A simple autograd-backprop library heavily inspired from Andrej Karpathy's micrograd. Users can build a simple neural network using the library . Tests involve comparison with pytorch's autodiff functions as well.
git clone https://github.com/0x-d15c0/revgrad.git
cd revgrad
pip install - r requirements.txta = Value(-3.0)
b = Value(1.0)
c = a + b
d = a * b + b**2
c += c + 1
c += 1 + c + (-a)
d += d * 3 + (b + a).relu()
d += 2 * d + (b - a).relu()
e = c - d
f = e**2
g = f / 2.0
g += 10.0 / f
print(f'{g.data:.4f}') # outcome of this forward pass = 162.0309
g.backward()
print(f'{a.grad:.4f}') # value of dg/da = -143.9726
print(f'{b.grad:.4f}') # value of dg/db =269.9486python test.py- Visualization using graphviz
- More activation functions