-
Notifications
You must be signed in to change notification settings - Fork 1
/
optimizers.py
204 lines (166 loc) · 5.99 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
"""Defines gradiant-based optimizers."""
import numpy as np
from .operation import backprop as _backprop
class Optimizer(object):
"""Base class of all Optimizers.
Subclasses must implement abstract method `apply_gradients`.
"""
def __init__(self, **params):
"""Constructor.
Args:
params: a dict mapping from parameter names to parameters.
"""
self._params = params
self._params_str = ', '.join(['%s=%s' % (k, v) for k, v in params.items()])
def __repr__(self):
"""Displays the initializer name and list of parameter name and value pairs.
"""
if self._params_str:
return '<%s:%s>' % (type(self).__name__, self._params_str)
else:
return '<%s>' % type(self).__name__
def compute_gradients(self, tensor, variables):
"""Compute the gradients w.r.t. the variables.
Args:
tensor (Tensor): the Tensor from which gradients will be backpropped (
i.e. the loss Tensor).
variables (List[Variable]): list of `Variable`s.
Returns:
grads_and_vars (List[Tuple[Tensor, Tensor]]): list of (gradient,
creator_id) tuples.
"""
variables = [(v.weight, v.handle) for v in variables if v.trainable]
var_weights = list(list(zip(*variables))[0])
var_handles = list(list(zip(*variables))[1])
gradients = _backprop([tensor], var_weights)
grads_and_vars = list(zip(gradients, var_handles))
return grads_and_vars
class GradientDescentOptimizer(Optimizer):
"""The Vanilla Gradient Descent Optimizer."""
def apply_gradients(self, grads_and_vars, reset_runtime=True):
"""Apply the computed gradient w.r.t. trainable variables.
Args:
grads_and_vars (List[Tuple[Tensor, Tensor]]): list of (gradient,
creator_id) tuples.
reset_runtime (bool): (Optional) whether to reset runtime after variables
are updated. Defaults to True.
"""
runtime = grads_and_vars[0][0].op.graph.runtime
for grad, var in grads_and_vars:
var_id = int(var.eval())
var_value = runtime.get_variable_value(var_id).astype("float32")
grad_value = grad.eval().astype("float32")
runtime.set_variable_value(
var_id, var_value - self._params["alpha"] * grad_value,
)
if reset_runtime:
runtime.reset()
class AdamOptimizer(Optimizer):
"""Adam optimizer"""
def __init__(self, **params):
"""Constructor.
Args:
params: a dict mapping from parameter names to parameters.
"""
self._params = params
self._params_str = ', '.join([
'%s=%s' % (k, v)
for k, v in params.items()
if k in ('alpha', 'beta1', 'beta2', 'epsilon')
])
self._t = 0
self._m = dict()
self._v = dict()
def apply_gradients(self, grads_and_vars, reset_runtime=True):
"""Apply the computed gradient w.r.t. trainable variables.
Args:
grads_and_vars (List[Tuple[Tensor, Tensor]]): list of (gradient,
creator_id) tuples.
reset_runtime (bool): (Optional) whether to reset runtime after variables
are updated. Defaults to True.
"""
alpha, beta1, beta2, epsilon = (
np.asarray(
self._params['alpha'],
"float32",
), np.asarray(self._params['beta1'], "float32"),
np.asarray(
self._params['beta2'],
"float32",
), np.asarray(self._params['epsilon'], "float32"),
)
t = self._t + 1
m = self._m
v = self._v
alpha_t = alpha * np.sqrt(1 - np.power(beta2, t)) / (1 - np.power(beta1, t))
alpha_t = alpha_t.astype('float32')
runtime = grads_and_vars[0][0].op.graph.runtime
for grad, var in grads_and_vars:
var_id = int(var.eval())
var_value = runtime.get_variable_value(var_id).astype("float32")
grad_value = grad.eval().astype("float32")
var_shape = var_value.shape
m[var_id] = beta1 * m.get(
var_id, np.zeros(var_shape, dtype="float32"),
) + (1 - beta1) * grad_value
v[var_id] = beta2 * v.get(
var_id, np.zeros(var_shape, dtype="float32"),
) + (1 - beta2) * grad_value * grad_value
runtime.set_variable_value(
var_id,
var_value - alpha_t * m[var_id] / (np.sqrt(v[var_id]) + epsilon),
)
self._m = m
self._v = v
self._t = t
if reset_runtime:
runtime.reset()
class RMSPropOptimizer(Optimizer):
"""RMSProp Optimizer"""
def __init__(self, **params):
"""Constructor.
Args:
params: a dict mapping from parameter names to parameters.
"""
self._params = params
self._params_str = ', '.join([
'%s=%s' % (k, v)
for k, v in params.items()
if k in ('alpha', 'rho', 'momentum', 'epsilon')
])
self._mean_square = dict()
self._moment = dict()
def apply_gradients(self, grads_and_vars, reset_runtime=True):
"""Apply the computed gradient w.r.t. trainable variables.
Args:
grads_and_vars (List[Tuple[Tensor, Tensor]]): list of (gradient,
creator_id) tuples.
reset_runtime (bool): (Optional) whether to reset runtime after variables
are updated. Defaults to True.
"""
alpha, rho, momentum, epsilon = (
self._params['alpha'], self._params['rho'], self._params['momentum'],
self._params['epsilon'],
)
mean_square = self._mean_square
moment = self._moment
runtime = grads_and_vars[0][0].op.graph.runtime
for grad, var in grads_and_vars:
var_id = var.eval().item().id
var_shape = var.eval().item().shape
var_value = runtime.get_variable_value(var_id)
grad_value = grad.eval().astype("float32")
mean_square[var_id] = (
rho * mean_square.get(var_id, np.zeros(var_shape)) +
(1 - rho) * grad_value * grad_value
)
moment[var_id] = momentum * moment.get(
var_id, np.zeros(var_shape),
) + alpha * grad_value / (
np.sqrt(mean_square[var_id]) + epsilon
)
runtime.set_variable_value(var_id, var_value - moment[var_id])
self._mean_square = mean_square
self._moment = moment
if reset_runtime:
runtime.reset()