Normalizations

四种 Norm 对比

• BN：垂直于 C 维度归一化
• LN：垂直于 N 维度归一化
• IN：垂直于 N, C 维度归一化
• GN：$GN = \begin{cases} LN,\ \ \ g=1\\ IN,\ \ \ g=c \end{cases}$，g 表示每个 group 覆盖的 channel 数

四种 Norm 的 Affine 参数 Shape

import torch
inp = torch.randn(3, 4, 5, 6)
# batch norm
batchnorm = torch.nn.BatchNorm2d(num_features=4)
batchnorm.weight.shape                              # torch.Size([4])
out = batchnorm(inp)
out.shape                                           # torch.Size([3, 4, 5, 6])
out.permute(1,0,2,3).reshape(4, -1).mean(1).data    # tensor([-5.9605e-09,  1.5895e-08, -5.2982e-09, -7.9473e-09])  几乎是 0
out.permute(1,0,2,3).reshape(4, -1).std(1).data     # tensor([1.0056, 1.0056, 1.0056, 1.0056])  几乎是 1
# layer norm
layernorm = torch.nn.LayerNorm(normalized_shape=(4, 5, 6))
layernorm.weight.shape              # torch.Size([4, 5, 6])
out = layernorm(inp)
out.shape                           # torch.Size([3, 4, 5, 6])
out.reshape(3, -1).mean(1).data     # tensor([ 3.9736e-09, -5.9605e-09,  2.7816e-08])  几乎是 0
out.reshape(3, -1).std(1).data      # tensor([1.0042, 1.0042, 1.0042])  几乎是 1
# instance norm
instancenorm = torch.nn.InstanceNorm2d(num_features=4)
out = instancenorm(inp)
out.shape                           # torch.Size([3, 4, 5, 6])
out.reshape(12, -1).mean(1).data    # tensor([ 1.7881e-08,  1.7881e-08, -2.3842e-08, -1.9868e-08,  0.0000e+00, -3.9736e-09,  0.0000e+00,  0.0000e+00,  0.0000e+00,  7.9473e-09, 1.9868e-08, -2.1855e-08])  几乎是 0
out.reshape(12, -1).std(1).data     # tensor([1.0171, 1.0171, 1.0171, 1.0171, 1.0171, 1.0171, 1.0171, 1.0171, 1.0171, 1.0171, 1.0171, 1.0171])  几乎是 1
# group norm
groupnorm = torch.nn.GroupNorm(num_groups=2, num_channels=4)
groupnorm.weight.shape              # torch.Size([4, 5, 6])
out = groupnorm(inp)
out.shape                           # torch.Size([3, 4, 5, 6])
out.reshape(6, -1).mean(1).data     # tensor([ 1.9868e-09,  3.1789e-08, -1.9868e-08,  0.0000e+00,  1.5895e-08, 5.9605e-08])  几乎是 0
out.reshape(6, -1).std(1).data      # tensor([1.0084, 1.0084, 1.0084, 1.0084, 1.0084, 1.0084])  几乎是 1