Bug when using powers?

[jcockayne]

Hi guys,

I'm seeing a strange error when I use x**2 vs. x*x in a function I'm differentiating with autograd. I'm on v1.1.4. Reproducing code is below:

def fun(x):
    return np.exp(-(x*x))
d1 = autograd.grad(fun)
d2 = autograd.grad(d1)
d3 = autograd.grad(d2)
d4 = autograd.grad(d3)
# prints -0.0, -2.0, 0.0, 12.0
print d1(0.), d2(0.), d3(0.), d4(0.)

def fun(x):
    return np.exp(-(x**2))
d1 = autograd.grad(fun)
d2 = autograd.grad(d1)
d3 = autograd.grad(d2)
d4 = autograd.grad(d3)
# prints -0.0, -2.0, nan, nan
print d1(0.), d2(0.), d3(0.), d4(0.)

Any idea why this could be happening?

Cheers!
Jon

[mattjj]

Good catch. I think it's because the formula used the call to anp.power.defgrad is not valid at zero, but I am afk and can't check right now. If that's the case then the fix would be to test if y is 0.