AttributeError: cannot assign module before Module.__init_

AttributeError: cannot assign module before Module.__init__() call

原因

调用了自定义的类，但是在自定义的类的__init__函数下面没有写super( XXX, self ).init()

错误案例

import torch
import torch.nn as nnclass SelfAttention(nn.Module):""" Self-Attention """def __init__(self, n_head, d_k, d_v, d_x, d_o):#super(SelfAttention, self).__init__()#SelfAttention 是类名self.wq = nn.Parameter(torch.Tensor(d_x, d_k))self.wk = nn.Parameter(torch.Tensor(d_x, d_k))self.wv = nn.Parameter(torch.Tensor(d_x, d_v))self.mha = MultiHeadAttention(n_head=n_head, d_k_=d_k, d_v_=d_v, d_k=d_k, d_v=d_v, d_o=d_o)self.init_parameters()def init_parameters(self):for param in self.parameters():stdv = 1. / np.power(param.size(-1), 0.5)param.data.uniform_(-stdv, stdv)def forward(self, x, mask=None):q = torch.matmul(x, self.wq)   k = torch.matmul(x, self.wk)v = torch.matmul(x, self.wv)attn, output = self.mha(q, k, v, mask=mask)return attn, outputif __name__ == "__main__":n_x = 4d_x = 80batch = 2x = torch.randn(batch, n_x, d_x)mask = torch.zeros(batch, n_x, n_x).bool()selfattn = SelfAttention(n_head=8, d_k=128, d_v=64, d_x=80, d_o=80)attn, output = selfattn(x, mask=mask)print(attn.size())print(output.size())

在这里插入图片描述

正确案例

import torch
import torch.nn as nnclass SelfAttention(nn.Module):""" Self-Attention """def __init__(self, n_head, d_k, d_v, d_x, d_o):super(SelfAttention, self).__init__()#SelfAttention 是类名self.wq = nn.Parameter(torch.Tensor(d_x, d_k))self.wk = nn.Parameter(torch.Tensor(d_x, d_k))self.wv = nn.Parameter(torch.Tensor(d_x, d_v))self.mha = MultiHeadAttention(n_head=n_head, d_k_=d_k, d_v_=d_v, d_k=d_k, d_v=d_v, d_o=d_o)self.init_parameters()def init_parameters(self):for param in self.parameters():stdv = 1. / np.power(param.size(-1), 0.5)param.data.uniform_(-stdv, stdv)def forward(self, x, mask=None):q = torch.matmul(x, self.wq)   k = torch.matmul(x, self.wk)v = torch.matmul(x, self.wv)attn, output = self.mha(q, k, v, mask=mask)return attn, outputif __name__ == "__main__":n_x = 4d_x = 80batch = 2x = torch.randn(batch, n_x, d_x)mask = torch.zeros(batch, n_x, n_x).bool()selfattn = SelfAttention(n_head=8, d_k=128, d_v=64, d_x=80, d_o=80)attn, output = selfattn(x, mask=mask)print(attn.size())print(output.size())