forked from probml/pyprobml
-
Notifications
You must be signed in to change notification settings - Fork 0
/
binary_fa_demo.py
185 lines (160 loc) · 6.47 KB
/
binary_fa_demo.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
import numpy as np
import matplotlib.pyplot as plt
from scipy.stats import multivariate_normal
from jax import vmap
class BinaryFA:
def __init__(self, input_dim, latent, max_iter, conv_tol=1e-4, compute_ll=True):
self.W = 0.1 * np.random.randn(latent, input_dim) # 2x16
self.b = 0.01 * np.random.randn(input_dim, 1) # 16x1
self.mu_prior = np.zeros((latent,1)) # 2x1
self.sigma_prior = np.eye(latent) # 2x2
self.input_dim = input_dim
self.latent = latent
self.max_iter = max_iter
self.compute_ll = compute_ll
if compute_ll :
self.ll_hist = np.zeros((max_iter + 1, 1)) # 51x1
def variational_em(self, data):
ll_hist = np.zeros((self.max_iter + 1, 1))
i = 0
while i < 3:
S1, S2, ll = self.estep(data)
ll_hist[i,0] = ll
self.mstep(S1, S2)
if i!=0:
delta_fval = abs(ll_hist[i] - ll_hist[i-1])
avg_fval = (abs(ll_hist[i]) + abs(ll_hist[i-1]) + np.finfo(float).eps)/2
if (delta_fval / avg_fval) < conv_tol:
break
i += 1
return ll_hist[:i]
def estep(self, data):
S1 = np.zeros((self.latent + 1, self.input_dim)) # 3x16
S2 = np.zeros((self.latent + 1, self.latent + 1, self.input_dim)) # 3x3x16
W, b, mu_prior = self.W , self.b, self.mu_prior
ll = 0
for i in range(data.T.shape[1]):
mu_post, sigma_post, logZ, lambd = self.compute_latent_posterior_statistics(data.T[:,i], max_iter=3)
ll += logZ
EZZ = np.zeros((self.latent+1, self.latent+1))
EZZ[:self.latent,:self.latent] = sigma_post + np.outer(mu_post, mu_post)
EZZ[self.latent,:self.latent] = mu_post.T
EZZ[:self.latent,self.latent] = np.squeeze(np.asarray(mu_post))
EZZ[self.latent,self.latent] = 1
EZ = np.append(mu_post,np.ones((1,1)))
for j in range(self.input_dim):
S1[:,j] = S1[:,j] + (data.T[j,i] - 0.5) * EZ
S2[:,:,j] = S2[:,:,j] - 2* lambd[j] * EZZ
return S1, S2, ll
def mstep(self, S1, S2):
for i in range(self.input_dim):
what = np.linalg.lstsq(S2[:,:,i] , S1[:,i])[0]
self.W[:,i] = what[:self.latent]
self.b[i] = what[self.latent]
def compute_latent_posterior_statistics(self, y, output=[0,0,0,0], max_iter=3):
W, b = np.copy(self.W), np.copy(self.b)
y = y.reshape((-1,1))
# variational parameters
mu_prior = self.mu_prior
xi = (2 * y -1) * (W.T @ mu_prior + b)
xi[xi==0] = 0.01 * np.random.rand(np.count_nonzero(xi==0)) # 16x1
sigma_inv, iter = np.linalg.inv(self.sigma_prior), 0
for iter in range(max_iter):
lambd = (0.5 - sigmoid(xi)) / (2*xi)
tmp = W @ np.diagflat(lambd) @ W.T # 2x2
sigma_post = np.linalg.inv(sigma_inv - (2 * tmp))
tmp = y -0.5 + 2* lambd * b
tmp2 = np.sum(W @ np.diagflat(tmp), axis=1).reshape((2,1))
mu_post = sigma_post @ (sigma_inv @ mu_prior + tmp2)
tmp = np.diag(W.T @ (sigma_post + mu_post @ mu_post.T) @ W)
tmp = tmp.reshape((tmp.shape[0],1))
tmp2 = 2*(W @ np.diagflat(b)).T @ mu_post
xi = np.sqrt(tmp + tmp2 + b**2)
logZ = 0
if self.compute_ll:
lam = -lambd
A = np.diagflat(2*lam)
invA = np.diagflat(1/(2*lam))
bb = -0.5 * np.ones((y.shape[0],1))
c = -lam * xi**2 - 0.5 * xi + np.log(1+ np.exp(xi))
ytilde = invA @ (bb + y)
B = W.T
logconst1 = -0.5* np.sum(np.log(lam/np.pi))
logconst2 = 0.5 * ytilde.T @ A @ ytilde - np.sum(c)
gauss = multivariate_normal.logpdf(np.squeeze(np.asarray(ytilde)), mean=np.squeeze(np.asarray(B @ mu_prior + b)), cov=(invA + B @ sigma_post @ B.T))
logZ = logconst1 + logconst2 + gauss
output = [mu_post, sigma_post, logZ,lambd]
return output
def predict_missing(self, y):
N, T = y.shape # 150 x 16
prob_on = np.zeros(y.shape) # 150 x 16
post_pred = np.zeros((N,T,2)) # 150 x 16 x 2
L,p = self.W.shape # 16 x 3
B = np.c_[np.copy(self.b),self.W.T] # 16 x 3
for n in range(N):
mu_post, sigma_post, logZ, lambd = self.compute_latent_posterior_statistics(y[n,:].T, False)
mu1 = np.r_[np.ones((1,1)), mu_post]
sigma1 = np.zeros((L+1,L+1))
sigma1[1:,1:] = sigma_post
prob_on[n,:] = sigmoid_times_gauss(B, mu1, sigma1)
return prob_on
def infer_latent(self, y):
N, T = y.shape
W, b, mu_prior = self.W, self.b, self.mu_prior
K, T2 = self.W.shape
mu_post, loglik = np.zeros((K,N)),np.zeros((1,N))
sigma_post = np.zeros((K,K,N))
for n in range(N):
mu_p , sigma_p, loglik[0,n] , _ = self.compute_latent_posterior_statistics(y[n,:].T)
mu_post[:,n] = np.squeeze(np.asarray(mu_p))
sigma_post[:,:,n] = np.squeeze(np.asarray(sigma_p))
return mu_post, sigma_post, loglik
def sigmoid_times_gauss(X, wMAP, C):
vv = lambda x, y: jnp.vdot(x, y)
mv = vmap(vv, (None, 0), 0)
mm = vmap(mv, (0, None), 0)
vm = vmap(vv, (0, 0), 0)
mu = X @ wMAP;
n = X.shape[1]
if n < 1000:
sigma2 = np.diag(X @ C @ X.T)
else:
sigma2 = vm(X , mm(C,X))
kappa = 1 / np.sqrt(1 + np.pi * sigma2 /8);
p = sigmoid(kappa * mu.reshape(kappa.shape))
return p
np.random.seed(1)
max_iter, conv_tol = 50, 1e-4
sigmoid = lambda x : 1/(1 + np.exp(-1 * x))
d, k, m = 16, 3, 50
noise_level = 0.5
proto = np.random.rand(d, k) < noise_level
src = np.concatenate((np.tile(proto[:,0], (1,m)), np.tile(proto[:,1],(1,m)), np.tile(proto[:,2],(1,m))),axis=1)
clean_data = np.concatenate((np.tile(proto[:,0], (m,1)), np.tile(proto[:,1],(m,1)), np.tile(proto[:,2],(m,1))), axis=0)
n = clean_data.shape[0]
mask, noisy_data, missing_data, = np.random.rand(n,d) < 0.05, np.copy(clean_data), np.copy(clean_data)
noisy_data[mask] = 1 - noisy_data[mask]
missing_data[mask] = np.nan
ax = plt.gca()
plt.imshow(noisy_data, aspect='auto', interpolation='none',
origin='lower', cmap="gray")
plt.title('Noisy Binary Data')
plt.show()
binaryFA = BinaryFA(d, 2, 50, 1e-4, True)
binaryFA.variational_em(noisy_data)
mu_post, sigma_post, loglik = binaryFA.infer_latent(noisy_data)
symbols = ['ro', 'gs', 'k*']
plt.plot(mu_post[0,:m], mu_post[1,0:m], symbols[0])
plt.plot(mu_post[0,m:2*m], mu_post[1,m:2*m], symbols[1])
plt.plot(mu_post[0,2*m:], mu_post[1,2*m:], symbols[2])
plt.title('Latent Embedding')
plt.show()
prob_on = binaryFA.predict_missing(noisy_data)
plt.imshow(prob_on, aspect='auto', interpolation='none',
origin='lower', cmap="gray")
plt.title('Posterior Predictive')
plt.show()
plt.imshow(prob_on>0.5, aspect='auto', interpolation='none',
origin='lower', cmap="gray")
plt.title('Reconstruction')
plt.show()