-
Notifications
You must be signed in to change notification settings - Fork 8
/
ram.py
187 lines (149 loc) · 5.68 KB
/
ram.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
import chainer
from chainer import cuda
import chainer.functions as F
import chainer.links as L
import numpy as np
from crop import crop
class RAM(chainer.Chain):
def __init__(
self, g_size=8, n_steps=6, n_scales=1, var=0.03, use_lstm=False
):
d_glm = 128
d_core = 256
super(RAM, self).__init__(
emb_l=L.Linear(2, d_glm),
emb_x=L.Linear(g_size*g_size*n_scales, d_glm),
fc_lg=L.Linear(d_glm, d_core),
fc_xg=L.Linear(d_glm, d_core),
fc_ha=L.Linear(d_core, 10),
fc_hl=L.Linear(d_core, 2),
fc_hb=L.Linear(d_core, 1),
)
if use_lstm:
self.add_link(name='core_lstm', link=L.LSTM(d_core, d_core))
else:
self.add_link(name='core_hh', link=L.Linear(d_core, d_core))
self.add_link(name='core_gh', link=L.Linear(d_core, d_core))
self.use_lstm = use_lstm
self.d_core = d_core
self.g_size = g_size
self.n_steps = n_steps
self.n_scales = n_scales
self.var = var
def clear(self, bs, train):
self.loss = None
self.accuracy = None
# init internal state of core RNN
if self.use_lstm:
self.core_lstm.reset_state()
else:
self.h = self.xp.zeros(shape=(bs,self.d_core), dtype=np.float32)
self.h = chainer.Variable(self.h, volatile=not train)
def __call__(self, x, t, train=True):
x = chainer.Variable(self.xp.asarray(x), volatile=not train)
t = chainer.Variable(self.xp.asarray(t), volatile=not train)
bs = x.data.shape[0] # batch size
self.clear(bs, train)
# init mean location
l = np.random.uniform(-1, 1, size=(bs,2)).astype(np.float32)
l = chainer.Variable(self.xp.asarray(l), volatile=not train)
# forward n_steps time
sum_ln_pi = 0
self.forward(x, train, action=False, init_l=l)
for i in range(1, self.n_steps):
action = True if (i == self.n_steps - 1) else False
l, ln_pi, y, b = self.forward(x, train, action)
if train: sum_ln_pi += ln_pi
# loss with softmax cross entropy
self.loss_action = F.softmax_cross_entropy(y, t)
self.loss = self.loss_action
self.accuracy = F.accuracy(y, t)
if train:
# reward
conditions = self.xp.argmax(y.data, axis=1) == t.data
r = self.xp.where(conditions, 1., 0.).astype(np.float32)
# squared error between reward and baseline
self.loss_base = F.mean_squared_error(r, b)
self.loss += self.loss_base
# loss with reinforce rule
mean_ln_pi = sum_ln_pi / (self.n_steps - 1)
self.loss_reinforce = F.sum(-mean_ln_pi * (r-b))/bs
self.loss += self.loss_reinforce
return self.loss
def forward(self, x, train, action, init_l=None):
if init_l is None:
# Location Net @t-1
m = F.tanh(self.fc_hl(self.h))
if train:
eps = (self.xp.random.normal(0, 1, size=m.data.shape)
).astype(np.float32)
l = m.data + np.sqrt(self.var)*eps
# do not backward reinforce loss via l
# log(location policy)
ln_pi = -0.5 * F.sum((l-m)*(l-m), axis=1) / self.var
l = chainer.Variable(l, volatile=not train)
else:
l = m
ln_pi = None
else:
l = init_l
ln_pi = None
# Retina Encoding
x.volatile = 'on' # do not backward
if self.xp == np:
loc = l.data
else:
loc = self.xp.asnumpy(l.data)
rho = crop(x, center=loc, size=self.g_size)
# multi-scale glimpse
for k in range(1, self.n_scales):
s = np.power(2, k)
patch = crop(x, center=loc, size=self.g_size*s)
patch = F.average_pooling_2d(patch, ksize=s)
rho = F.concat((rho, patch), axis=1)
if train: rho.volatile = 'off' # backward up to link emb_x
hg = F.relu(self.emb_x(rho))
# Location Encoding
hl = F.relu(self.emb_l(l))
# Glimpse Net
g = F.relu(self.fc_lg(hl) + self.fc_xg(hg))
# Core Net
if self.use_lstm:
self.h = self.core_lstm(g)
else:
self.h = F.relu(self.core_hh(self.h) + self.core_gh(g))
# Action Net
if action:
y = self.fc_ha(self.h)
else:
y = None
# Baseline
if train and action:
b = F.sigmoid(self.fc_hb(self.h))
b = F.reshape(b, (-1,))
else:
b = None
return l, ln_pi, y, b
def infer(self, x, init_l):
train = False
x = chainer.Variable(self.xp.asarray(x), volatile=not train)
bs = 1 # batch size
self.clear(bs, train)
ys = self.xp.zeros(shape=(self.n_steps,10), dtype=np.float32)
locs = self.xp.zeros(shape=(self.n_steps,2), dtype=np.float32)
locs[0] = np.array(init_l)
# forward
l = init_l.reshape(bs,2).astype(np.float32)
l = chainer.Variable(self.xp.asarray(l), volatile=not train)
l, ln_pi, y, b = self.forward(x, train, action=True, init_l=l)
ys[0] = F.softmax(y).data[0]
locs[0] = l.data[0]
for i in range(1, self.n_steps):
l, ln_pi, y, b = self.forward(x, train, action=True)
locs[i] = l.data[0]
ys[i] = F.softmax(y).data[0]
y = self.xp.argmax(ys[-1])
if self.xp != np:
ys = self.xp.asnumpy(ys)
locs = self.xp.asnumpy(locs)
return y, ys, locs