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
|
# Copyright 2019 Kakao Brain
#
# Copyright (c) Facebook, Inc. and its affiliates. All rights reserved.
#
# This source code is licensed under the BSD license found in the
# LICENSE file in the root directory of this source tree.
"""Tracks the running statistics per mini-batch instead of micro-batch."""
from typing import TypeVar, cast
import torch
from torch import Tensor, nn
from torch.nn.functional import batch_norm
from torch.nn.modules.batchnorm import _BatchNorm
from .checkpoint import is_recomputing
__all__ = ["DeferredBatchNorm"]
TModule = TypeVar("TModule", bound=nn.Module)
class DeferredBatchNorm(_BatchNorm):
"""A BatchNorm layer tracks multiple micro-batches to update running
statistics per mini-batch.
"""
sum: Tensor
sum_squares: Tensor
running_mean: Tensor
running_var: Tensor
num_batches_tracked: Tensor
def __init__(
self,
num_features: int,
eps: float = 1e-5,
momentum: float = 0.1,
affine: bool = True,
chunks: int = 1,
) -> None:
super().__init__(num_features, eps, momentum, affine, track_running_stats=True)
self.register_buffer("sum", torch.zeros_like(self.running_mean))
self.register_buffer("sum_squares", torch.zeros_like(self.running_var))
self.counter = 0
self.tracked = 0
self.chunks = chunks
def _check_input_dim(self, input: Tensor) -> None:
# It's the typical _check_input_dim() implementation in PyTorch.
if input.dim() <= 2:
raise ValueError("expected at least 3D input (got %dD input)" % input.dim())
def _track(self, input: Tensor) -> bool:
"""Tracks statistics of a micro-batch."""
# Dimensions except channel. For example, (0, 2, 3) is for BatchNorm2d.
dim = [0]
dim.extend(range(2, input.dim()))
with torch.no_grad():
self.sum += input.sum(dim)
self.sum_squares += (input ** 2).sum(dim)
size = input.size().numel() // input.size(1)
self.counter += size
self.tracked += 1
return self.tracked == self.chunks
def _commit(self) -> None:
"""Updates the running statistics of a mini-batch."""
exponential_average_factor = 0.0
self.num_batches_tracked += 1
if self.momentum is None: # use cumulative moving average
exponential_average_factor = 1.0 / float(self.num_batches_tracked)
else: # use exponential moving average
exponential_average_factor = self.momentum
mean = self.sum / self.counter
var = self.sum_squares / self.counter - mean ** 2
# Calculate the exponential moving average here.
m = exponential_average_factor
self.running_mean *= 1 - m
self.running_mean += mean * m
self.running_var *= 1 - m
self.running_var += var * m
self.sum.zero_()
self.sum_squares.zero_()
self.counter = 0
self.tracked = 0
def forward(self, input: Tensor) -> Tensor:
if not self.training:
# Don't train parameters on the evaluation mode.
return batch_norm(
input,
running_mean=self.running_mean,
running_var=self.running_var,
weight=self.weight,
bias=self.bias,
training=False,
momentum=0.0,
eps=self.eps,
)
if not is_recomputing():
# Track a micro-batch on the training mode
# but not under a recomputation.
tracked_enough = self._track(input)
# Update the running statistics for a mini-batch
# if it has tracked enough micro-batches.
if tracked_enough:
self._commit()
# Normalize a micro-batch and train the parameters.
return batch_norm(
input,
running_mean=None,
running_var=None,
weight=self.weight,
bias=self.bias,
training=True,
momentum=0.0,
eps=self.eps,
)
@classmethod
def convert_deferred_batch_norm(cls, module: TModule, chunks: int = 1) -> TModule:
"""Converts a :class:`nn.BatchNorm` or underlying
:class:`nn.BatchNorm`s into :class:`DeferredBatchNorm`::
from torchvision.models.resnet import resnet101
from torchpipe.batchnorm import DeferredBatchNorm
model = resnet101()
model = DeferredBatchNorm.convert_deferred_batch_norm(model)
"""
if isinstance(module, DeferredBatchNorm) and module.chunks is chunks:
return cast(TModule, module)
module_output: nn.Module = module
if isinstance(module, _BatchNorm) and module.track_running_stats:
module_output = DeferredBatchNorm(module.num_features, module.eps, module.momentum, module.affine, chunks)
if module.affine:
module_output.register_parameter("weight", module.weight)
module_output.register_parameter("bias", module.bias)
module_output.register_buffer("running_mean", module.running_mean)
module_output.register_buffer("running_var", module.running_var)
module_output.register_buffer("num_batches_tracked", module.num_batches_tracked)
for name, child in module.named_children():
module_output.add_module(name, cls.convert_deferred_batch_norm(child, chunks))
return cast(TModule, module_output)
|
