This repository has been archived by the owner on Jul 17, 2020. It is now read-only.
/
matrix_layers.py
208 lines (179 loc) · 10.2 KB
/
matrix_layers.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
import theano
import theano.tensor as T
import theano.tensor.nlinalg as Tn
import numpy as np
import nn_utils as nnu
from layers import Layer, sample_gauss, sample_mult_noise, eps_ind, c1, c2, c3
class MatrixGaussDiagLayerFF(Layer):
def __init__(self, dim_in, dim_out, name, priors=(0., 0., 0.), N=1, nonlin='relu', type_init='he2', n_inducing=50,
noise_lvl=0.01):
sigma_in = 0.01
params = [nnu.randmat(dim_in, dim_out, 'mu_' + name, type_init=type_init, type_dist='normal'),
nnu.randvector(dim_in, 'sigma_row_mgauss_' + name, sigma=sigma_in),
nnu.randvector(dim_out, 'sigma_col_mgauss_' + name, sigma=sigma_in),
nnu.randmat2(n_inducing, dim_in, 'inducing_x_' + name, sigma=sigma_in, type_dist='uniform'),
nnu.randmat2(n_inducing, dim_out, 'inducing_y_' + name, sigma=sigma_in, type_dist='uniform'),
nnu.multvector(dim_in, np.log(np.sqrt(noise_lvl)), name='dropout_alpha_ffdrop_x_' + name),
nnu.multvector(dim_out, np.log(np.sqrt(noise_lvl)), name='dropout_alpha_ffdrop_y_' + name),
nnu.multvector(n_inducing, np.log(np.sqrt(noise_lvl)), name='dropout_alpha_ffdrop_pd_' + name)]
self.dim_in = dim_in
self.dim_out = dim_out
self.name = name
self.type_init = type_init
self.n_inducing = n_inducing
self.noise_lvl = noise_lvl
super(MatrixGaussDiagLayerFF, self).__init__(params, N=N, nonlin=nonlin, priors=priors)
def __str__(self):
return 'inducing m_gauss ffdrop layer ' + self.name
def _get_stds(self):
dx, dy, dpp = T.exp(self.params[-3]), T.exp(self.params[-2]), T.exp(self.params[-1])
stdx, stdy = T.outer(dpp, dx), T.outer(dpp, dy)
return stdx, stdy
def f(self, x, sampling=True, **kwargs):
x /= np.cast[theano.config.floatX](np.sqrt(self.dim_in))
indx, indy = self.params[3], self.params[4]
indx /= np.cast[theano.config.floatX](np.sqrt(self.dim_in))
if sampling:
stdx, stdy = self._get_stds()
noisex, noisey = sample_mult_noise(stdx, indx.shape), sample_mult_noise(stdy, indy.shape)
indy *= noisey; indx *= noisex
Rr, Rc = T.exp(self.params[1]), T.exp(self.params[2])
U = T.sqr(Rr)
sigma11 = T.dot(indx * U.dimshuffle('x', 0), indx.T) + eps_ind * T.eye(self.n_inducing)
sigma22 = T.dot(x * U.dimshuffle('x', 0), x.T)
sigma12 = T.dot(indx * U.dimshuffle('x', 0), x.T)
mu_ind = T.dot(indx, self.params[0])
inv_sigma11 = Tn.matrix_inverse(sigma11)
mu_x = T.dot(x, self.params[0]) + T.dot(sigma12.T, inv_sigma11).dot(indy - mu_ind)
if not sampling:
return mu_x
sigma_x = Tn.extract_diag(sigma22 - T.dot(sigma12.T, inv_sigma11).dot(sigma12))
std = T.outer(T.sqrt(sigma_x), Rc)
out_sample = sample_gauss(mu_x, std)
return out_sample
def get_reg_ind(self):
nsl = self.noise_lvl**2
constant = .5 * np.log(nsl) + c1 * nsl + c2 * (nsl**2) + c3 * (nsl**3)
stdx, stdy = self._get_stds()
drop_ax, drop_ay = T.pow(stdx, 2), T.pow(stdy, 2)
reg_indx = .5 * T.log(drop_ax) + c1 * drop_ax + c2 * T.pow(drop_ax, 2) + c3 * T.pow(drop_ax, 3) - constant
reg_indy = .5 * T.log(drop_ay) + c1 * drop_ay + c2 * T.pow(drop_ay, 2) + c3 * T.pow(drop_ay, 3) - constant
reg_ind = T.sum(reg_indx) + T.sum(reg_indy)
return reg_ind
def get_reg(self):
amount_reg = 1. / self.N
reg_ind = self.get_reg_ind()
reg = [amount_reg * self.kldiv_m(self.params[0], T.exp(self.params[1]), T.exp(self.params[2])),
amount_reg * reg_ind]
return reg
def get_priors(self):
pstdr = T.exp(self.priors[1])
pstdc = T.exp(self.priors[2])
return self.priors[0], pstdr, pstdc
def kldiv_m(self, mu, std_r, std_c):
pmu, pstdr, pstdc = self.get_priors()
var_r, var_c = T.sqr(std_r), T.sqr(std_c)
# first kl term
fa = T.sum((1./(pstdc**2)) * var_c)*T.sum((1./(pstdr**2))*var_r)
# second kl term
prior_sigma = T.outer(T.ones((mu.shape[0],))*(pstdr**2), T.ones((mu.shape[1],))*(pstdc**2))
fb = T.sum(T.sqr(mu - pmu) / prior_sigma)
# third kl term
fc = mu.shape[1]*(mu.shape[0]*T.log(pstdr**2) - T.sum(T.log(var_r))) + \
mu.shape[0]*(mu.shape[1]*T.log(pstdc**2) - T.sum(T.log(var_c)))
return - 0.5 * (fa + fb - T.prod(mu.shape) + fc)
def __copy__(self):
cpl = MatrixGaussDiagLayerFF(self.dim_in, self.dim_out, self.name, priors=self.priors, N=self.N,
nonlin=self.nonlin, type_init=self.type_init, n_inducing=self.n_inducing,
noise_lvl=self.noise_lvl)
return cpl
class MatrixGaussDiagLayerLearnP(Layer):
def __init__(self, dim_in, dim_out, name, priors=(0., 1., 1.), N=1, nonlin='relu', type_init='he2', n_inducing=50,
noise_lvl=0.01):
sigma_in = 0.01
self.a0, self.b0 = 1., .5
params = [nnu.randmat(dim_in, dim_out, 'mu_' + name, type_init=type_init),
nnu.randvector(dim_in, 'sigma_row_mgauss_' + name, sigma=sigma_in),
nnu.randvector(dim_out, 'sigma_col_mgauss_' + name, sigma=sigma_in),
nnu.randmat2(n_inducing, dim_in, 'inducing_x_' + name, sigma=sigma_in, type_dist='uniform'),
nnu.randmat2(n_inducing, dim_out, 'inducing_y_' + name, sigma=sigma_in, type_dist='uniform'),
nnu.multvector(1, np.log(self.a0), 'row_a_' + name), nnu.multvector(1, np.log(self.b0), 'row_b_' + name),
nnu.multvector(1, np.log(self.a0), 'col_a_' + name), nnu.multvector(1, np.log(self.b0), 'col_b_' + name),
nnu.tscalar(np.log(np.sqrt(noise_lvl)), 'scalar_dropout_alpha_x_' + name),
nnu.tscalar(np.log(np.sqrt(noise_lvl)), 'scalar_dropout_alpha_y_' + name)]
self.dim_in = dim_in
self.dim_out = dim_out
self.name = name
self.type_init = type_init
self.n_inducing = n_inducing
self.noise_lvl = noise_lvl
super(MatrixGaussDiagLayerLearnP, self).__init__(params, N=N, nonlin=nonlin, priors=priors)
def __str__(self):
return 'inducing m_gauss learn_p layer'
def f(self, x, sampling=True, **kwargs):
x /= np.cast[theano.config.floatX](np.sqrt(self.dim_in))
indx, indy = self.params[3], self.params[4]
indx /= np.cast[theano.config.floatX](np.sqrt(self.dim_in))
if sampling:
noisex = sample_mult_noise(T.exp(self.params[-2]), indx.shape)
noisey = sample_mult_noise(T.exp(self.params[-1]), indy.shape)
indy *= noisey; indx *= noisex
Rr, Rc = T.exp(self.params[1]), T.exp(self.params[2])
U = T.sqr(Rr)
sigma11 = T.dot(indx * U.dimshuffle('x', 0), indx.T) + eps_ind * T.eye(self.n_inducing)
sigma22 = T.dot(x * U.dimshuffle('x', 0), x.T)
sigma12 = T.dot(indx * U.dimshuffle('x', 0), x.T)
mu_ind = T.dot(indx, self.params[0])
inv_sigma11 = Tn.matrix_inverse(sigma11)
mu_x = T.dot(x, self.params[0]) + T.dot(sigma12.T, inv_sigma11).dot(indy - mu_ind)
if not sampling:
return mu_x
sigma_x = Tn.extract_diag(sigma22 - T.dot(sigma12.T, inv_sigma11).dot(sigma12))
std = T.outer(T.sqrt(sigma_x), Rc)
out_sample = sample_gauss(mu_x, std)
return out_sample
def get_reg_ind(self):
drop_ax, drop_ay = T.pow(T.exp(self.params[-2]), 2), T.pow(T.exp(self.params[-1]), 2)
constant = np.cast[theano.config.floatX](.5 * np.log(self.noise_lvl) + c1 * self.noise_lvl + c2 * (self.noise_lvl**2) + c3 * (self.noise_lvl**3))
reg_indx = .5 * T.log(drop_ax) + c1 * drop_ax + c2 * T.pow(drop_ax, 2) + c3 * T.pow(drop_ax, 3) - constant
reg_indy = .5 * T.log(drop_ay) + c1 * drop_ay + c2 * T.pow(drop_ay, 2) + c3 * T.pow(drop_ay, 3) - constant
reg_ind = T.cast(T.prod(self.params[3].shape), theano.config.floatX) * reg_indx + T.cast(T.prod(self.params[4].shape), theano.config.floatX) * reg_indy
return reg_ind
def get_reg(self, anneal=None):
amount_reg = 1. / self.N
reg_ind = self.get_reg_ind()
reg = [amount_reg * self.kldiv_m(self.params[0], T.exp(self.params[1]), T.exp(self.params[2])),
amount_reg * reg_ind,
amount_reg * (self.kldiv_r(*self._get_row_pl()) + self.kldiv_r(*self._get_col_pl()))]
return reg
def get_priors(self):
pstdr = T.exp(self.priors[1])
pstdc = T.exp(self.priors[2])
return self.priors[0], pstdr, pstdc
def _get_row_pl(self):
return T.exp(self.params[5]), T.exp(self.params[6])
def _get_col_pl(self):
return T.exp(self.params[7]), T.exp(self.params[8])
def kldiv_m(self, mu, std_r, std_c):
ar, br = self._get_row_pl()
ac, bc = self._get_col_pl()
etaur, etauc, elogtaur, elogtauc = (ar/br)[0], (ac/bc)[0], (nnu.Psi()(ar) - T.log(br))[0], (nnu.Psi()(ac) - T.log(bc))[0]
# first kl term
fa = T.sum(etauc * T.sqr(std_c))*T.sum(etaur * T.sqr(std_r))
# second kl term
prior_sigma = etaur * etauc
fb = prior_sigma * T.sum(T.sqr(mu))
# third kl term
fc = T.cast(- mu.shape[1]*(mu.shape[0]*elogtaur + T.sum(T.log(T.sqr(std_r)))) + \
- mu.shape[0]*(mu.shape[1]*elogtauc + T.sum(T.log(T.sqr(std_c)))), theano.config.floatX)
return - np.cast[theano.config.floatX](0.5) * T.cast((fa + fb - T.cast(T.prod(mu.shape), theano.config.floatX) + fc),
theano.config.floatX)
def kldiv_r(self, a1, b1):
return - ((a1 - self.a0)*nnu.Psi()(a1) - T.gammaln(a1) + T.gammaln(self.a0) +
self.a0*(T.log(b1) - T.log(self.b0)) + a1*((self.b0 - b1)/b1))[0]
def __copy__(self):
cpl = MatrixGaussDiagLayerLearnP(self.dim_in, self.dim_out, self.name, priors=self.priors, N=self.N,
nonlin=self.nonlin, type_init=self.type_init, n_inducing=self.n_inducing,
noise_lvl=self.noise_lvl)
return cpl
layers_def = {'mgdl_lp': MatrixGaussDiagLayerLearnP, 'mgdl_ff': MatrixGaussDiagLayerFF}