pytorch中参数的无效组合

Question

I'm having some trouble to identify the error. This is the error message that I'm getting:

return torch.exp(-((x-2)**2)) + 0.8*torch.exp(-(x+2)**2)

TypeError: torch.exp received an invalid combination of arguments - got (!float!), but expected (torch.FloatTensor source)

import torch
import time
dtype = torch.FloatTensor

def functionFit(x):
    return torch.exp(-((x-2)**2)) + 0.8*torch.exp(-(x+2)**2)

def createSample(mu,sig,N):
    return  mu + sig*torch.randn(N,1).type(dtype)

def updateMU(alpha,rho,x,N,I,mu):
    return alpha*torch.mean(x[I[int((1-rho)*N):N,0]]) + (1-alpha)*mu

def updateSIG(alpha,rho,x,N,I,sig):
    return alpha*torch.std(x[I[int((1-rho)*N):N,0]])  + (1-alpha)*sig  

def CE(N,rho,alpha,epsilon,mu,sig): # initial std dev.
    start = time.time()
    k = 0
    while (sig > epsilon):
        x            = createSample(mu,sig,N)
        S            = functionFit(x)
        sorted_v , I = torch.sort(S,0)
        mu           = updateMU(alpha,rho,x,N,I,mu)
        sig          = updateSIG(alpha,rho,x,N,I,sig)
        k = k + 1

    end = time.time()
    xm  = torch.mean(x)
    ym  = functionFit(xm)
    print('x =',xm)
    print('y =',ym)
    print('time =',end - start,'s')
    print('iter =',k)


if __name__ == '__main__':
    N       = 50
    rho     = 0.5
    alpha   = 0.9
    epsilon = 0.001
    mu      = 20*torch.rand(1,1).type(dtype)-10 # init mu
    sig     = 5
    CE(N,rho,alpha,epsilon,mu,sig)

Answer 1

In the CE function, the following two lines are causing the error.

xm  = torch.mean(x)
ym  = functionFit(xm)

Here, x is a 50 x 1 FloatTensor but when you call torch.mean() , it returns a float value which is causing error when you call functionFit(xm) .

By the way, just for your information, torch.mean() returns a float value and torch.exp() expects a tensor as input. You can simply check the type of the parameter in functionFit() and compute exponent using numpy if the parameter is a float value instead of a tensor.