def forward(self, x):
B, T, N = x.size()
x = x.permute(0,2,1)
print(x.shape)
out1 = self.inception1(x)
out2 = self.inception2(x)
out3 = self.inception3(x)
out4 = self.inception4(x)
out5 = self.inception5(x)
# UEB
out_m = torch.mean(torch.stack([out1,out2,out3,out5,out4]), dim=0)
print(out_m.shape)
out_v = 1 - F.softmax(torch.var(torch.stack([out1,out2,out3,out5,out4]), dim=0, unbiased=False))
uncertainty = self.conv_u(out_v)
loss_ae = torch.mean(torch.abs(torch.stack([out1,out2,out3,out5,out4]) - out_m.unsqueeze(0).repeat(5,1,1,1)))
loss = torch.abs(torch.sum(0.5*(torch.exp((-1)*uncertainty)) * loss_ae**2 + 0.5*uncertainty))
kernel = self.linear_k((out_m * uncertainty).view(-1,T*N))
kernel = kernel[0,:].view(-1,N,self.kernel_lenth).repeat(N,1,1)
res = F.conv1d(out_m, kernel, padding = (self.kernel_lenth-1)//2)
out = out_m + res
out = out.permute(0,2,1)
return out, loss
上述代码不会报错。
def forward(self, x):
B, T, N = x.size()
x = x.permute(0,2,1)
print(x.shape)
out1 = self.inception1(x)
out2 = self.inception2(x)
out3 = self.inception3(x)
out4 = self.inception4(x)
out5 = self.inception5(x)
# UEB
out_m = torch.mean(torch.stack([out1,out2,out3,out5,out4]), dim=0)
print(out_m.shape)
out_v = 1 - F.softmax(torch.var(torch.stack([out1,out2,out3,out5,out4]), dim=0, unbiased=False))
uncertainty = self.conv_u(out_v)
loss_ae = torch.mean(torch.abs(torch.stack([out1,out2,out3,out5,out4]) - out_m.unsqueeze(0).repeat(5,1,1,1)))
loss = torch.abs(torch.sum(0.5*(torch.exp((-1)*uncertainty)) * loss_ae**2 + 0.5*uncertainty))
kernel = self.linear_k((x * uncertainty).view(-1,T*N))
kernel = kernel[0,:].view(-1,N,self.kernel_lenth).repeat(N,1,1)
res = F.conv1d(x, kernel, padding = (self.kernel_lenth-1)//2)
out = x + res
out = out.permute(0,2,1)
return out, loss
把out_m改成x会报错。
但打印出来shape是一样的。
出现这个原因主要就是因为view()需要Tensor中的元素地址是连续的，因为可能出现Tensor不连续的情况，所以修改方法为：
在.view前加.contiguous（），使其变为连续就ok。
不知道为啥x不是连续的。
加了contiguous确实好了。