Gradient of re-used variable

autograd
#1

Hi all,

I’m a newbie of pytorch and I like it a lot! But I have a problem with gradient of re-used variable

here is my code:

import torch as t
from torch.autograd import Variable as v

step =2
x0 = t.FloatTensor([2, 1]).view(1, 2)
x0 = v(x0, requires_grad=True)
for _ in range(step):
x2 = t.FloatTensor([[1, 2], [1, 2]])
x2 = v(x2, requires_grad=True)
x0 = t.mm(x0, x2)
y = v(t.FloatTensor([[1, 2], [3, 4]]))

z = t.mm(x0, y)
z.backward(t.FloatTensor([[1, 1]]))
print(x2.grad)
print(x0.grad)

is gradient of x2 accumulated by every step? why gradient of x0 is None? and how can we get gradient of x0 for every step? Thanks a lot!

#2

because you redefined x0. the second x0 is not really the x0 you defined the first time. it is not really reusing thje same variable, but redefining a symbol to be a different variable. your new variable here is an intermediate results, which do not retain gradients by default. you can call retain_grad() on it to get grad http://pytorch.org/docs/master/autograd.html#torch.autograd.Variable.retain_grad.

#3

Hi SimonW,

Thanks for your reply! Now I know how to obtain gradient for every step.

But I’m still wondering, after I changed x0 for every step, when I do backward and apply gradient to variables by step(), still the original x0 (first defined x0) will be updated but just I can’t print out its gradient, right?

#4

That’s right. Just think of them as regular Python objects. The symbol name is just a pointer.

#5

You solve my confusion, thanks a lot!