pytroch tensor初始化权重、改变tensor某行列为指定值

1.几种不同的初始化方法

import torch.nn as nn

embedding = torch.Tensor(3, 5)
#如下6种初始化方法

#正态分布
nn.init.normal_(embedding)
#均匀分布
nn.init.uniform_(embedding)

#凯明均匀分布,mode可为fan_in 或 fan_out, fan_in正向传播时，方差一致;fan_out反向传播时，方差一致;nonlinearity为对应的激活函数
nn.init.kaiming_uniform_(embedding, mode='fan_in', nonlinearity='leaky_relu')
#凯明正态分布,mode可为fan_in 或 fan_out, fan_in正向传播时，方差一致;fan_out反向传播时，方差一致;nonlinearity为对应的激活函数
nn.init.kaiming_normal_(embedding, mode='fan_in', nonlinearity='leaky_relu')

#xavier初始化方法中服从正态分布，mean=0,std = gain * sqrt(2/fan_in + fan_out)
nn.init.xavier_normal_(embedding)
#avier初始化方法中服从均匀分布U(−a,a) ，分布的参数a = gain * sqrt(6/fan_in+fan_out)
nn.init.xavier_uniform_(embedding)
embedding.requires_grad=True

2.加载预训练的词向量1

import torch
import torch.nn as nn
embedding = torch.Tensor(3, 5)
nn.init.xavier_normal_(embedding)
#embedding = Variable(tensor)
data=torch.Tensor([-0.5736, -3.6566,  3.0850,  3.4097,  2.6072])#已有的词向量，
embedding[1, :] = data#data必须是tensor
embedding = nn.Parameter(embedding)#默认是可训练的
print(embedding[1])

3.加载预训练得的词向量2

import torch
import torch.nn as nn
word_embeds = nn.Embedding(vocab_size, embedding_dim)
pretrained_weight = np.array(pretrained_weight)#预训练的词向量
embed.weight.data.copy_(torch.from_numpy(pretrained_weight))