|
1
|
|
|
#!/usr/bin/env python |
|
2
|
|
|
# -*- coding: utf-8 -*- |
|
3
|
|
|
|
|
4
|
|
|
import theano |
|
5
|
|
|
import theano.tensor as T |
|
6
|
|
|
|
|
7
|
|
|
from deepy.conf import TrainerConfig |
|
8
|
|
|
from deepy.trainers.base import NeuralTrainer |
|
9
|
|
|
from deepy.trainers.optimize import optimize_updates |
|
10
|
|
|
|
|
11
|
|
|
from logging import getLogger |
|
12
|
|
|
logging = getLogger(__name__) |
|
13
|
|
|
|
|
14
|
|
|
class GeneralNeuralTrainer(NeuralTrainer): |
|
15
|
|
|
""" |
|
16
|
|
|
General neural network trainer. |
|
17
|
|
|
""" |
|
18
|
|
|
def __init__(self, network, config=None, method=None): |
|
19
|
|
|
|
|
20
|
|
|
if method: |
|
21
|
|
|
logging.info("changing optimization method to '%s'" % method) |
|
22
|
|
|
if not config: |
|
23
|
|
|
config = TrainerConfig() |
|
24
|
|
|
elif isinstance(config, dict): |
|
25
|
|
|
config = TrainerConfig(config) |
|
26
|
|
|
config.method = method |
|
27
|
|
|
|
|
28
|
|
|
super(GeneralNeuralTrainer, self).__init__(network, config) |
|
29
|
|
|
|
|
30
|
|
|
logging.info('compiling %s learning function', self.__class__.__name__) |
|
31
|
|
|
|
|
32
|
|
|
self._learning_func = None |
|
33
|
|
|
|
|
34
|
|
|
def learn(self, *variables): |
|
35
|
|
|
if not self._learning_func: |
|
36
|
|
|
self._learning_func = self.learning_function() |
|
37
|
|
|
return self._learning_func(*variables) |
|
38
|
|
|
|
|
39
|
|
|
def _learning_updates(self): |
|
40
|
|
|
""" |
|
41
|
|
|
Return updates in the training. |
|
42
|
|
|
""" |
|
43
|
|
|
params = self.training_params() |
|
44
|
|
|
gradients = self.get_gradients(params) |
|
45
|
|
|
return self.optimization_updates(params, gradients) |
|
46
|
|
|
|
|
47
|
|
|
def training_params(self): |
|
48
|
|
|
""" |
|
49
|
|
|
Get parameters to be optimized. |
|
50
|
|
|
""" |
|
51
|
|
|
params = self.network.parameters |
|
52
|
|
|
# Freeze parameters |
|
53
|
|
|
if self.config.fixed_parameters: |
|
54
|
|
|
logging.info("fixed parameters: %s" % ", ".join(map(str, self.config.fixed_parameters))) |
|
55
|
|
|
params = [p for p in params if p not in self.config.fixed_parameters] |
|
56
|
|
|
return params |
|
57
|
|
|
|
|
58
|
|
|
def get_gradients(self, params): |
|
59
|
|
|
""" |
|
60
|
|
|
Get gradients from given parameters. |
|
61
|
|
|
""" |
|
62
|
|
|
return T.grad(self.cost, params) |
|
63
|
|
|
|
|
64
|
|
|
def optimization_updates(self, params, gradients): |
|
65
|
|
|
""" |
|
66
|
|
|
Return updates from optimization. |
|
67
|
|
|
""" |
|
68
|
|
|
updates, free_parameters = optimize_updates(params, gradients, self.config) |
|
69
|
|
|
self.network.free_parameters.extend(free_parameters) |
|
70
|
|
|
logging.info("Added %d free parameters for optimization" % len(free_parameters)) |
|
71
|
|
|
return updates |
|
72
|
|
|
|
|
73
|
|
|
def learning_function(self): |
|
74
|
|
|
""" |
|
75
|
|
|
Get the learning function. |
|
76
|
|
|
:param func: |
|
77
|
|
|
:return: |
|
78
|
|
|
""" |
|
79
|
|
|
network_updates = list(self.network.updates) + list(self.network.training_updates) |
|
80
|
|
|
learning_updates = list(self._learning_updates()) |
|
81
|
|
|
update_list = network_updates + learning_updates |
|
82
|
|
|
|
|
83
|
|
|
logging.info("network updates: %s" % " ".join(map(str, [x[0] for x in network_updates]))) |
|
84
|
|
|
logging.info("learning updates: %s" % " ".join(map(str, [x[0] for x in learning_updates]))) |
|
85
|
|
|
|
|
86
|
|
|
variables = self.network.input_variables + self.network.target_variables |
|
87
|
|
|
givens = None |
|
88
|
|
|
return theano.function( |
|
89
|
|
|
variables, |
|
90
|
|
|
map(lambda v: theano.Out(v, borrow=True), self.training_variables), |
|
91
|
|
|
updates=update_list, allow_input_downcast=True, |
|
92
|
|
|
mode=self.config.get("theano_mode", None), |
|
93
|
|
|
givens=givens) |
|
94
|
|
|
|
|
95
|
|
|
|
|
96
|
|
|
class SGDTrainer(GeneralNeuralTrainer): |
|
97
|
|
|
""" |
|
98
|
|
|
SGD trainer. |
|
99
|
|
|
""" |
|
100
|
|
|
def __init__(self, network, config=None): |
|
101
|
|
|
super(SGDTrainer, self).__init__(network, config, "SGD") |
|
102
|
|
|
|
|
103
|
|
|
class AdaDeltaTrainer(GeneralNeuralTrainer): |
|
104
|
|
|
""" |
|
105
|
|
|
AdaDelta trainer. |
|
106
|
|
|
""" |
|
107
|
|
|
def __init__(self, network, config=None): |
|
108
|
|
|
super(AdaDeltaTrainer, self).__init__(network, config, "ADADELTA") |
|
109
|
|
|
|
|
110
|
|
|
|
|
111
|
|
|
class AdaGradTrainer(GeneralNeuralTrainer): |
|
112
|
|
|
""" |
|
113
|
|
|
AdaGrad trainer. |
|
114
|
|
|
""" |
|
115
|
|
|
def __init__(self, network, config=None): |
|
116
|
|
|
super(AdaGradTrainer, self).__init__(network, config, "ADAGRAD") |
|
117
|
|
|
|
|
118
|
|
|
class FineTuningAdaGradTrainer(GeneralNeuralTrainer): |
|
119
|
|
|
""" |
|
120
|
|
|
AdaGrad trainer. |
|
121
|
|
|
""" |
|
122
|
|
|
def __init__(self, network, config=None): |
|
123
|
|
|
super(FineTuningAdaGradTrainer, self).__init__(network, config, "FINETUNING_ADAGRAD") |
|
124
|
|
|
|
|
125
|
|
|
class AdamTrainer(GeneralNeuralTrainer): |
|
126
|
|
|
""" |
|
127
|
|
|
AdaGrad trainer. |
|
128
|
|
|
""" |
|
129
|
|
|
def __init__(self, network, config=None): |
|
130
|
|
|
super(AdamTrainer, self).__init__(network, config, "ADAM") |
|
131
|
|
|
|
|
132
|
|
|
class RmspropTrainer(GeneralNeuralTrainer): |
|
133
|
|
|
""" |
|
134
|
|
|
RmsProp trainer. |
|
135
|
|
|
""" |
|
136
|
|
|
def __init__(self, network, config=None): |
|
137
|
|
|
super(RmspropTrainer, self).__init__(network, config, "RMSPROP") |
|
138
|
|
|
|
|
139
|
|
|
class MomentumTrainer(GeneralNeuralTrainer): |
|
140
|
|
|
""" |
|
141
|
|
|
Momentum trainer. |
|
142
|
|
|
""" |
|
143
|
|
|
def __init__(self, network, config=None): |
|
144
|
|
|
super(MomentumTrainer, self).__init__(network, config, "MOMENTUM") |
|
145
|
|
|
|
|
146
|
|
|
class FakeTrainer(GeneralNeuralTrainer): |
|
147
|
|
|
""" |
|
148
|
|
|
Fake Trainer does nothing. |
|
149
|
|
|
""" |
|
150
|
|
|
|
|
151
|
|
|
def _learning_updates(self): |
|
152
|
|
|
return [] |
zomux /
deepy
