-
Notifications
You must be signed in to change notification settings - Fork 0
/
optimizers.py
253 lines (202 loc) · 8.17 KB
/
optimizers.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
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
'''
Optimizers
'''
import numpy
from collections import OrderedDict
import theano
import theano.tensor as tensor
from theano.sandbox.rng_mrg import MRG_RandomStreams as RandomStreams
from util import *
from theano_util import *
# Calling convention:
# f_update = name(hyperp, tparams, grads, inputs (list), cost)
# with profile as an optional argument
def adam(lr, tparams, grads, inp, cost, beta1=0.9, beta2=0.999, e=1e-8, optimizer_params={}, profile=False):
PREFIX='adam_'
updates = []
optimizer_tparams = {}
# Avoid underflow of e with float16
if (floatX == "float16") and (e > 0.0):
e = max(e, 1e-6)
t_prev_name = PREFIX + 't_prev'
if t_prev_name in optimizer_params:
t_prev_init = optimizer_params[t_prev_name]
else:
t_prev_init = 0.
t_prev = theano.shared(numpy_floatX(t_prev_init), t_prev_name)
optimizer_tparams[t_prev_name] = t_prev
t = t_prev + 1.
lr_t = lr * tensor.sqrt(1. - beta2**t) / (1. - beta1**t)
for p, g in zip(tparams.values(), grads):
# Create/Load variable for first moment
m_name = PREFIX + p.name + '_mean'
if m_name in optimizer_params:
m_init = optimizer_params[m_name]
else:
m_init = p.get_value() * 0.
m = theano.shared(m_init, m_name)
optimizer_tparams[m_name] = m
# Create/Load variable for second moment
v_name = PREFIX + p.name + '_variance'
if v_name in optimizer_params:
v_init = optimizer_params[v_name]
else:
v_init = p.get_value() * 0.
v = theano.shared(v_init, v_name)
optimizer_tparams[v_name] = v
# Define updates on shared vars
m_t = beta1 * m + (1. - beta1) * g
v_t = beta2 * v + (1. - beta2) * g**2
step = lr_t * m_t / (tensor.sqrt(v_t) + e)
p_t = p - step
updates.append((m, m_t))
updates.append((v, v_t))
updates.append((p, p_t))
updates.append((t_prev, t))
f_update = theano.function([lr]+inp, cost, updates=updates,
on_unused_input='ignore', profile=profile)
return f_update, optimizer_tparams
def adam_disc(lr, tparams, grads, cost, beta1=0.9, beta2=0.999, e=1e-8, optimizer_params={}, profile=False):
PREFIX='adam_'
updates = []
optimizer_tparams = {}
# Avoid underflow of e with float16
if (floatX == "float16") and (e > 0.0):
e = max(e, 1e-6)
t_prev_name = PREFIX + 't_prev'
if t_prev_name in optimizer_params:
t_prev_init = optimizer_params[t_prev_name]
else:
t_prev_init = 0.
t_prev = theano.shared(numpy_floatX(t_prev_init), t_prev_name)
optimizer_tparams[t_prev_name] = t_prev
t = t_prev + 1.
lr_t = lr * tensor.sqrt(1. - beta2**t) / (1. - beta1**t)
for p, g in zip(tparams.values(), grads):
# Create/Load variable for first moment
m_name = PREFIX + p.name + '_mean'
if m_name in optimizer_params:
m_init = optimizer_params[m_name]
else:
m_init = p.get_value() * 0.
m = theano.shared(m_init, m_name)
optimizer_tparams[m_name] = m
# Create/Load variable for second moment
v_name = PREFIX + p.name + '_variance'
if v_name in optimizer_params:
v_init = optimizer_params[v_name]
else:
v_init = p.get_value() * 0.
v = theano.shared(v_init, v_name)
optimizer_tparams[v_name] = v
# Define updates on shared vars
m_t = beta1 * m + (1. - beta1) * g
v_t = beta2 * v + (1. - beta2) * g**2
step = lr_t * m_t / (tensor.sqrt(v_t) + e)
p_t = p - step
updates.append((m, m_t))
updates.append((v, v_t))
updates.append((p, p_t))
updates.append((t_prev, t))
# f_update = theano.function([lr]+inp, cost, updates=updates,
# on_unused_input='ignore', profile=profile)
return updates #f_update, optimizer_tparams
def adadelta(lr, tparams, grads, inp, cost, optimizer_params={}, profile=False):
PREFIX = 'adadelta_'
updates = []
optimizer_tparams = {}
for p, g in zip(tparams.values(), grads):
zg_name = PREFIX + p.name + '_zg'
if zg_name in optimizer_params:
zg_init = optimizer_params[zg_name]
else:
zg_init = p.get_value() * 0.
zg = theano.shared(zg_init, zg_name)
optimizer_tparams[zg_name] = zg
ru2_name = PREFIX + p.name + '_ru2'
if ru2_name in optimizer_params:
ru2_init = optimizer_params[ru2_name]
else:
ru2_init = p.get_value() * 0.
ru2 = theano.shared(ru2_init, ru2_name)
optimizer_tparams[ru2_name] = ru2
rg2_name = PREFIX + p.name + '_rg2'
if rg2_name in optimizer_params:
rg2_init = optimizer_params[rg2_name]
else:
rg2_init = p.get_value() * 0.
rg2 = theano.shared(rg2_init, rg2_name)
optimizer_tparams[rg2_name] = rg2
ud = -tensor.sqrt(ru2 + 1e-6) / tensor.sqrt(rg2 + 1e-6) * zg
updates.append((zg, g))
updates.append((rg2, 0.95 * rg2 + 0.05 * (g ** 2)))
updates.append((ru2, 0.95 * ru2 + 0.05 * (ud ** 2)))
updates.append((p, p + ud))
f_update = theano.function([lr]+inp, cost, updates=updates,
on_unused_input='ignore', profile=profile)
return f_update, optimizer_tparams
def rmsprop(lr, tparams, grads, inp, cost, optimizer_params={}, profile=False):
PREFIX = 'rmsprop_'
updates = []
optimizer_tparams = {}
for p, g in zip(tparams.values(), grads):
zg_name = PREFIX + p.name + '_zg'
if zg_name in optimizer_params:
zg_init = optimizer_params[zg_name]
else:
zg_init = p.get_value() * 0.
zg = theano.shared(zg_init, zg_name)
optimizer_tparams[zg_name] = zg
rg_name = PREFIX + p.name + '_rg'
if rg_name in optimizer_params:
rg_init = optimizer_params[rg_name]
else:
rg_init = p.get_value() * 0.
rg = theano.shared(rg_init, rg_name)
optimizer_tparams[rg_name] = rg
rg2_name = PREFIX + p.name + '_rg2'
if rg2_name in optimizer_params:
rg2_init = optimizer_params[rg2_name]
else:
rg2_init = p.get_value() * 0.
rg2 = theano.shared(rg2_init, rg2_name)
optimizer_tparams[rg2_name] = rg2
ud_name = PREFIX + p.name + '_ud'
if ud_name in optimizer_params:
ud_init = optimizer_params[ud_name]
else:
ud_init = p.get_value() * 0.
ud = theano.shared(ud_init, ud_name)
optimizer_tparams[ud_name] = ud
updates.append((zg, g))
updates.append((rg, 0.95 * rg + 0.05 * g))
updates.append((rg2, 0.95 * rg2 + 0.05 * (g ** 2)))
udn = 0.9 * ud - 1e-4 * zg / tensor.sqrt(rg2 - rg ** 2 + 1e-4)
updates.append((ud, udn))
updates.append((p, p + udn))
f_update = theano.function([lr]+inp, cost, updates=updates,
on_unused_input='ignore', profile=profile)
return f_update, optimizer_tparams
def sgd(lr, tparams, grads, inp, cost, optimizer_params=None, profile=False):
updates = [(p, p - lr * g) for p, g in zip(tparams.values(), grads)]
f_update = theano.function([lr]+inp, cost, updates=updates, profile=profile)
return f_update, {}
def sgdmomentum(lr, tparams, grads, inp, cost, momentum=0.9, optimizer_params={}, profile=False):
assert momentum >= 0 and momentum < 1
PREFIX = 'sgdmomentum_'
updates = []
optimizer_tparams = {}
for p, g in zip(tparams.values(), grads):
prev_name = PREFIX + p.name + '_prev'
if prev_name in optimizer_params:
prev_init = optimizer_params[prev_name]
else:
prev_init = p.get_value() * 0.
prev = theano.shared(prev_init, prev_name)
optimizer_tparams[prev_name] = prev
step = momentum * prev - lr * g
updates.append((prev, step))
updates.append((p, p + step))
f_update = theano.function([lr]+inp, cost, updates=updates,
on_unused_input='ignore', profile=profile)
return f_update, optimizer_tparams