torch.nn.modules.batchnorm 的源代码

from typing import Optional, Any

import torch
from torch import Tensor
from torch.nn.parameter import Parameter, UninitializedParameter, UninitializedBuffer

from .. import functional as F
from .. import init
from ._functions import SyncBatchNorm as sync_batch_norm
from .lazy import LazyModuleMixin
from .module import Module

__all__ = ['BatchNorm1d', 'LazyBatchNorm1d', 'BatchNorm2d', 'LazyBatchNorm2d', 'BatchNorm3d',
           'LazyBatchNorm3d', 'SyncBatchNorm']


class _NormBase(Module):
    """_InstanceNorm 和 _BatchNorm 的通用基类。"""

    _version = 2
    __constants__ = ["track_running_stats", "momentum", "eps", "num_features", "affine"]
    num_features: int
    eps: float
    momentum: float
    affine: bool
    track_running_stats: bool
    # 警告：weight 和 bias 故意不在这里定义。
    # 参见 https://github.com/pytorch/pytorch/issues/39670

    def __init__(
        self,
        num_features: int,
        eps: float = 1e-5,
        momentum: float = 0.1,
        affine: bool = True,
        track_running_stats: bool = True,
        device=None,
        dtype=None
    ) -> None:
        factory_kwargs = {'device': device, 'dtype': dtype}
        super().__init__()
        self.num_features = num_features
        self.eps = eps
        self.momentum = momentum
        self.affine = affine
        self.track_running_stats = track_running_stats
        if self.affine:
            self.weight = Parameter(torch.empty(num_features, **factory_kwargs))
            self.bias = Parameter(torch.empty(num_features, **factory_kwargs))
        else:
            self.register_parameter("weight", None)
            self.register_parameter("bias", None)
        if self.track_running_stats:
            self.register_buffer('running_mean', torch.zeros(num_features, **factory_kwargs))
            self.register_buffer('running_var', torch.ones(num_features, **factory_kwargs))
            self.running_mean: Optional[Tensor]
            self.running_var: Optional[Tensor]
            self.register_buffer('num_batches_tracked',
                                 torch.tensor(0, dtype=torch.long,
                                              **{k: v for k, v in factory_kwargs.items() if k != 'dtype'}))
            self.num_batches_tracked: Optional[Tensor]
        else:
            self.register_buffer("running_mean", None)
            self.register_buffer("running_var", None)
            self.register_buffer("num_batches_tracked", None)
        self.reset_parameters()

    def reset_running_stats(self) -> None:
        if self.track_running_stats:
            # running_mean/running_var/num_batches... 是在运行时根据
            # self.track_running_stats 是否开启来注册的
            self.running_mean.zero_()  # type: ignore[union-attr]
            self.running_var.fill_(1)  # type: ignore[union-attr]
            self.num_batches_tracked.zero_()  # type: ignore[union-attr,operator]

    def reset_parameters(self) -> None:
        self.reset_running_stats()
        if self<