-
Notifications
You must be signed in to change notification settings - Fork 0
/
kmeans.py
184 lines (137 loc) · 5.52 KB
/
kmeans.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
#! /usr/bin/env python
# -*- coding: utf-8 -*-
# vim:fenc=utf-8
#
# Copyright © 2016 romancpodolski <roman.podolski@tum.de>
#
# Distributed under terms of the MIT license.
"""
TODO
References:
- papers: "Learning Feature Represantations with K-means" -
Adam Coates and Andreq Y. Ng
"""
from __future__ import print_function
import theano
import theano.tensor as T
import numpy as np
import scipy.misc
import matplotlib.pyplot as plt
import sys
import os
import six.moves.cPickle as pickle
sys.path.append(os.path.join(os.path.split(__file__)[0], '..', 'data'))
from data import load_data
from data import shared_dataset
from utils import tile_raster_images
try:
import PIL.Image as Image
except ImportError:
import Image
class K_Means(object):
def __init__(self, n_dim, n_samples, k = 500, input = None):
"""Instance of k-Means object
:type n_dim: int
:param n_dim: TODO
:type n_samples: int
:param n_samples: TODO
:type k: int
:param k: number of centroids used, columns in the Dictionary D
"""
self.k = k
self.input = input
self.D = theano.shared(
name = 'D',
value = np.asarray(
np.random.normal(0, size = (n_dim, k)),
dtype = theano.config.floatX
),
borrow = True
)
self.S = theano.shared(
name = 'S',
value = np.zeros(
(k, n_samples),
dtype = theano.config.floatX
),
borrow = True
)
max_position = T.argmax(abs(T.dot(self.D.T, self.input)), axis = 0)
max_values = T.dot(self.D.T, self.input)[max_position, T.arange(max_position.shape[0])]
zero_sub = T.zeros_like(self.S)[max_position, T.arange(max_position.shape[0])]
self.S_update = T.set_subtensor(zero_sub, max_values)
D_update = T.dot(self.input, self.S.T) + self.D
# self.D_norm = D_update / D_update.norm(2)
self.D_norm = D_update / T.sqrt(T.sum(T.sqr(D_update), axis = 0))
self.cost = T.sum(T.sqrt(T.sum(T.sqr(T.dot(self.D, self.S) - self.input), axis = 0)))
def train(dataset = 'cifar-10-python.tar.gz', n_classes = 500, max_iter = 10, batch_size = 600):
datasets = load_data(dataset)
train_set_x, train_set_y = datasets[0]
n_train_batches = train_set_x.shape[0] // batch_size
print(batch_size)
print('... resizing the input from 32x32 pixels to 12x12 to pixels')
resized = []
for i in xrange(train_set_x.shape[0]):
resized.append(scipy.misc.imresize(np.reshape(train_set_x[i], (32, 32)), (12, 12)).flatten() )
train_set_x = np.asarray(resized)
print('... normalize input')
epsilon = 10
train_set_x = (train_set_x - np.mean(train_set_x, axis = 1)[:, np.newaxis]) / np.sqrt(np.var(train_set_x, axis = 1) + epsilon)[:, np.newaxis]
print('... whiten input')
d, V = np.linalg.eig(np.cov(train_set_x.T))
D = np.diag(d)
I = np.eye(d.shape[0])
epsilon = 0.01
train_set_x = np.dot(np.dot(np.dot(V, np.linalg.inv(np.sqrt(D + epsilon * I))), V.T), train_set_x.T)
train_set = (train_set_x.T, train_set_y)
train_set = shared_dataset(train_set)
train_set_x = train_set[0]
print('... building the model')
x = T.dmatrix('X')
classifier = K_Means(n_dim = 12 * 12, n_samples = batch_size, k = n_classes, input = x)
updates = [(classifier.S, classifier.S_update),
(classifier.D, classifier.D_norm)]
index = T.lscalar('index')
train_model = theano.function(
inputs = [index],
outputs = classifier.cost,
updates = updates,
givens = {
x: train_set_x[index * batch_size: (index + 1) * batch_size].T,
}
)
print('... training the model')
for epoch in range(max_iter):
cost = []
for minibatch_index in range(n_train_batches):
cost = train_model(minibatch_index)
print('epoch %i, minibatch %i, %i, cost %.2f ' % (epoch + 1, minibatch_index + 1, n_train_batches, cost))
with open(os.path.join(os.path.split(__file__)[0], 'best_model.pkl'), 'wb') as f:
pickle.dump(classifier, f)
def plot():
import scipy.ndimage
classifier = pickle.load(open(os.path.join(os.path.split(__file__)[0], 'best_model.pkl')))
print('... magnifying the receptive fields using bilinear interpolation to 36x36 pixels')
resized = []
for i in xrange(classifier.D.get_value(borrow = True).T.shape[0]):
resized.append(scipy.ndimage.zoom(np.reshape(classifier.D.get_value(borrow = True).T[i], (12, 12)), 3, order = 1).flatten() )
repflds = np.asarray(resized)
print('... ploting the receptive fields with 36x36 pixels')
image = Image.fromarray(tile_raster_images(X = repflds, img_shape = (36, 36), tile_shape = (10, 50), tile_spacing=(1, 1)))
print('... saving image to %s' % os.path.join(os.path.split(__file__)[0], 'repflds.png'))
image.save(os.path.join(os.path.split(__file__)[0], 'repflds.png'))
def main(argv):
if len(argv) < 1:
print("please call with at least 1 argument")
return -1
command = argv[0]
if command == 'train':
return train()
elif command == 'plot':
return plot()
else:
print('unknown command: %' % command)
print("either use 'train' or 'plot'")
return -1
if __name__ == "__main__":
sys.exit(main(sys.argv[1:]))