-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathhhmm.py
167 lines (156 loc) · 8.33 KB
/
hhmm.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
import pickle
import time
import numpy as np
import tensorflow as tf
from scipy.misc import logsumexp
class PHPTrainer(object):
def __init__(self, theta, o, a, hierarchy, opt=tf.train.AdamOptimizer(1e-4)):
"""
:param o: [T, O]
:param a: [T, A]
hierarchy.log_eta: [T + 1, O] -> [T, H, TAU, H+K] ; log_eta[t, h, tau, h'] = log eta_h^tau(h' | o_t)
hierarchy.log_pi: [T + 1, O], [T, A] -> [T, K, TAU] ; log_pi[t, h, tau] = log pi_h^tau(a_t | o_t)
hierarchy.log_psi: [T + 1, O] -> [T, H+K, TAU, 2] ; log_psi[t, h, tau, 1] = log psi_h^tau(o_{t+1})
hierarchy.log_init: [T + 1, O] -> [M] ; log_init[m] = log Pr(m_0 = m | o_0)
hierarchy.log_tran: [T + 1, O] -> [T-1, M, M] ; log_tran[t, m, m'] = log Pr(m_{t+1} = m' | m_t = m, o_{t+1})
hierarchy.log_term: [T + 1, O] -> [M] ; log_term[m] = log Pr(m_T = [] | m_{T-1} = m, o_T)
hierarchy.log_emit: [T + 1, O], [T, A] -> [T, M] ; log_emit[t, m] = log Pr(a_t | m_t = m, o_t)
"""
self.theta = theta
self.o = o
self.a = a
self.hierarchy = hierarchy
M = self.hierarchy.num_states
self.vij = tf.placeholder(tf.float32, [None, M, M]) # [T-1, M, M]
self.vi = tf.placeholder(tf.float32, [None, M]) # [T, M]
self.entropy_weight = tf.placeholder(tf.float32, [])
self.gradients = list(zip(*opt.compute_gradients(self.make_loss(), self.theta)))[0]
self.opt = opt.apply_gradients(zip(self.gradients, self.theta))
def make_loss(self):
w_eta = tf.tensordot(self.vi, self.hierarchy.get_w_eta_selector(), 1) # [T, H, TAU, H+K]
w_pi = tf.tensordot(self.vi, self.hierarchy.get_w_pi_selector(), 1) # [T, K, TAU]
w_psi = tf.concat([
tf.tensordot(self.vij, self.hierarchy.get_w_psi_selector(), [[1, 2], [0, 1]]),
tf.tensordot(self.vi[-1:, :], self.hierarchy.get_w_psi_term_selector(), 1)], 0) # [T, H+K, TAU, 2]
log_prob = (tf.reduce_sum(w_eta * self.hierarchy.log_eta)
+ tf.reduce_sum(w_pi * self.hierarchy.log_pi)
+ tf.reduce_sum(w_psi * self.hierarchy.log_psi))
entropy = -(tf.reduce_sum(tf.exp(self.hierarchy.log_eta) * self.hierarchy.log_eta)
+ tf.reduce_sum(tf.exp(self.hierarchy.log_pi) * self.hierarchy.log_pi)
+ tf.reduce_sum(tf.exp(self.hierarchy.log_psi) * self.hierarchy.log_psi))
return -log_prob - self.entropy_weight * entropy
@staticmethod
def forward(results):
log_init = results['log_init']
log_tran = results['log_tran']
log_emit = results['log_emit']
T, M = log_emit.shape
log_phi = np.full([T, M], -np.inf, np.float32)
log_phi[0, :] = log_init[:]
for t in range(T - 1):
log_phi[t + 1, :] = logsumexp(log_phi[t, :, None] + log_emit[t, :, None] + log_tran[t, :, :], 0)
results['log_phi'] = log_phi
@staticmethod
def backward(results):
log_tran = results['log_tran']
log_term = results['log_term']
log_emit = results['log_emit']
T, M = log_emit.shape
log_omega = np.full([T, M], -np.inf, np.float32)
log_omega[T - 1, :] = log_emit[T - 1, :] + log_term[:]
for t in range(T - 2, -1, -1):
log_omega[t, :] = log_emit[t, :] + logsumexp(log_tran[t, :, :] + log_omega[t + 1, None, :], 1)
results['log_omega'] = log_omega
def opt_step(self, traj, entropy_weight):
sess = tf.get_default_session()
results = sess.run(self.hierarchy.all, feed_dict={self.o: traj.o, self.a: traj.a})
log_tran = results['log_tran']
log_emit = results['log_emit']
T, M = log_emit.shape
if hasattr(traj, 'm'):
vij = np.zeros([T - 1, M, M])
vi = np.zeros([T, M])
for t in range(T - 1):
vij[t, traj.m[t], traj.m[t + 1]] = 1.
vi[:-1, :] = vij.sum(2)
vi[-1, traj.m[-1]] = 1.
else:
self.forward(results)
self.backward(results)
log_phi = results['log_phi']
log_omega = results['log_omega']
log_likelihood = logsumexp(log_phi[0, :] + log_omega[0, :])
vij = np.exp(log_phi[:-1, :, None] + log_emit[:-1, :, None] + log_tran[:, :, :] + log_omega[1:, None, :] - log_likelihood)
vi = np.zeros([T, M])
vi[:-1, :] = vij.sum(2)
vi[-1, :] = np.exp(log_phi[-1, :] + log_omega[-1, :] - log_likelihood)
return sess.run(self.gradients, feed_dict={self.o: traj.o, self.a: traj.a, self.vij: vij, self.vi: vi, self.entropy_weight: entropy_weight})
def score(self, traj, correctness=True):
sess = tf.get_default_session()
results = sess.run(self.hierarchy.all, feed_dict={self.o: traj.o, self.a: traj.a})
self.backward(results)
log_init = results['log_init']
log_tran = results['log_tran']
log_term = results['log_term']
log_emit = results['log_emit']
log_omega = results['log_omega']
log_likelihood = logsumexp(log_init[:] + log_omega[0, :])
if not correctness:
return log_likelihood
T = traj.length
m_llik = log_init[traj.m[0]] + sum(log_tran[t, traj.m[t], traj.m[t + 1]] for t in range(T - 1)) + log_term[traj.m[-1]]
a_llik = sum(log_emit[t, traj.m[t]] for t in range(T))
m_corr = (log_init.argmax() == traj.m[0]) + (log_term.argmax() == traj.m[-1]) + sum(
(log_tran[t].argmax() == [traj.m[t], traj.m[t + 1]]).all()
for t in range(T - 1))
a_corr = sum(log_emit[t].argmax() == traj.m[t] for t in range(T))
return log_likelihood, m_llik, a_llik, m_corr, a_corr
def train(self, out_path, train_trajs, valid_trajs, num_steps=1000):
sess = tf.get_default_session()
start_time = time.time()
for step in range(num_steps + 1):
entropy_weight = 1e-2 / (step + 1.)
if step % 10 == 0 or step == num_steps:
its_per_sec = (step + 1) / (time.time() - start_time)
train_log_likelihood = 0.
total_length = 0
for traj in train_trajs:
train_log_likelihood += self.score(traj, False)
total_length += traj.length
train_log_likelihood /= total_length
total_log_likelihood = 0.
total_m_llik = 0
total_a_llik = 0
total_m_corr = 0
total_a_corr = 0
total_length = 0
for traj in valid_trajs:
log_likelihood, m_llik, a_llik, m_corr, a_corr = self.score(traj)
total_log_likelihood += log_likelihood
total_m_llik += m_llik
total_a_llik += a_llik
total_m_corr += m_corr
total_a_corr += a_corr
total_length += traj.length
mean_log_likelihood = total_log_likelihood / total_length
mean_m_llik = total_m_llik / (total_length + len(valid_trajs))
mean_a_llik = total_a_llik / total_length
mean_m_corr = total_m_corr / (total_length + len(valid_trajs))
mean_a_corr = total_a_corr / total_length
print('step #{} train ll: {}, test ll: {}, m llik: {}, a llik: {}, m corr: {}, a corr: {}, ent w: {} ({} it/s)'
.format(step, train_log_likelihood, mean_log_likelihood, mean_m_llik, mean_a_llik, mean_m_corr, mean_a_corr, entropy_weight, its_per_sec))
if step % 100 == 0:
pickle.dump(sess.run(self.theta), open('{}/snapshot{:06d}.{}.pkl'.format(out_path, step, round(mean_log_likelihood, 2)), 'wb'))
if step < num_steps:
grad = None
for traj in train_trajs:
grad = self.acc_grad(grad, self.opt_step(traj, entropy_weight))
sess.run(self.opt, feed_dict={self.gradients: grad})
pickle.dump(sess.run(self.theta), open('{}/params.pkl'.format(out_path), 'wb'))
return self.theta
@staticmethod
def acc_grad(acc, grad):
if acc is None:
return grad
assert (len(acc) == len(grad))
return [acc[i] + grad[i] for i in range(len(acc))]