cadmea 0.1.0

Symbolic parametric CAD value kernel with reverse-mode autodiff

cadmea

Cadmea is currently a symbolic parametric CAD kernel prototype. The packaged Python surface contains fixed-size symbolic values, reverse-mode autodiff, and a small optimizer layer.

Current Python API

import math

import cadmea as cd

x = cd.Parameter(0.2, name="x")
opt = cd.optim.Adam([x], lr=0.02)

for _ in range(520):
    residual = cd.sin(x) - 0.5
    loss = residual * residual
    opt.zero_grad()
    loss.backward()
    opt.step()

assert abs(math.sin(x.value[0]) - 0.5) < 1e-8

The active value model is:

• Constant: fixed scalar/vec1, vec2, vec3, or vec4 value.
• Parameter: user-owned differentiable input.
• Undefined: solver-owned unknown placeholder with optional hint.
• Expression: derived symbolic value.

Constructors infer dimension from arguments:

radius = cd.Parameter(15.0)
center = cd.Parameter(10.0, 20.0)
origin = cd.Constant(0.0, 0.0)
unknown = cd.Undefined(hint=(0.0, 0.0))
packed = cd.vec2(radius, cd.sin(radius))

Reverse-mode differentiation currently supports arithmetic, unary negation, sin, and cos. Gradients accumulate on Parameter leaves:

w = cd.Parameter(190.0)
h = cd.Parameter(240.0)
area = w * h

area.backward()
assert w.grad == (240.0,)
assert h.grad == (190.0,)

Optimizers live under cadmea.optim:

• SGD
• RMSProp
• Adam
• AdamW
• Adagrad

Optimizers mutate only Parameter values. They do not assign Undefined values and are separate from the future sketch constraint solver.