-
Notifications
You must be signed in to change notification settings - Fork 0
/
custom_adam.py
199 lines (180 loc) · 8.45 KB
/
custom_adam.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
# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Adam for TensorFlow."""
from __future__ import absolute_import, division, print_function
from tensorflow import constant
from tensorflow.python.framework import ops
from tensorflow.python.keras import backend_config
from tensorflow.python.keras.optimizer_v2 import optimizer_v2
from tensorflow.python.ops import array_ops, control_flow_ops, math_ops, state_ops
from tensorflow.python.training import training_ops
class CustomAdam(optimizer_v2.OptimizerV2):
"""Optimizer that implements the Adam algorithm.
Adam optimization is a stochastic gradient descent method that is based on
adaptive estimation of first-order and second-order moments.
According to the paper
[Adam: A Method for Stochastic Optimization. Kingma et al.,
2014](http://arxiv.org/abs/1412.6980), the method is "*computationally
efficient, has little memory requirement, invariant to diagonal rescaling of
gradients, and is well suited for problems that are large in terms of
data/parameters*".
For AMSGrad see [On The Convergence Of Adam And Beyond.
Reddi et al., 5-8](https://openreview.net/pdf?id=ryQu7f-RZ).
"""
_HAS_AGGREGATE_GRAD = True
def __init__(
self,
learning_rate=1e-3,
learning_rate_deconv=1e-4,
beta_1=0.9,
beta_2=0.999,
epsilon=1e-7,
amsgrad=False,
name="Adam",
**kwargs
):
super(CustomAdam, self).__init__(name, **kwargs)
self._set_hyper("learning_rate", kwargs.get("lr", learning_rate))
self._set_hyper("learning_rate_deconv", learning_rate_deconv)
self._set_hyper("decay", self._initial_decay)
self._set_hyper("beta_1", beta_1)
self._set_hyper("beta_2", beta_2)
self.epsilon = epsilon or backend_config.epsilon()
self.amsgrad = amsgrad
def _create_slots(self, var_list):
# Create slots for the first and second moments.
# Separate for-loops to respect the ordering of slot variables from v1.
for var in var_list:
self.add_slot(var, "m")
for var in var_list:
self.add_slot(var, "v")
if self.amsgrad:
for var in var_list:
self.add_slot(var, "vhat")
def _prepare_local(self, var_device, var_dtype, apply_state):
super(CustomAdam, self)._prepare_local(var_device, var_dtype, apply_state)
local_step = math_ops.cast(self.iterations + 1, var_dtype)
beta_1_t = array_ops.identity(self._get_hyper("beta_1", var_dtype))
beta_2_t = array_ops.identity(self._get_hyper("beta_2", var_dtype))
beta_1_power = math_ops.pow(beta_1_t, local_step)
beta_2_power = math_ops.pow(beta_2_t, local_step)
lr = apply_state[(var_device, var_dtype)]["lr_t"] * (math_ops.sqrt(1 - beta_2_power) / (1 - beta_1_power))
apply_state[(var_device, var_dtype)].update(
dict(
lr=lr,
epsilon=ops.convert_to_tensor_v2(self.epsilon, var_dtype),
beta_1_t=beta_1_t,
beta_1_power=beta_1_power,
one_minus_beta_1_t=1 - beta_1_t,
beta_2_t=beta_2_t,
beta_2_power=beta_2_power,
one_minus_beta_2_t=1 - beta_2_t,
)
)
def set_weights(self, weights):
params = self.weights
# If the weights are generated by Keras V1 optimizer, it includes vhats
# even without amsgrad, i.e, V1 optimizer has 3x + 1 variables, while V2
# optimizer has 2x + 1 variables. Filter vhats out for compatibility.
num_vars = int((len(params) - 1) / 2)
if len(weights) == 3 * num_vars + 1:
weights = weights[: len(params)]
super(CustomAdam, self).set_weights(weights)
def _resource_apply_dense(self, grad, var, apply_state=None):
var_device, var_dtype = var.device, var.dtype.base_dtype
coefficients = (apply_state or {}).get((var_device, var_dtype)) or self._fallback_apply_state(
var_device, var_dtype
)
m = self.get_slot(var, "m")
v = self.get_slot(var, "v")
lr = coefficients["lr_t"]
if str(var.name).find("transpose") != -1:
lr = constant(self._serialize_hyperparameter("learning_rate_deconv"))
if not self.amsgrad:
return training_ops.resource_apply_adam(
var.handle,
m.handle,
v.handle,
coefficients["beta_1_power"],
coefficients["beta_2_power"],
lr,
coefficients["beta_1_t"],
coefficients["beta_2_t"],
coefficients["epsilon"],
grad,
use_locking=self._use_locking,
)
else:
vhat = self.get_slot(var, "vhat")
return training_ops.resource_apply_adam_with_amsgrad(
var.handle,
m.handle,
v.handle,
vhat.handle,
coefficients["beta_1_power"],
coefficients["beta_2_power"],
lr,
coefficients["beta_1_t"],
coefficients["beta_2_t"],
coefficients["epsilon"],
grad,
use_locking=self._use_locking,
)
def _resource_apply_sparse(self, grad, var, indices, apply_state=None):
var_device, var_dtype = var.device, var.dtype.base_dtype
coefficients = (apply_state or {}).get((var_device, var_dtype)) or self._fallback_apply_state(
var_device, var_dtype
)
# m_t = beta1 * m + (1 - beta1) * g_t
m = self.get_slot(var, "m")
m_scaled_g_values = grad * coefficients["one_minus_beta_1_t"]
m_t = state_ops.assign(m, m * coefficients["beta_1_t"], use_locking=self._use_locking)
with ops.control_dependencies([m_t]):
m_t = self._resource_scatter_add(m, indices, m_scaled_g_values)
# v_t = beta2 * v + (1 - beta2) * (g_t * g_t)
v = self.get_slot(var, "v")
v_scaled_g_values = (grad * grad) * coefficients["one_minus_beta_2_t"]
v_t = state_ops.assign(v, v * coefficients["beta_2_t"], use_locking=self._use_locking)
with ops.control_dependencies([v_t]):
v_t = self._resource_scatter_add(v, indices, v_scaled_g_values)
if not self.amsgrad:
v_sqrt = math_ops.sqrt(v_t)
var_update = state_ops.assign_sub(
var, coefficients["lr"] * m_t / (v_sqrt + coefficients["epsilon"]), use_locking=self._use_locking
)
return control_flow_ops.group(*[var_update, m_t, v_t])
else:
v_hat = self.get_slot(var, "vhat")
v_hat_t = math_ops.maximum(v_hat, v_t)
with ops.control_dependencies([v_hat_t]):
v_hat_t = state_ops.assign(v_hat, v_hat_t, use_locking=self._use_locking)
v_hat_sqrt = math_ops.sqrt(v_hat_t)
var_update = state_ops.assign_sub(
var, coefficients["lr"] * m_t / (v_hat_sqrt + coefficients["epsilon"]), use_locking=self._use_locking
)
return control_flow_ops.group(*[var_update, m_t, v_t, v_hat_t])
def get_config(self):
config = super(CustomAdam, self).get_config()
config.update(
{
"learning_rate": self._serialize_hyperparameter("learning_rate"),
"decay": self._serialize_hyperparameter("decay"),
"beta_1": self._serialize_hyperparameter("beta_1"),
"beta_2": self._serialize_hyperparameter("beta_2"),
"epsilon": self.epsilon,
"amsgrad": self.amsgrad,
}
)
return config