resnext/nutszebra_optimizer.py at master · nutszebra/resnext · GitHub

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
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
import chainer
from chainer import optimizers
import nutszebra_basic_print


class Optimizer(object):

    def __init__(self, model=None):
        self.model = model
        self.optimizer = None

    def __call__(self, i):
        pass

    def update(self):
        self.optimizer.update()


class OptimizerResnet(Optimizer):

    def __init__(self, model=None, schedule=(int(32000. / (50000. / 128)), int(48000. / (50000. / 128))), lr=0.1, momentum=0.9, weight_decay=1.0e-4, warm_up_lr=0.01):
        super(OptimizerResnet, self).__init__(model)
        optimizer = optimizers.MomentumSGD(warm_up_lr, momentum)
        weight_decay = chainer.optimizer.WeightDecay(weight_decay)
        optimizer.setup(self.model)
        optimizer.add_hook(weight_decay)
        self.optimizer = optimizer
        self.schedule = schedule
        self.lr = lr
        self.warmup_lr = warm_up_lr
        self.momentum = momentum
        self.weight_decay = weight_decay

    def __call__(self, i):
        if i == 1:
            lr = self.lr
            print('finishded warming up')
            print('lr is changed: {} -> {}'.format(self.optimizer.lr, lr))
            self.optimizer.lr = lr
        if i in self.schedule:
            lr = self.optimizer.lr / 10
            print('lr is changed: {} -> {}'.format(self.optimizer.lr, lr))
            self.optimizer.lr = lr


class OptimizerDense(Optimizer):

    def __init__(self, model=None, schedule=(150, 225), lr=0.1, momentum=0.9, weight_decay=1.0e-4):
        super(OptimizerDense, self).__init__(model)
        optimizer = optimizers.MomentumSGD(lr, momentum)
        weight_decay = chainer.optimizer.WeightDecay(weight_decay)
        optimizer.setup(self.model)
        optimizer.add_hook(weight_decay)
        self.optimizer = optimizer
        self.schedule = schedule
        self.lr = lr
        self.momentum = momentum
        self.weight_decay = weight_decay

    def __call__(self, i):
        if i in self.schedule:
            lr = self.optimizer.lr / 10
            print('lr is changed: {} -> {}'.format(self.optimizer.lr, lr))
            self.optimizer.lr = lr


class OptimizerWideRes(Optimizer):

    def __init__(self, model=None, schedule=(60, 120, 160), lr=0.1, momentum=0.9, weight_decay=5.0e-4):
        super(OptimizerWideRes, self).__init__(model)
        optimizer = optimizers.MomentumSGD(lr, momentum)
        weight_decay = chainer.optimizer.WeightDecay(weight_decay)
        optimizer.setup(self.model)
        optimizer.add_hook(weight_decay)
        self.optimizer = optimizer
        self.schedule = schedule
        self.lr = lr
        self.momentum = momentum
        self.weight_decay = weight_decay

    def __call__(self, i):
        if i in self.schedule:
            lr = self.optimizer.lr * 0.2
            print('lr is changed: {} -> {}'.format(self.optimizer.lr, lr))
            self.optimizer.lr = lr


class OptimizerSwapout(Optimizer):

    def __init__(self, model=None, schedule=(196, 224), lr=0.1, momentum=0.9, weight_decay=1.0e-4):
        super(OptimizerSwapout, self).__init__(model)
        optimizer = optimizers.MomentumSGD(lr, momentum)
        weight_decay = chainer.optimizer.WeightDecay(weight_decay)
        optimizer.setup(self.model)
        optimizer.add_hook(weight_decay)
        self.optimizer = optimizer
        self.schedule = schedule
        self.lr = lr
        self.momentum = momentum
        self.weight_decay = weight_decay

    def __call__(self, i):
        if i in self.schedule:
            lr = self.optimizer.lr / 10
            print('lr is changed: {} -> {}'.format(self.optimizer.lr, lr))
            self.optimizer.lr = lr


class OptimizerXception(Optimizer):

    def __init__(self, model=None, lr=0.045, momentum=0.9, weight_decay=1.0e-5, period=2):
        super(OptimizerXception, self).__init__(model)
        optimizer = optimizers.MomentumSGD(lr, momentum)
        weight_decay = chainer.optimizer.WeightDecay(weight_decay)
        optimizer.setup(self.model)
        optimizer.add_hook(weight_decay)
        self.optimizer = optimizer
        self.lr = lr
        self.momentum = momentum
        self.weight_decay = weight_decay
        self.period = int(period)

    def __call__(self, i):
        if i % self.period == 0:
            lr = self.optimizer.lr * 0.94
            print('lr is changed: {} -> {}'.format(self.optimizer.lr, lr))
            self.optimizer.lr = lr


class OptimizerVGG(Optimizer):

    def __init__(self, model=None, lr=0.01, momentum=0.9, weight_decay=5.0e-4):
        super(OptimizerVGG, self).__init__(model)
        optimizer = optimizers.MomentumSGD(lr, momentum)
        weight_decay = chainer.optimizer.WeightDecay(weight_decay)
        optimizer.setup(self.model)
        optimizer.add_hook(weight_decay)
        self.optimizer = optimizer
        self.lr = lr
        self.momentum = momentum
        self.weight_decay = weight_decay

    def __call__(self, i):
        # 150 epoch means (0.94 ** 75) * lr
        # if lr is 0.01, then (0.94 ** 75) * 0.01 is 0.0001 at the end
        if i % 2 == 0:
            lr = self.optimizer.lr * 0.94
            print('lr is changed: {} -> {}'.format(self.optimizer.lr, lr))
            self.optimizer.lr = lr


class OptimizerGooglenet(Optimizer):

    def __init__(self, model=None, lr=0.0015, momentum=0.9, weight_decay=2.0e-4):
        super(OptimizerGooglenet, self).__init__(model)
        optimizer = optimizers.MomentumSGD(lr, momentum)
        weight_decay = chainer.optimizer.WeightDecay(weight_decay)
        optimizer.setup(self.model)
        optimizer.add_hook(weight_decay)
        self.optimizer = optimizer
        self.lr = lr
        self.momentum = momentum
        self.weight_decay = weight_decay

    def __call__(self, i):
        if i % 8 == 0:
            lr = self.optimizer.lr * 0.96
            print('lr is changed: {} -> {}'.format(self.optimizer.lr, lr))
            self.optimizer.lr = lr


class OptimizerNetworkInNetwork(Optimizer):

    def __init__(self, model=None, lr=0.1, momentum=0.9, weight_decay=1.0e-4, schedule=(int(1.0e5 / (50000. / 128)), )):
        super(OptimizerNetworkInNetwork, self).__init__(model)
        optimizer = optimizers.MomentumSGD(lr, momentum)
        weight_decay = chainer.optimizer.WeightDecay(weight_decay)
        optimizer.setup(self.model)
        optimizer.add_hook(weight_decay)
        self.optimizer = optimizer
        self.lr = lr
        self.momentum = momentum
        self.weight_decay = weight_decay
        self.schedule = schedule

    def __call__(self, i):
        if i in self.schedule:
            lr = self.optimizer.lr / 10
            print('lr is changed: {} -> {}'.format(self.optimizer.lr, lr))
            self.optimizer.lr = lr


class OptimizerGooglenetV2(Optimizer):

    def __init__(self, model=None, lr=0.045, momentum=0.9, weight_decay=4.0e-5):
        super(OptimizerGooglenetV2, self).__init__(model)
        optimizer = optimizers.MomentumSGD(lr, momentum)
        weight_decay = chainer.optimizer.WeightDecay(weight_decay)
        optimizer.setup(self.model)
        optimizer.add_hook(weight_decay)
        self.optimizer = optimizer
        self.lr = lr
        self.momentum = momentum
        self.weight_decay = weight_decay

    def __call__(self, i):
        if i % 2 == 0:
            lr = self.optimizer.lr * 0.94
            print('lr is changed: {} -> {}'.format(self.optimizer.lr, lr))
            self.optimizer.lr = lr


class OptimizerGooglenetV3(Optimizer):

    def __init__(self, model=None, lr=0.045, decay=0.9, eps=1.0, weight_decay=4.0e-5, clip=2.0):
        super(OptimizerGooglenetV3, self).__init__(model)
        optimizer = optimizers.RMSprop(lr, decay, eps)
        weight_decay = chainer.optimizer.WeightDecay(weight_decay)
        clip = chainer.optimizer.GradientClipping(clip)
        optimizer.setup(self.model)
        optimizer.add_hook(weight_decay)
        optimizer.add_hook(clip)
        self.optimizer = optimizer

    def __call__(self, i):
        if i % 2 == 0:
            lr = self.optimizer.lr * 0.94
            print('lr is changed: {} -> {}'.format(self.optimizer.lr, lr))
            self.optimizer.lr = lr


class OptimizerResNext(Optimizer):

    def __init__(self, model=None, lr=0.1, momentum=0.9, weight_decay=5.0e-4, schedule=(150, 225)):
        super(OptimizerResNext, self).__init__(model)
        optimizer = optimizers.MomentumSGD(lr, momentum)
        weight_decay = chainer.optimizer.WeightDecay(weight_decay)
        optimizer.setup(self.model)
        optimizer.add_hook(weight_decay)
        self.optimizer = optimizer
        self.schedule = schedule
        self.lr = lr
        self.momentum = momentum
        self.weight_decay = weight_decay

    def __call__(self, i):
        if i in self.schedule:
            lr = self.optimizer.lr / 10
            print('lr is changed: {} -> {}'.format(self.optimizer.lr, lr))
            self.optimizer.lr = lr