【小工具】参数初始化

import torch
import numpy as np

'参数初始化'
params = torch.nn.Conv2d(1,1,3,1)
# x = list(params.parameters())[0].data
x = list(params.weight)[0].data
print("nn.Conv2d:\n",x,x.mean(),x.std(),"\n")

data = torch.randn([1,1,3,3],dtype=torch.float32)
print("torch.randn:\n",data,data.mean(),data.std(),"\n")

x = torch.nn.init.constant_(data, 0.5)#初始化为指定常数
print("init.constant_:\n",x,x.mean(),x.std(),"\n")

x = torch.nn.init.normal_(data, mean=0, std=0.002)#初始化为指定分布的参数
print("init.normal_:\n",x,x.mean(),x.std(),"\n")

x = torch.nn.init.uniform_(data, 0, 1)#初始化为0到1之间的参数
print("init.uniform_:\n",x,x.mean(),x.std(),"\n")

# "k = \frac{1}{C_\text{in} * \prod_{i=0}^{1}\text{kernel\_size}[i]}"
# "\mathcal{U}(-\sqrt{k}, \sqrt{k})"
# x = torch.nn.init.uniform_(data, -k, k)#初始化为-k到k之间的参数，pytorch中卷积核的初始化方法
# print("init.uniform_:\n",x,x.mean(),x.std(),"\n")

x = torch.nn.init.xavier_normal_(data)
print("init.xavier_normal_:\n",x,x.mean(),x.std(),"\n")

x = torch.nn.init.xavier_uniform_(data)
print("init.xavier_uniform_:\n",x,x.mean(),x.std(),"\n")

x = torch.nn.init.kaiming_normal_(data)#凯明初始化方法
print("init.kaiming_normal_:\n",x,x.mean(),x.std(),"\n")

x = torch.nn.init.kaiming_uniform_(data)#凯明初始化方法
print("init.kaiming_uniform_:\n",x,x.mean(),x.std(),"\n")

noise_d = torch.normal(0, 0.002, (self.batch_size, 128, 1, 1))