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
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
|
import unittest
import torch
import torch.optim as optim
import torch.legacy.optim as old_optim
from torch.autograd import Variable
from common import TestCase
def rosenbrock(tensor):
x, y = tensor
return (1 - x) ** 2 + 100 * (y - x ** 2) ** 2
def drosenbrock(tensor):
x, y = tensor
return torch.DoubleTensor((-400 * x * (y - x**2) - 2 * (1 - x), 200 * (y - x**2)))
def wrap_old_fn(old_fn, **config):
def wrapper(closure, params, state):
return old_fn(closure, params, config, state)
return wrapper
class TestOptim(TestCase):
def _test_rosenbrock(self, constructor, old_fn):
params_t = torch.Tensor([1.5, 1.5])
state = {}
params = Variable(torch.Tensor([1.5, 1.5]), requires_grad=True)
optimizer = constructor([params])
solution = torch.Tensor([1, 1])
initial_dist = params.data.dist(solution)
def eval():
loss = rosenbrock(params)
loss.backward()
return loss
for i in range(2000):
optimizer.zero_grad()
optimizer.step(eval)
old_fn(lambda _: (rosenbrock(params_t), drosenbrock(params_t)),
params_t, state)
self.assertEqual(params.data, params_t)
self.assertLessEqual(params.data.dist(solution), initial_dist)
def _test_basic_cases_template(self, weight, bias, input, constructor):
weight = Variable(weight, requires_grad=True)
bias = Variable(bias, requires_grad=True)
input = Variable(input, requires_grad=False)
optimizer = constructor(weight, bias)
def fn():
y = weight.mv(input)
if y.is_cuda and bias.is_cuda and y.get_device() != bias.get_device():
y = y.cuda(bias.get_device())
return (y + bias).abs().sum()
initial_value = fn().data[0]
for i in range(200):
weight.grad.data.zero_()
bias.grad.data.zero_()
fn().backward()
optimizer.step()
self.assertLessEqual(fn().data[0], initial_value)
def _test_basic_cases(self, constructor):
self._test_basic_cases_template(
torch.randn(10, 5),
torch.randn(10),
torch.randn(5),
constructor
)
# non-contiguous parameters
self._test_basic_cases_template(
torch.randn(10, 5, 2)[...,0],
torch.randn(10, 2)[...,0],
torch.randn(5),
constructor
)
# CUDA
if not torch.cuda.is_available():
return
self._test_basic_cases_template(
torch.randn(10, 5).cuda(),
torch.randn(10).cuda(),
torch.randn(5).cuda(),
constructor
)
# Multi-GPU
if not torch.cuda.device_count() > 1:
return
self._test_basic_cases_template(
torch.randn(10, 5).cuda(),
torch.randn(10).cuda(),
torch.randn(5).cuda(),
constructor
)
def _build_params_dict(self, weight, bias, **kwargs):
return [dict(params=[weight]), dict(params=[bias], **kwargs)]
def test_sgd(self):
self._test_rosenbrock(
lambda params: optim.SGD(params, lr=1e-3),
wrap_old_fn(old_optim.sgd, learningRate=1e-3)
)
self._test_rosenbrock(
lambda params: optim.SGD(params, lr=1e-3, momentum=0.9,
dampening=0, weight_decay=1e-4),
wrap_old_fn(old_optim.sgd, learningRate=1e-3, momentum=0.9,
dampening=0, weightDecay=1e-4)
)
self._test_basic_cases(
lambda weight, bias: optim.SGD([weight, bias], lr=1e-3)
)
self._test_basic_cases(
lambda weight, bias: optim.SGD(
self._build_params_dict(weight, bias, lr=1e-2),
lr=1e-3)
)
def test_adam(self):
self._test_rosenbrock(
lambda params: optim.Adam(params, lr=1e-2),
wrap_old_fn(old_optim.adam, learningRate=1e-2)
)
self._test_rosenbrock(
lambda params: optim.Adam(params, lr=1e-2, weight_decay=1e-2),
wrap_old_fn(old_optim.adam, learningRate=1e-2, weightDecay=1e-2)
)
self._test_basic_cases(
lambda weight, bias: optim.Adam([weight, bias], lr=1e-3)
)
self._test_basic_cases(
lambda weight, bias: optim.Adam(
self._build_params_dict(weight, bias, lr=1e-2),
lr=1e-3)
)
def test_adadelta(self):
self._test_rosenbrock(
lambda params: optim.Adadelta(params),
wrap_old_fn(old_optim.adadelta)
)
self._test_rosenbrock(
lambda params: optim.Adadelta(params, rho=0.95),
wrap_old_fn(old_optim.adadelta, rho=0.95)
)
self._test_rosenbrock(
lambda params: optim.Adadelta(params, weight_decay=1e-2),
wrap_old_fn(old_optim.adadelta, weightDecay=1e-2)
)
self._test_basic_cases(
lambda weight, bias: optim.Adadelta([weight, bias])
)
self._test_basic_cases(
lambda weight, bias: optim.Adadelta(
self._build_params_dict(weight, bias, rho=0.95))
)
def test_adagrad(self):
self._test_rosenbrock(
lambda params: optim.Adagrad(params, lr=1e-1),
wrap_old_fn(old_optim.adagrad, learningRate=1e-1)
)
self._test_rosenbrock(
lambda params: optim.Adagrad(params, lr=1e-1, lr_decay=1e-3),
wrap_old_fn(old_optim.adagrad, learningRate=1e-1, learningRateDecay=1e-3)
)
self._test_rosenbrock(
lambda params: optim.Adagrad(params, lr=1e-1, weight_decay=1e-2),
wrap_old_fn(old_optim.adagrad, learningRate=1e-1, weightDecay=1e-2)
)
self._test_basic_cases(
lambda weight, bias: optim.Adagrad([weight, bias], lr=1e-1)
)
self._test_basic_cases(
lambda weight, bias: optim.Adagrad(
self._build_params_dict(weight, bias, lr=1e-2),
lr=1e-1)
)
def test_adamax(self):
self._test_rosenbrock(
lambda params: optim.Adamax(params, lr=1e-1),
wrap_old_fn(old_optim.adamax, learningRate=1e-1)
)
self._test_rosenbrock(
lambda params: optim.Adamax(params, lr=1e-1, weight_decay=1e-2),
wrap_old_fn(old_optim.adamax, learningRate=1e-1, weightDecay=1e-2)
)
self._test_rosenbrock(
lambda params: optim.Adamax(params, lr=1e-1, betas=(0.95, 0.998)),
wrap_old_fn(old_optim.adamax, learningRate=1e-1, beta1=0.95, beta2=0.998)
)
self._test_basic_cases(
lambda weight, bias: optim.Adagrad([weight, bias], lr=1e-1)
)
self._test_basic_cases(
lambda weight, bias: optim.Adagrad(
self._build_params_dict(weight, bias, lr=1e-2),
lr=1e-1)
)
def test_rmsprop(self):
self._test_rosenbrock(
lambda params: optim.RMSprop(params, lr=1e-2),
wrap_old_fn(old_optim.rmsprop, learningRate=1e-2)
)
self._test_rosenbrock(
lambda params: optim.RMSprop(params, lr=1e-2, weight_decay=1e-2),
wrap_old_fn(old_optim.rmsprop, learningRate=1e-2, weightDecay=1e-2)
)
self._test_rosenbrock(
lambda params: optim.RMSprop(params, lr=1e-2, alpha=0.95),
wrap_old_fn(old_optim.rmsprop, learningRate=1e-2, alpha=0.95)
)
self._test_basic_cases(
lambda weight, bias: optim.Adagrad([weight, bias], lr=1e-2)
)
self._test_basic_cases(
lambda weight, bias: optim.Adagrad(
self._build_params_dict(weight, bias, lr=1e-3),
lr=1e-2)
)
def test_asgd(self):
self._test_rosenbrock(
lambda params: optim.ASGD(params, lr=1e-3),
wrap_old_fn(old_optim.asgd, eta0=1e-3)
)
self._test_rosenbrock(
lambda params: optim.ASGD(params, lr=1e-3, alpha=0.8),
wrap_old_fn(old_optim.asgd, eta0=1e-3, alpha=0.8)
)
self._test_rosenbrock(
lambda params: optim.ASGD(params, lr=1e-3, t0=1e3),
wrap_old_fn(old_optim.asgd, eta0=1e-3, t0=1e3)
)
self._test_basic_cases(
lambda weight, bias: optim.ASGD([weight, bias], lr=1e-3, t0=100)
)
self._test_basic_cases(
lambda weight, bias: optim.ASGD(
self._build_params_dict(weight, bias, lr=1e-2),
lr=1e-3, t0=100)
)
def test_rprop(self):
self._test_rosenbrock(
lambda params: optim.Rprop(params, lr=1e-3),
wrap_old_fn(old_optim.rprop, stepsize=1e-3)
)
self._test_rosenbrock(
lambda params: optim.Rprop(params, lr=1e-3, etas=(0.6, 1.1)),
wrap_old_fn(old_optim.rprop, stepsize=1e-3, etaminus=0.6, etaplus=1.1)
)
self._test_rosenbrock(
lambda params: optim.Rprop(params, lr=1e-3, step_sizes=(1e-4, 3)),
wrap_old_fn(old_optim.rprop, stepsize=1e-3, stepsizemin=1e-4, stepsizemax=3)
)
self._test_basic_cases(
lambda weight, bias: optim.Rprop([weight, bias], lr=1e-3)
)
self._test_basic_cases(
lambda weight, bias: optim.Rprop(
self._build_params_dict(weight, bias, lr=1e-2),
lr=1e-3)
)
def test_invalid_param_type(self):
with self.assertRaises(TypeError):
optim.SGD(Variable(torch.randn(5, 5)), lr=3)
if __name__ == '__main__':
unittest.main()
|