Maximizing Gaussian Likelihhood Using Pytorch

1

Hey I am trying to maximize Gaussian likelihood in PyTorch, Unfortunately getting very unreasonable results.

Start of the code

mu= Variable(torch.rand(1), requires_grad=True)
sigma= Variable(torch.rand(1), requires_grad=True)
learning_rate = 0.00002
for t in range(100000):
    NLL = -torch.sum(torch.log((1/(np.sqrt(2*3.14)*sigma))*torch.exp((x-mu)**2/(2*sigma**2))))
    NLL.backward()

    if t % 1000 == 0:
        print("loglik  =", NLL.data.numpy(),"sigma",sigma.data.numpy(),"mu",mu.data.numpy())
    
        mu.data -= learning_rate * mu.grad.data
        sigma.data -= learning_rate * sigma.grad.data
    
        mu.grad.data.zero_()
        sigma.grad.data.zero_()
        
Not sure where the problem goes wrong? Do you think this can implemented and solved using techniques like adam optimizers?