-
Notifications
You must be signed in to change notification settings - Fork 8
/
optim.py
226 lines (185 loc) · 8.03 KB
/
optim.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
import tensorflow as tf
import tfops as Z
import horovod.tensorflow as hvd
# Optimizers
'''
Polyak averaging op
'''
class Optimizer(object):
def __init__(self):
super(Optimizer, self).__init__()
def polyak(self, params, beta):
#params = tf.trainable_variables()
ema = tf.train.ExponentialMovingAverage(decay=beta, zero_debias=True)
avg_op = tf.group(ema.apply(params))
# Swapping op
updates = []
for i in range(len(params)):
p = params[i]
avg = ema.average(p)
tmp = 0. + avg * 1.
with tf.control_dependencies([tmp]):
update1 = avg.assign(p)
with tf.control_dependencies([update1]):
update2 = p.assign(tmp)
updates += [update1, update2]
swap_op = tf.group(*updates)
return avg_op, swap_op, ema
def adam(self, params, cost_or_grads, alpha=3e-4, hps=None, epsilon=1e-8):
updates = []
if type(cost_or_grads) is not list:
gs = tf.gradients(cost_or_grads, params)
else:
gs = cost_or_grads
beta2 = 1-1./(hps.train_its*hps.polyak_epochs)
# all-reduce
grads = [Z.allreduce_mean(g) for g in gs]
t = tf.Variable(1., 'adam_t')
alpha_t = alpha * tf.sqrt((1. - tf.pow(beta2, t))) / \
(1. - tf.pow(hps.beta1, t))
updates.append(t.assign_add(1))
for w, g in zip(params, grads):
mom2 = tf.Variable(tf.zeros(w.get_shape()), w.name + '_adam_m2')
if hps.beta1 > 0:
mom1 = tf.Variable(tf.zeros(w.get_shape()), w.name + '_adam_m1')
mom1_new = hps.beta1 * mom1 + (1. - hps.beta1) * g
updates.append(mom1.assign(mom1_new))
else:
mom1_new = g
m2_new = beta2 * mom2 + (1. - beta2) * tf.square(g)
delta_t = mom1_new / (tf.sqrt(m2_new) + epsilon)
w_new = hps.weight_decay * w - alpha_t * delta_t
updates.append(mom2.assign(m2_new))
updates.append(w.assign(w_new))
# Polyak averaging
polyak_avg_op, polyak_swap_op, ema = self.polyak(params, beta2)
train_op = tf.group(polyak_avg_op, *updates)
return train_op, polyak_swap_op, ema
'''
Adam optimizer
Version whose learning rate could, in theory, be scaled linearly (like SGD+momentum).
(It doesn't seem to work yet, though.)
'''
def adam2(self, params, cost_or_grads, alpha=3e-4, hps=None, epsilon=1e-8):
updates = []
if type(cost_or_grads) is not list:
gs = tf.gradients(cost_or_grads, params)
else:
gs = cost_or_grads
beta2 = 1-1./(hps.train_its*hps.polyak_epochs)
# all-reduce
grads1 = [Z.allreduce_mean(g) for g in gs]
grads2 = [Z.allreduce_mean(g**2) for g in gs]
t = tf.Variable(1., 'adam_t')
alpha_t = alpha * tf.sqrt((1. - tf.pow(beta2, t))) / \
(1. - tf.pow(hps.beta1, t))
updates.append(t.assign_add(1))
for w, g1, g2 in zip(params, grads1, grads2):
mom2 = tf.Variable(tf.zeros(w.get_shape()), w.name + '_adam_m2')
if hps.beta1 > 0:
mom1 = tf.Variable(tf.zeros(w.get_shape()), w.name + '_adam_m1')
mom1_new = hps.beta1 * mom1 + (1. - hps.beta1) * g1
updates.append(mom1.assign(mom1_new))
else:
mom1_new = g1
m2_new = beta2 * mom2 + (1. - beta2) * g2
delta_t = mom1_new / (tf.sqrt(m2_new) + epsilon)
w_new = hps.weight_decay * w - alpha_t * delta_t
updates.append(mom2.assign(m2_new))
updates.append(w.assign(w_new))
# Polyak averaging
polyak_avg_op, polyak_swap_op, ema = self.polyak(params, beta2)
train_op = tf.group(polyak_avg_op, *updates)
return train_op, polyak_swap_op, ema
'''
Adam optimizer
Version whose learning rate could, in theory, be scaled linearly (like SGD+momentum).
It doesn't seem to work though.
'''
def adam2_old(self, params, cost_or_grads, lr=3e-4, mom1=0.9, mom2=0.999, epsilon=1e-8):
updates = []
if type(cost_or_grads) is not list:
gs = tf.gradients(cost_or_grads, params)
else:
gs = cost_or_grads
# all-reduce
grads1 = [Z.allreduce_mean(g) for g in gs]
grads2 = [Z.allreduce_mean(tf.square(g)) for g in gs]
mom2 = tf.maximum(0., 1. - (hvd.size() * (1 - mom2)))
t = tf.Variable(1., 'adam_t')
lr_t = lr * tf.sqrt((1. - tf.pow(mom2, t))) / (1. - tf.pow(mom1, t))
updates.append(t.assign_add(1))
for p, g1, g2 in zip(params, grads1, grads2):
mg = tf.Variable(tf.zeros(p.get_shape()), p.name + '_adam_mg')
if mom1 > 0:
v = tf.Variable(tf.zeros(p.get_shape()), p.name + '_adam_v')
v_t = mom1 * v + (1. - mom1) * g1
updates.append(v.assign(v_t))
else:
v_t = g1
mg_t = mom2 * mg + (1. - mom2) * g2
delta_t = v_t / (tf.sqrt(mg_t) + epsilon)
p_t = p - lr_t * delta_t
updates.append(mg.assign(mg_t))
updates.append(p.assign(p_t))
return tf.group(*updates)
def adamax(self, params, cost_or_grads, alpha=3e-4, hps=None, epsilon=1e-8):
updates = []
if type(cost_or_grads) is not list:
gs = tf.gradients(cost_or_grads, params)
else:
gs = cost_or_grads
beta2 = 1-1./(hps.train_its*hps.polyak_epochs)
# all-reduce
grads = [Z.allreduce_mean(g) for g in gs]
t = tf.Variable(1., 'adam_t')
alpha_t = alpha * tf.sqrt((1. - tf.pow(beta2, t))) / \
(1. - tf.pow(hps.beta1, t))
updates.append(t.assign_add(1))
for w, g in zip(params, grads):
mom2 = tf.Variable(tf.zeros(w.get_shape()), w.name + '_adam_m2')
if hps.beta1 > 0:
mom1 = tf.Variable(tf.zeros(w.get_shape()), w.name + '_adam_m1')
mom1_new = hps.beta1 * mom1 + (1. - hps.beta1) * g
updates.append(mom1.assign(mom1_new))
else:
mom1_new = g
m2_new = tf.maximum(beta2 * mom2, abs(g))
delta_t = mom1_new / (m2_new + epsilon)
w_new = hps.weight_decay * w - alpha_t * delta_t
updates.append(mom2.assign(m2_new))
updates.append(w.assign(w_new))
# Polyak averaging
polyak_avg_op, polyak_swap_op, ema = self.polyak(params, beta2)
train_op = tf.group(polyak_avg_op, *updates)
return train_op, polyak_swap_op, ema
def adam(self, params, cost_or_grads, alpha=3e-4, hps=None, epsilon=1e-8):
updates = []
if type(cost_or_grads) is not list:
gs = tf.gradients(cost_or_grads, params)
else:
gs = cost_or_grads
beta2 = 1-1./(hps.train_its*hps.polyak_epochs)
# all-reduce
grads = [Z.allreduce_mean(g) for g in gs]
t = tf.Variable(1., 'adam_t')
alpha_t = alpha * tf.sqrt((1. - tf.pow(beta2, t))) / \
(1. - tf.pow(hps.beta1, t))
updates.append(t.assign_add(1))
for w, g in zip(params, grads):
mom2 = tf.Variable(tf.zeros(w.get_shape()), w.name + '_adam_m2')
if hps.beta1 > 0:
mom1 = tf.Variable(tf.zeros(w.get_shape()), w.name + '_adam_m1')
mom1_new = hps.beta1 * mom1 + (1. - hps.beta1) * g
updates.append(mom1.assign(mom1_new))
else:
mom1_new = g
m2_new = beta2 * mom2 + (1. - beta2) * tf.square(g)
delta_t = mom1_new / (tf.sqrt(m2_new) + epsilon)
w_new = hps.weight_decay * w - alpha_t * delta_t
updates.append(mom2.assign(m2_new))
updates.append(w.assign(w_new))
# Polyak averaging
polyak_avg_op, polyak_swap_op, ema = self.polyak(params, beta2)
train_op = tf.group(polyak_avg_op, *updates)
return train_op, polyak_swap_op, ema