网络参数初始化的优劣在极大程度上决定了网络的最终性能。比较推荐的初始化方式有 He 初始化是,将参数初始化为服从高斯分布或均匀分布的较小随机整数,同时对参数方差加以规范化。
处于稳定状态下的神经网络,其参数和数据均值为 0。
PyTorch 中实现了较多的参数初始化 API, 主要如下:
- 均匀分布初始化
- 正态分布初始化
- 全0初始化
- 全1初始化
- 固定值初始化
- kaiming 初始化
- xavier 初始化
示例代码:
import torch import torch.nn.functional as F import torch.nn as nn # 1. 均匀分布随机初始化 def test01(): linear = nn.Linear(5, 3) # 从0-1均匀分布产生参数 nn.init.uniform_(linear.weight) print(linear.weight.data) # 2. 固定初始化 def test02(): linear = nn.Linear(5, 3) nn.init.constant_(linear.weight, 5) print(linear.weight.data) # 3. 全0初始化 def test03(): linear = nn.Linear(5, 3) nn.init.zeros_(linear.weight) print(linear.weight.data) # 4. 全1初始化 def test04(): linear = nn.Linear(5, 3) nn.init.ones_(linear.weight) print(linear.weight.data) # 5. 正态分布随机初始化 def test05(): linear = nn.Linear(5, 3) nn.init.normal_(linear.weight, mean=0, std=1) print(linear.weight.data) # 6. kaiming 初始化 def test06(): # kaiming 正态分布初始化 linear = nn.Linear(5, 3) nn.init.kaiming_normal_(linear.weight) print(linear.weight.data) # kaiming 均匀分布初始化 linear = nn.Linear(5, 3) nn.init.kaiming_uniform_(linear.weight) print(linear.weight.data) # 7. xavier 初始化 def test07(): # xavier 正态分布初始化 linear = nn.Linear(5, 3) nn.init.xavier_normal_(linear.weight) print(linear.weight.data) # xavier 均匀分布初始化 linear = nn.Linear(5, 3) nn.init.xavier_uniform_(linear.weight) print(linear.weight.data) if __name__ == '__main__': test07()