-
Notifications
You must be signed in to change notification settings - Fork 11
/
Copy pathadasopt.py
169 lines (152 loc) · 7.64 KB
/
adasopt.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
# MIT License
#
# Copyright (c) 2020 YanaiEliyahu
#
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
"""Implementation of Adas."""
from tensorflow.keras.optimizers import Optimizer
from tensorflow.keras import backend as K
from tensorflow.python.ops import math_ops
from tensorflow.python.ops import control_flow_ops
from tensorflow.python.ops import state_ops
from tensorflow.python.training import moving_averages
import tensorflow as tf
class AdasOptimizer(Optimizer):
"""
Introduction:
For the mathematical part see https://github.com/YanaiEliyahu/AdasOptimizer,
the `Theory` section contains the major innovation,
and then `How ADAS works` contains more low level details that are still somewhat related to the theory.
Compatibility:
Tested under tensorflow versions 1.5.4 and 2.3.1.
Arguments:
lr: float > 0. Initial learning rate that is per feature/input (e.g. dense layer with N inputs and M outputs, will have N learning rates).
lr2: float >= 0. lr's Initial learning rate. (just ~1-2 per layer, additonal one because of bias)
lr3: float >= 0. lr2's fixed learning rate. (global)
beta_1: 0 < float < 1. Preferably close to 1. Second moments decay factor to update lr and lr2 weights.
beta_2: 0 < float < 1. Preferably close to 1. 1/(1 - beta_2) steps back in time that `lr`s will be optimized for, larger dataset might require more nines.
beta_3: 0 < float < 1. Preferably close to 1. Same as beta_2, but for `lr2`s.
epsilon: float >= 0. Fuzz factor. If `None`, defaults to `K.epsilon()`.
"""
def __init__(self,
lr = 0.001, lr2 = .005, lr3 = .0005,
beta_1 = 0.999, beta_2 = 0.999, beta_3 = 0.9999,
epsilon = None, **kwargs):
super(AdasOptimizer, self).__init__('Adas',**kwargs)
# TODO: use self._set_hyper and self._get_hyper
self._lr = lr
self._lr2 = lr2
self._lr3 = lr3
self._beta_1 = beta_1
self._beta_2 = beta_2
self._beta_3 = beta_3
if epsilon is None:
epsilon = K.epsilon()
self._epsilon = epsilon
self._tf1 = tf.__version__.split(".")[0] == '1'
if not self._tf1:
self._varn = None
self._is_create_slots = None
self._curr_var = None
def _assign(self,x,y):
if self._tf1:
result = K.update(x,y)
else:
result = state_ops.assign(x,y,use_locking=self._use_locking)
self._updates.append(result)
return result
def _add(self,x,y):
if self._tf1:
result = K.update_add(x,y)
else:
result = state_ops.assign_add(x,y,use_locking=self._use_locking)
self._updates.append(result)
return result
def _moving_average(self,var,value,momentum):
if self._tf1:
return self._assign(var,var * momentum + value * (1 - momentum))
result = K.moving_average_update(var,value,momentum)
self._updates.append(result)
return result
# TODO: fix variables' names being too convoluted in _derivatives_normalizer and _get_updates_universal_impl
def _derivatives_normalizer(self,derivative,beta):
if self._tf1:
self._iterations = self._make_variable()
self._add(self._iterations,1)
t = K.cast(self._iterations if self._tf1 else self.iterations, K.floatx()) + 1
lr_t = K.sqrt(1. - K.pow(self._beta_1, t))
m = self._make_variable(0,K.int_shape(derivative),K.dtype(derivative))
old_moments = self._make_variable(0,K.int_shape(derivative),K.dtype(derivative))
old_moments2 = self._make_variable(0,K.int_shape(derivative),K.dtype(derivative))
res = self._assign(old_moments2,old_moments)
v_t = self._moving_average(m,K.square(derivative),self._beta_1)
with tf.control_dependencies([v_t]):
np_t = derivative * lr_t / (K.sqrt(v_t) + self._epsilon)
with tf.control_dependencies([np_t]):
m_t = self._moving_average(old_moments,np_t,beta)
return (res,np_t)
def _make_variable(self,value = 0,shape = (),dtype = K.floatx()):
if self._tf1:
return K.variable(K.constant(value,shape=shape, dtype=dtype))
self._varn += 1
name = 'unnamed_variable' + str(self._varn)
if self._is_create_slots:
return self.add_slot(self._curr_var,name,initializer = K.constant(value,shape=shape, dtype=dtype))
else:
return self.get_slot(self._curr_var,name)
def _get_updates_universal_impl(self, grad, param):
self._updates = []
lr = self._make_variable(value = self._lr,shape=K.int_shape(param)[:-1], dtype=K.dtype(param))
moment, deriv = self._derivatives_normalizer(grad,self._beta_3)
param_t = self._add(param, - K.expand_dims(lr,len(K.int_shape(param)) - 1) * deriv)
with tf.control_dependencies([moment]):
lr_deriv = math_ops.reduce_sum(moment * grad,len(K.int_shape(param)) - 1)
master_lr = self._make_variable(self._lr2)
m2,d2 = self._derivatives_normalizer(lr_deriv,self._beta_2)
lr_t = self._add(lr,master_lr * lr * d2)
with tf.control_dependencies([m2]):
master_lr_deriv2 = math_ops.reduce_sum(m2 * lr_deriv)
m3,d3 = self._derivatives_normalizer(master_lr_deriv2,0.)
with tf.control_dependencies([m3]):
self._add(master_lr,self._lr3 * master_lr * d3)
return self._updates
def _get_updates_universal(self, param, grad = None, is_create_slots = False):
self._curr_var = param
self._is_create_slots = is_create_slots
self._varn = 0
return self._get_updates_universal_impl(grad if grad is not None else K.constant(0,shape=param.shape,dtype=K.dtype(param)),param)
def get_updates(self, loss, params):
return sum([self._get_updates_universal_impl(grad,var) for (grad,var) in zip(self.get_gradients(loss, params),params)],[])
def _create_slots(self, var_list):
for var in var_list:
self._get_updates_universal(var,is_create_slots = True)
def _resource_apply_dense(self, grad, var):
return control_flow_ops.group(*self._get_updates_universal(var,grad))
def get_config(self):
config = {
'lr': float(self._lr),
'lr2': float(self._lr2),
'lr3': float(self._lr3),
'beta_1': float(K.get_value(self._beta_1)),
'beta_2': float(K.get_value(self._beta_2)),
'beta_3': float(K.get_value(self._beta_3)),
'epsilon': self._epsilon
}
base_config = super(AdasOptimizer, self).get_config()
return dict(list(base_config.items()) + list(config.items()))