Retrodiff

Standalone reverse autodiff library.

Usage

Dag api

1. Subclass retrodiff.Function to create base functions (or use presets from retrodiff.utils.function).
2. Make the dag function by composing Nodes and Functions.
3. Set values on input nodes.
4. Call node.forward() and node.backward() on the output node to calulate output values and gradients.
5. Gradients and valuesa are stored in node.grad and in node.value. You can also see the full function with node.show_tree().

Example:

class Mul(Function):
    def forward(self, x, y): return x * y
    def backward(self, grad, wrt, x, y): return (y, x)[wrt] * grad

class Add(Function):
    def forward(self, x, y): return x + y
    def backward(self, grad, wrt, x, y): return grad

mul, add = Mul(), Add()
x, y, z = (Node(), Node(), Node())
f = add(mul(x, y), z)

x.value = 2
y.value = 3
z.value = 1

f.forward()
f.backward(1)

out = f.value
grads = x.grad, y.grad, z.grad, f.grad

See also the examples.

Models api

See retrodiff.utils.nn.