rasmusbergpalm / nanograd

A miniscule implementation of reverse mode auto-differentiation

nanograd

A minimal implementation of reverse mode auto-differentiation in ~40 lines of python.

May be useful for educational purposes.

Example

from nanograd import Var

a = Var(3.0)
b = Var(5.0)
c = Var(9.0)
d = b * c
e = a + (-d) ** 2
f = (e / d).relu()
f.backward()

for v in [a, b, c, d, e, f]:
    print(v)

Outputs

Var(v=3.0000, grad=0.0222)
Var(v=5.0000, grad=8.9867)
Var(v=9.0000, grad=4.9926)
Var(v=45.0000, grad=0.9985)
Var(v=2028.0000, grad=0.0222)
Var(v=45.0667, grad=1.0000)

Running tests

The tests use pytorch to check the gradients.

python -m pytest