|
1
|
|
|
#!/usr/bin/env python |
|
2
|
|
|
# -*- coding: utf-8 -*- |
|
3
|
|
|
|
|
4
|
|
|
import numpy as np |
|
5
|
|
|
|
|
6
|
|
|
from controllers import TrainingController |
|
7
|
|
|
from deepy.core.env import FLOATX |
|
8
|
|
|
from deepy.core import graph |
|
9
|
|
|
|
|
10
|
|
|
import logging as loggers |
|
11
|
|
|
logging = loggers.getLogger(__name__) |
|
12
|
|
|
|
|
13
|
|
|
class LearningRateAnnealer(TrainingController): |
|
14
|
|
|
""" |
|
15
|
|
|
Learning rate annealer. |
|
16
|
|
|
""" |
|
17
|
|
|
|
|
18
|
|
|
def __init__(self, patience=3, anneal_times=4): |
|
19
|
|
|
""" |
|
20
|
|
|
:type trainer: deepy.trainers.base.NeuralTrainer |
|
21
|
|
|
""" |
|
22
|
|
|
self._iter = 0 |
|
23
|
|
|
self._annealed_iter = 0 |
|
24
|
|
|
self._patience = patience |
|
25
|
|
|
self._anneal_times = anneal_times |
|
26
|
|
|
self._annealed_times = 0 |
|
27
|
|
|
self._learning_rate = 0 |
|
28
|
|
|
if type(self._learning_rate) == float: |
|
29
|
|
|
raise Exception("use shared_scalar to wrap the value in the config.") |
|
30
|
|
|
|
|
31
|
|
|
def bind(self, trainer): |
|
32
|
|
|
super(LearningRateAnnealer, self).bind(trainer) |
|
33
|
|
|
self._learning_rate = self._trainer.config.learning_rate |
|
34
|
|
|
self._iter = 0 |
|
35
|
|
|
self._annealed_iter = 0 |
|
36
|
|
|
|
|
37
|
|
|
def invoke(self): |
|
38
|
|
|
""" |
|
39
|
|
|
Run it, return whether to end training. |
|
40
|
|
|
""" |
|
41
|
|
|
self._iter += 1 |
|
42
|
|
|
if self._iter - max(self._trainer.best_iter, self._annealed_iter) >= self._patience: |
|
43
|
|
|
if self._annealed_times >= self._anneal_times: |
|
44
|
|
|
logging.info("ending") |
|
45
|
|
|
self._trainer.exit() |
|
46
|
|
|
else: |
|
47
|
|
|
self._trainer.set_params(*self._trainer.best_params) |
|
48
|
|
|
self._learning_rate.set_value(self._learning_rate.get_value() * 0.5) |
|
49
|
|
|
self._annealed_times += 1 |
|
50
|
|
|
self._annealed_iter = self._iter |
|
51
|
|
|
logging.info("annealed learning rate to %f" % self._learning_rate.get_value()) |
|
52
|
|
|
|
|
53
|
|
|
@staticmethod |
|
54
|
|
|
def learning_rate(value=0.01): |
|
55
|
|
|
""" |
|
56
|
|
|
Wrap learning rate. |
|
57
|
|
|
""" |
|
58
|
|
|
return graph.shared(value, name="learning_rate") |
|
59
|
|
|
|
|
60
|
|
|
|
|
61
|
|
|
class ScheduledLearningRateAnnealer(TrainingController): |
|
62
|
|
|
""" |
|
63
|
|
|
Anneal learning rate according to pre-scripted schedule. |
|
64
|
|
|
""" |
|
65
|
|
|
|
|
66
|
|
|
def __init__(self, start_halving_at=5, end_at=10, halving_interval=1, rollback=False): |
|
67
|
|
|
logging.info("iteration to start halving learning rate: %d" % start_halving_at) |
|
68
|
|
|
self.epoch_start_halving = start_halving_at |
|
69
|
|
|
self.end_at = end_at |
|
70
|
|
|
self._halving_interval = halving_interval |
|
71
|
|
|
self._rollback = rollback |
|
72
|
|
|
self._last_halving_epoch = 0 |
|
73
|
|
|
self._learning_rate = None |
|
74
|
|
|
|
|
75
|
|
|
def bind(self, trainer): |
|
76
|
|
|
super(ScheduledLearningRateAnnealer, self).bind(trainer) |
|
77
|
|
|
self._learning_rate = self._trainer.config.learning_rate |
|
78
|
|
|
self._last_halving_epoch = 0 |
|
79
|
|
|
|
|
80
|
|
|
def invoke(self): |
|
81
|
|
|
epoch = self._trainer.epoch() |
|
82
|
|
|
if epoch >= self.epoch_start_halving and epoch >= self._last_halving_epoch + self._halving_interval: |
|
83
|
|
|
if self._rollback: |
|
84
|
|
|
self._trainer.set_params(*self._trainer.best_params) |
|
85
|
|
|
self._learning_rate.set_value(self._learning_rate.get_value() * 0.5) |
|
86
|
|
|
logging.info("halving learning rate to %f" % self._learning_rate.get_value()) |
|
87
|
|
|
self._trainer.network.train_logger.record("set learning rate to %f" % self._learning_rate.get_value()) |
|
88
|
|
|
self._last_halving_epoch = epoch |
|
89
|
|
|
if epoch >= self.end_at: |
|
90
|
|
|
logging.info("ending") |
|
91
|
|
|
self._trainer.exit() |
|
92
|
|
|
|
|
93
|
|
|
|
|
94
|
|
|
class ExponentialLearningRateAnnealer(TrainingController): |
|
95
|
|
|
""" |
|
96
|
|
|
Exponentially decay learning rate after each update. |
|
97
|
|
|
""" |
|
98
|
|
|
|
|
99
|
|
|
def __init__(self, decay_factor=1.000004, min_lr=.000001, debug=False): |
|
100
|
|
|
logging.info("exponentially decay learning rate with decay factor = %f" % decay_factor) |
|
101
|
|
|
self.decay_factor = np.array(decay_factor, dtype=FLOATX) |
|
102
|
|
|
self.min_lr = np.array(min_lr, dtype=FLOATX) |
|
103
|
|
|
self.debug = debug |
|
104
|
|
|
self._learning_rate = self._trainer.config.learning_rate |
|
105
|
|
|
if type(self._learning_rate) == float: |
|
106
|
|
|
raise Exception("use shared_scalar to wrap the value in the config.") |
|
107
|
|
|
self._trainer.network.training_callbacks.append(self.update_callback) |
|
108
|
|
|
|
|
109
|
|
|
def update_callback(self): |
|
110
|
|
|
if self._learning_rate.get_value() > self.min_lr: |
|
111
|
|
|
self._learning_rate.set_value(self._learning_rate.get_value() / self.decay_factor) |
|
112
|
|
|
|
|
113
|
|
|
def invoke(self): |
|
114
|
|
|
if self.debug: |
|
115
|
|
|
logging.info("learning rate: %.8f" % self._learning_rate.get_value()) |
|
116
|
|
|
|
|
117
|
|
|
|
|
118
|
|
|
class SimpleScheduler(TrainingController): |
|
119
|
|
|
|
|
120
|
|
|
""" |
|
121
|
|
|
Simple scheduler with maximum patience. |
|
122
|
|
|
""" |
|
123
|
|
|
|
|
124
|
|
|
def __init__(self, end_at=10): |
|
125
|
|
|
""" |
|
126
|
|
|
:type trainer: deepy.trainers.base.NeuralTrainer |
|
127
|
|
|
""" |
|
128
|
|
|
self._iter = 0 |
|
129
|
|
|
self._patience = end_at |
|
130
|
|
|
|
|
131
|
|
|
def invoke(self): |
|
132
|
|
|
""" |
|
133
|
|
|
Run it, return whether to end training. |
|
134
|
|
|
""" |
|
135
|
|
|
self._iter += 1 |
|
136
|
|
|
logging.info("{} epochs left to run".format(self._patience - self._iter)) |
|
137
|
|
|
if self._iter >= self._patience: |
|
138
|
|
|
self._trainer.exit() |
zomux /
deepy
