forked from open-mmlab/mmengine
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtest_averaged_model.py
263 lines (237 loc) · 10.7 KB
/
test_averaged_model.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
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
# Copyright (c) OpenMMLab. All rights reserved.
import itertools
from unittest import TestCase
import torch
from mmengine.logging import MMLogger
from mmengine.model import (ExponentialMovingAverage, MomentumAnnealingEMA,
StochasticWeightAverage)
from mmengine.testing import assert_allclose
class TestAveragedModel(TestCase):
"""Test the AveragedModel class.
Some test cases are referenced from https://github.com/pytorch/pytorch/blob/master/test/test_optim.py
""" # noqa: E501
def _test_swa_model(self, net_device, avg_device):
model = torch.nn.Sequential(
torch.nn.Conv2d(1, 5, kernel_size=3),
torch.nn.Linear(5, 10)).to(net_device)
averaged_model = StochasticWeightAverage(model, device=avg_device)
averaged_params = [
torch.zeros_like(param) for param in model.parameters()
]
n_updates = 2
for i in range(n_updates):
for p, p_avg in zip(model.parameters(), averaged_params):
p.detach().add_(torch.randn_like(p))
p_avg += p.detach() / n_updates
averaged_model.update_parameters(model)
for p_avg, p_swa in zip(averaged_params, averaged_model.parameters()):
# Check that AveragedModel is on the correct device
self.assertTrue(p_swa.device == avg_device)
self.assertTrue(p.device == net_device)
assert_allclose(p_avg, p_swa.to(p_avg.device))
self.assertTrue(averaged_model.steps.device == avg_device)
def test_averaged_model_all_devices(self):
cpu = torch.device('cpu')
self._test_swa_model(cpu, cpu)
if torch.cuda.is_available():
cuda = torch.device(0)
self._test_swa_model(cuda, cpu)
self._test_swa_model(cpu, cuda)
self._test_swa_model(cuda, cuda)
def test_swa_mixed_device(self):
if not torch.cuda.is_available():
return
model = torch.nn.Sequential(
torch.nn.Conv2d(1, 5, kernel_size=3), torch.nn.Linear(5, 10))
model[0].cuda()
model[1].cpu()
averaged_model = StochasticWeightAverage(model)
averaged_params = [
torch.zeros_like(param) for param in model.parameters()
]
n_updates = 10
for i in range(n_updates):
for p, p_avg in zip(model.parameters(), averaged_params):
p.detach().add_(torch.randn_like(p))
p_avg += p.detach() / n_updates
averaged_model.update_parameters(model)
for p_avg, p_swa in zip(averaged_params, averaged_model.parameters()):
assert_allclose(p_avg, p_swa)
# Check that AveragedModel is on the correct device
self.assertTrue(p_avg.device == p_swa.device)
def test_swa_state_dict(self):
model = torch.nn.Sequential(
torch.nn.Conv2d(1, 5, kernel_size=3), torch.nn.Linear(5, 10))
averaged_model = StochasticWeightAverage(model)
averaged_model2 = StochasticWeightAverage(model)
n_updates = 10
for i in range(n_updates):
for p in model.parameters():
p.detach().add_(torch.randn_like(p))
averaged_model.update_parameters(model)
averaged_model2.load_state_dict(averaged_model.state_dict())
for p_swa, p_swa2 in zip(averaged_model.parameters(),
averaged_model2.parameters()):
assert_allclose(p_swa, p_swa2)
self.assertTrue(averaged_model.steps == averaged_model2.steps)
def test_ema(self):
# test invalid momentum
with self.assertRaisesRegex(AssertionError,
'momentum must be in range'):
model = torch.nn.Sequential(
torch.nn.Conv2d(1, 5, kernel_size=3), torch.nn.Linear(5, 10))
ExponentialMovingAverage(model, momentum=3)
# Warning should be raised if the value of momentum in EMA is
# a large number
with self.assertLogs(MMLogger.get_current_instance(), level='WARNING'):
model = torch.nn.Sequential(
torch.nn.Conv2d(1, 5, kernel_size=3), torch.nn.Linear(5, 10))
ExponentialMovingAverage(model, momentum=0.9)
# test EMA
model = torch.nn.Sequential(
torch.nn.Conv2d(1, 5, kernel_size=3), torch.nn.Linear(5, 10))
momentum = 0.1
ema_model = ExponentialMovingAverage(model, momentum=momentum)
averaged_params = [
torch.zeros_like(param) for param in model.parameters()
]
n_updates = 10
for i in range(n_updates):
updated_averaged_params = []
for p, p_avg in zip(model.parameters(), averaged_params):
p.detach().add_(torch.randn_like(p))
if i == 0:
updated_averaged_params.append(p.clone())
else:
updated_averaged_params.append(
(p_avg * (1 - momentum) + p * momentum).clone())
ema_model.update_parameters(model)
averaged_params = updated_averaged_params
for p_target, p_ema in zip(averaged_params, ema_model.parameters()):
assert_allclose(p_target, p_ema)
def test_ema_update_buffers(self):
# Test EMA and update_buffers as True.
model = torch.nn.Sequential(
torch.nn.Conv2d(1, 5, kernel_size=3),
torch.nn.BatchNorm2d(5, momentum=0.3), torch.nn.Linear(5, 10))
momentum = 0.1
ema_model = ExponentialMovingAverage(
model, momentum=momentum, update_buffers=True)
averaged_params = [
torch.zeros_like(param)
for param in itertools.chain(model.parameters(), model.buffers())
if param.size() != torch.Size([])
]
n_updates = 10
for i in range(n_updates):
updated_averaged_params = []
params = [
param for param in itertools.chain(model.parameters(),
model.buffers())
if param.size() != torch.Size([])
]
for p, p_avg in zip(params, averaged_params):
p.detach().add_(torch.randn_like(p))
if i == 0:
updated_averaged_params.append(p.clone())
else:
updated_averaged_params.append(
(p_avg * (1 - momentum) + p * momentum).clone())
ema_model.update_parameters(model)
averaged_params = updated_averaged_params
ema_params = [
param for param in itertools.chain(ema_model.module.parameters(),
ema_model.module.buffers())
if param.size() != torch.Size([])
]
for p_target, p_ema in zip(averaged_params, ema_params):
assert_allclose(p_target, p_ema)
def test_momentum_annealing_ema(self):
model = torch.nn.Sequential(
torch.nn.Conv2d(1, 5, kernel_size=3),
torch.nn.BatchNorm2d(5, momentum=0.3), torch.nn.Linear(5, 10))
# Test invalid gamma
with self.assertRaisesRegex(AssertionError,
'gamma must be greater than 0'):
MomentumAnnealingEMA(model, gamma=-1)
# Test EMA with momentum annealing.
momentum = 0.1
gamma = 4
ema_model = MomentumAnnealingEMA(
model, gamma=gamma, momentum=momentum, update_buffers=True)
averaged_params = [
torch.zeros_like(param)
for param in itertools.chain(model.parameters(), model.buffers())
if param.size() != torch.Size([])
]
n_updates = 10
for i in range(n_updates):
updated_averaged_params = []
params = [
param for param in itertools.chain(model.parameters(),
model.buffers())
if param.size() != torch.Size([])
]
for p, p_avg in zip(params, averaged_params):
p.add(torch.randn_like(p))
if i == 0:
updated_averaged_params.append(p.clone())
else:
m = max(momentum, gamma / (gamma + i))
updated_averaged_params.append(
(p_avg * (1 - m) + p * m).clone())
ema_model.update_parameters(model)
averaged_params = updated_averaged_params
ema_params = [
param for param in itertools.chain(ema_model.module.parameters(),
ema_model.module.buffers())
if param.size() != torch.Size([])
]
for p_target, p_ema in zip(averaged_params, ema_params):
assert_allclose(p_target, p_ema)
def test_momentum_annealing_ema_with_interval(self):
# Test EMA with momentum annealing and interval
model = torch.nn.Sequential(
torch.nn.Conv2d(1, 5, kernel_size=3),
torch.nn.BatchNorm2d(5, momentum=0.3), torch.nn.Linear(5, 10))
momentum = 0.1
gamma = 4
interval = 3
ema_model = MomentumAnnealingEMA(
model,
gamma=gamma,
momentum=momentum,
interval=interval,
update_buffers=True)
averaged_params = [
torch.zeros_like(param)
for param in itertools.chain(model.parameters(), model.buffers())
if param.size() != torch.Size([])
]
n_updates = 10
for i in range(n_updates):
updated_averaged_params = []
params = [
param for param in itertools.chain(model.parameters(),
model.buffers())
if param.size() != torch.Size([])
]
for p, p_avg in zip(params, averaged_params):
p.add(torch.randn_like(p))
if i == 0:
updated_averaged_params.append(p.clone())
elif i % interval == 0:
m = max(momentum, gamma / (gamma + i))
updated_averaged_params.append(
(p_avg * (1 - m) + p * m).clone())
else:
updated_averaged_params.append(p_avg.clone())
ema_model.update_parameters(model)
averaged_params = updated_averaged_params
ema_params = [
param for param in itertools.chain(ema_model.module.parameters(),
ema_model.module.buffers())
if param.size() != torch.Size([])
]
for p_target, p_ema in zip(averaged_params, ema_params):
assert_allclose(p_target, p_ema)