-
Notifications
You must be signed in to change notification settings - Fork 2
/
tf_toposparse.py
235 lines (192 loc) · 9.12 KB
/
tf_toposparse.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
228
229
230
231
232
233
234
235
# -*- coding: utf-8 -*-
from __future__ import absolute_import, division, print_function
import tensorflow as tf
import numpy as np
from tf_sparsenet import Sparsenet as snet
try:
import matplotlib.pyplot as plt
except ImportError:
print("Can't import matplotlib. No plotting.")
def block_diag(*arrs):
"""
copied from scipy.linalg.block_diag to avoid scipy dependency because long story
"""
if arrs == ():
arrs = ([],)
arrs = [np.atleast_2d(a) for a in arrs]
bad_args = [k for k in range(len(arrs)) if arrs[k].ndim > 2]
if bad_args:
raise ValueError("arguments in the following positions have dimension "
"greater than 2: %s" % bad_args)
shapes = np.array([a.shape for a in arrs])
out = np.zeros(np.sum(shapes, axis=0), dtype=arrs[0].dtype)
r, c = 0, 0
for i, (rr, cc) in enumerate(shapes):
out[r:r + rr, c:c + cc] = arrs[i]
r += rr
c += cc
return out
class TopoSparsenet(snet):
"""Topographic Sparsenet with TensorFlow backend
and a few methods for defining topologies."""
def __init__(self, data, datatype="image", pca=None,
dict_shape=(30, 30), topo=None, lam_g=0.1,
**kwargs):
"""
Topographic Sparsenet inherits from Sparsenet. Its unique
attributes give the dictionary a shape and define the relative
weight of the topographic term in the cost function.
The topology matrix g is defined by the topology object topo.
Args:
lam_g : float, defines weight of topography term
dict_shape : tuple (len, wid) of ints specifying shape of dictionary
"""
self.lam_g = lam_g
self.epsilon = 0.0001 # to regularize derivative of square root
self.dict_shape = dict_shape
nunits = int(np.prod(self.dict_shape))
self.topo = topo or topology((nunits, nunits))
nunits = self.topo.ncomponents * nunits
try:
kwargs['lam']
snet.__init__(self, data, nunits=nunits, datatype=datatype, pca=pca, **kwargs)
except KeyError:
snet.__init__(self, data, nunits=nunits, datatype=datatype, pca=pca, lam=0, **kwargs)
def build_graph(self):
graph = tf.get_default_graph()
self.g = tf.constant(self.topo.get_matrix(), dtype=tf.float32)
self._infrate = tf.Variable(self.infrate, trainable=False)
self._learnrate = tf.Variable(self.learnrate, trainable=False)
self.phi = tf.Variable(self.Q)
self.acts = tf.Variable(tf.zeros([self.nunits, self.batch_size]))
self.reset_acts = self.acts.assign(tf.zeros([self.nunits, self.batch_size]))
self.x = tf.Variable(tf.zeros([self.batch_size, self.stims.datasize]), trainable=False)
self.xhat = tf.matmul(tf.transpose(self.acts), self.phi, name='xhat')
self.resid = self.x - self.xhat
self.mse = tf.reduce_sum(tf.square(self.resid))/self.batch_size/self.stims.datasize
self.meanL1 = tf.reduce_sum(tf.abs(self.acts))/self.batch_size
self.layer2 = tf.reduce_sum(tf.sqrt(tf.matmul(self.g, tf.square(self.acts),
name='g_times_acts') + self.epsilon))/self.batch_size
self.loss = 0.5*self.mse + (self.lam*self.meanL1 + self.lam_g*self.layer2)/self.stims.datasize
inffactor = self.batch_size*self.stims.datasize
inferer = tf.train.GradientDescentOptimizer(self._infrate*inffactor)
self.inf_op = inferer.minimize(self.loss, var_list=[self.acts])
learner = tf.train.GradientDescentOptimizer(self.learnrate)
learn_step = tf.Variable(0,name='learn_step', trainable=False)
self.learn_op = learner.minimize(self.loss, global_step=learn_step, var_list=[self.phi])
self._ma_variances = tf.Variable(self.ma_variances, trainable=False)
self._gains = tf.Variable(self.gains, trainable=False)
_, self.variances = tf.nn.moments(self.acts, axes=[1])
vareta = self.var_avg_rate
newvar = (1.-vareta)*self._ma_variances + vareta*self.variances
self.update_variance = self._ma_variances.assign(newvar)
newgain = self.gains*tf.pow(self.var_goal/self._ma_variances,
self.gain_rate)
self.update_gains = self._gains.assign(newgain)
normphi = (tf.expand_dims(self._gains,
dim=1)*tf.nn.l2_normalize(self.phi, dim=1))
self.renorm_phi = self.phi.assign(normphi)
self._init_op = tf.global_variables_initializer()
return graph
def show_dict(self, cmap='RdBu_r', layout=None, savestr=None):
Qs = self.Q
layout = layout or self.dict_shape
ncomp = self.topo.ncomponents
per_comp = np.prod(layout)
nn = 0
display = self.stims.stimarray(Qs[nn*per_comp:(nn+1)*per_comp], layout=layout)
for nn in range(1,ncomp):
display = np.concatenate([display, self.stims.stimarray(Qs[nn*per_comp:(nn+1)*per_comp], layout=layout)],
axis=0)
plt.figure()
arrayplot = plt.imshow(display, interpolation='nearest', cmap=cmap, aspect='auto', origin='lower')
plt.axis('off')
plt.colorbar()
if savestr is not None:
plt.savefig(savestr, bbox_inches='tight')
return display
def sort(self, *args, **kwargs):
print("The topographic order is meaningful, don't sort it away!")
def get_param_list(self):
params = snet.get_param_list(self)
params['lam_g'] = self.lam_g
return params
class topology():
def __init__(self, shape, discs=True, torus=True, binary=True, sigma = 1.0, ncomponents = 1):
"""
shape: (tuple) (nlayer2comp, nlayer1) shape of each component
sigma : (float) defines stdev of default gaussian neighborhoods
"""
self.shape = shape
dict_side = int(np.sqrt(self.shape[1]))
assert dict_side**2 == self.shape[1], 'Only square dictionaries supported.'
self.dict_shape = (dict_side, dict_side)
self.discs = discs
self.torus = torus
self.binary = binary
self.sigma = sigma
self.ncomponents = ncomponents
def get_matrix(self):
g = np.zeros(self.shape)
if self.discs:
g = self.make_discs(g, *self.shape)
if self.ncomponents > 1:
blocks = [g.copy() for ii in range(self.ncomponents)]
g = block_diag(*blocks)
if self.binary:
g = self.binarize(g)
return g
def make_discs(self, g, nlayer2, nlayer1):
sigsquared = self.sigma**2
for i in range(nlayer2):
for j in range(nlayer1):
g[i, j] = np.exp(-self.distance(i, j)/(2 * sigsquared))
return g
def distance(self, i, j):
""" This function measures the squared distance between element i and j. The distance
here is the distance between element i and j once the row vector has been
reshaped into a square matrix, treating the dictionary as a torus globally
if torus is True."""
rows, cols = self.dict_shape
rowi = i // cols
coli = i % cols
rowj = j // cols
colj = j % cols
if self.torus:
# global topology is a torus
rowj = [rowj - rows, rowj, rowj + rows]
colj = [colj - cols, colj, colj + cols]
dist = []
for r in rowj:
for c in colj:
dist.append((rowi - r)**2 + (coli - c)**2)
return np.min(dist)
else:
return (rowi - rowj)**2 + (coli - colj)**2
def block_membership(self, i, j, width=5):
"""This returns 1 if j is in the ith block, otherwise 0. Currently only
works for square dictionaries."""
# FIXME: I think there's a bug here that makes the boundary conditions
# and the sizes wrong
size = self.dict_shape[0]
if size != self.dict_shape[1]:
raise NotImplementedError
i = [i // size, i % size]
j = [j // size, j % size]
if (abs((i[0]%size)-(j[0]%size)) % (size-1) < width) and (abs((i[1]%size)-(j[1]%size)) % (size-1) < width):
return 1
else:
return 0
def set_blocks(self, width=5):
"""Change the topography by making each second layer unit respond to
a square block of layer one with given width. g becomes binary."""
# FIXME: doesn't work because block_membership doesn't work
self.g = np.zeros_like(self.g)
nunits = np.prod(self.dict_shape)
for i in range(nunits):
for j in range(nunits):
self.g[i, j] = self.block_membership(i, j, width)
def binarize(self, g, thresh=1/2, width=None):
if width is not None:
thresh = np.exp(-width**2/(2*self.sigma**2))
return np.array(g >= thresh, dtype=int)