File size: 6,361 Bytes
ee21b96
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.

import unittest
from copy import deepcopy
from dataclasses import dataclass
from typing import Optional

import torch
from fairseq.models.ema import EMA


class DummyModule(torch.nn.Module):
    def __init__(self) -> None:
        """LightningModule for testing purposes

        Args:
            epoch_min_loss_override (int, optional): Pass in an epoch that will be set to the minimum
                validation loss for testing purposes (zero based). If None this is ignored. Defaults to None.
        """
        super().__init__()
        self.layer = torch.nn.Linear(in_features=32, out_features=2)
        self.another_layer = torch.nn.Linear(in_features=2, out_features=2)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = self.layer(x)
        return self.another_layer(x)


@dataclass
class EMAConfig(object):
    ema_decay: float = 0.99
    ema_start_update: int = 0
    ema_fp32: bool = False
    ema_seed_model: Optional[str] = None


class TestEMAGPU(unittest.TestCase):
    def assertTorchAllClose(self, x, y, atol=1e-8, rtol=1e-5, msg=None):
        diff = x.float() - y.float()
        diff_norm = torch.norm(diff)
        other_norm = torch.norm(y.float())

        if msg is None:
            msg = "|input - other| > {} + {} * |other|".format(
                atol, rtol
            )

        self.assertLessEqual(
            diff_norm,
            atol + rtol * other_norm,
            msg=msg,
        )

    def test_ema(self):
        model = DummyModule()
        optimizer = torch.optim.SGD(model.parameters(), lr=0.01)
        state = deepcopy(model.state_dict())
        config = EMAConfig()
        ema = EMA(model, config)

        # set decay
        ema._set_decay(config.ema_decay)
        self.assertEqual(ema.get_decay(), config.ema_decay)

        # get model
        self.assertEqual(ema.get_model(), ema.model)

        # Since fp32 params is not used, it should be of size 0
        self.assertEqual(len(ema.fp32_params), 0)

        # EMA step
        x = torch.randn(32)
        y = model(x)
        loss = y.sum()
        loss.backward()
        optimizer.step()

        ema.step(model)

        ema_state_dict = ema.get_model().state_dict()

        for key, param in model.state_dict().items():
            prev_param = state[key]
            ema_param = ema_state_dict[key]

            if "version" in key:
                # Do not decay a model.version pytorch param
                continue
            self.assertTorchAllClose(
                ema_param,
                config.ema_decay * prev_param + (1 - config.ema_decay) * param,
            )

        # Since fp32 params is not used, it should be of size 0
        self.assertEqual(len(ema.fp32_params), 0)

        # Load EMA into model
        model2 = DummyModule()
        ema.reverse(model2)

        for key, param in model2.state_dict().items():
            ema_param = ema_state_dict[key]
            self.assertTrue(
                torch.allclose(ema_param, param)
            )

    def test_ema_fp32(self):
        model = DummyModule().half()
        optimizer = torch.optim.SGD(model.parameters(), lr=0.01)
        state = deepcopy(model.state_dict())
        config = EMAConfig(ema_fp32=True)
        ema = EMA(model, config)

        x = torch.randn(32)
        y = model(x.half())
        loss = y.sum()
        loss.backward()
        optimizer.step()

        ema.step(model)

        for key, param in model.state_dict().items():
            prev_param = state[key]
            ema_param = ema.get_model().state_dict()[key]

            if "version" in key:
                # Do not decay a model.version pytorch param
                continue
            self.assertIn(key, ema.fp32_params)

            # EMA update is done in fp32, and hence the EMA param must be
            # closer to the EMA update done in fp32 than in fp16.
            self.assertLessEqual(
                torch.norm(
                    ema_param.float() -
                    (config.ema_decay * prev_param.float() + (1 - config.ema_decay) * param.float()).half().float()
                ),
                torch.norm(
                    ema_param.float() -
                    (config.ema_decay * prev_param + (1 - config.ema_decay) * param).float()
                ),
            )
            self.assertTorchAllClose(
                ema_param,
                (config.ema_decay * prev_param.float() + (1 - config.ema_decay) * param.float()).half(),
            )

    def test_ema_fp16(self):
        model = DummyModule().half()
        optimizer = torch.optim.SGD(model.parameters(), lr=0.01)
        state = deepcopy(model.state_dict())
        config = EMAConfig(ema_fp32=False)
        ema = EMA(model, config)

        # Since fp32 params is not used, it should be of size 0
        self.assertEqual(len(ema.fp32_params), 0)

        x = torch.randn(32)
        y = model(x.half())
        loss = y.sum()
        loss.backward()
        optimizer.step()

        ema.step(model)

        for key, param in model.state_dict().items():
            prev_param = state[key]
            ema_param = ema.get_model().state_dict()[key]

            if "version" in key:
                # Do not decay a model.version pytorch param
                continue

            # EMA update is done in fp16, and hence the EMA param must be
            # closer to the EMA update done in fp16 than in fp32.
            self.assertLessEqual(
                torch.norm(
                    ema_param.float() -
                    (config.ema_decay * prev_param + (1 - config.ema_decay) * param).float()
                ),
                torch.norm(
                    ema_param.float() -
                    (config.ema_decay * prev_param.float() + (1 - config.ema_decay) * param.float()).half().float()
                ),
            )
            self.assertTorchAllClose(
                ema_param,
                config.ema_decay * prev_param + (1 - config.ema_decay) * param,
            )

        # Since fp32 params is not used, it should be of size 0
        self.assertEqual(len(ema.fp32_params), 0)


if __name__ == "__main__":
    unittest.main()