detectron2/layers源码笔记
from .wrappers import (
BatchNorm2d,
Conv2d, #在torch.conv2d的基础上集成了norm层和activation层
ConvTranspose2d,
cat,
interpolate,
Linear,
nonzero_tuple, #nonzero_tuple(x)得到tuple of 每个维度的索引
cross_entropy,
empty_input_loss_func_wrapper,
shapes_to_tensor, #等价于torch.as_tensor()
move_device_like(a,b), #a.to(b.device)
)
FrozenBatchNorm2d中提供了两个有力的对batchnorm进行frozen和解frozen的函数，并且只需要接收model，会递归地对模型中的batchnorm部分进行变换