
best = np.argmax(row)
nums = map(str, row)
nums[best] = '\\textbf{'+nums[best]+'}'
print >>fh, i, '&', ' & '.join(nums), '\\\\'
print >>fh, '\\end{tabular}'
## Simple edge propagator, hacked onto the sklearn version.
def edge_propagation(W, labels):
'''W is binary adj matrix, label -1 means unknown'''
W = W.astype(float)
np.fill_diagonal(W, 1)
P = np.array([row.nonzero()[0] for row in W])
return _LP(P=P).fit(W, labels).predict(W)
class _LP(LabelPropagation):
def __init__(self, P):
LabelPropagation.__init__(self, kernel='knn')
self.P = P
def _get_kernel(self, W, W2=None):
if W2 is None:
return W
return self.P
if __name__ == '__main__':
from optparse import OptionParser
op = OptionParser()
op.add_option('--digits', type=str, default='all',
help='comma-separated list of digits [all digits]')
op.add_option('--classify', type=int, metavar='N', default=20,
help='# of examples for MNIST classification test [20]')
op.add_option('--no-sparsity', action='store_false', dest='sparsity',
default=True, help="Don't plot sparsity figures.")
op.add_option('--cache', type=str, help='Path to cached .npy files')
opts, args = op.parse_args()
digits = None if opts.digits == 'all' else map(int,opts.digits.split(','))
mnist_experiment(digits, opts)