Torch can't be used in backward

autograd
#1 
class RBFfun(Function): 
    def forward(self,input,weight,sigma,center):
        t=torch.mv(weight,input)
        dtc=t-center
        dist2=torch.sum(dtc*dtc)
        output=-(sigma*sigma)*dist2/2
        output=torch.exp(output)
        self.save_for_backward(input,weight,sigma,center,output)
        self.dtc=dtc
        return output
    
    def backward(self,grad_output):
        input,weight,sigma,center,output=self.saved_tensors
        dtc=self.dtc
        grad_input=grad_weight=grad_sigma=grad_center=None
        if self.needs_input_grad[0]:
            m,n=weight.size()[0],weight.size()[1]
            Hess=Variable(torch.zeros(n,m))
            for i in range(n):
                Hess[i]=-(torch.dot(dtc,weight[:,i]))*output*pow(sigma,2)
            grad_input=torch.mv(Hess,grad_output) 
               
        return grad_input,grad_weight,grad_sigma,grad_center

the error message is “mv(): argument ‘vec’ (position 1) must be Variable, not torch.FloatTensor”
but I think the grad_output is Variable

#2

Hi,

Please check the doc here on how to implement your own Function.
For example, your code is missing the @staticmethod and thus does not create a proper Function.

#3

thanks so much, I missed “@staticmethod”