Vector hierarchical probabilistic drawing.
Vector: specify coordinates, not pixels.
Hierarchical: each object can be used as a reference frame for others.
Probabilistic: no need to specify everything exactly, distribution is enough.
Pyro is used to compute log-probabilities given the configuration of coordinates. But the inference itself runs Adam (for quick convergence close to the optimum) + Metropollis-Hastings (for randomness), because when I tried using Pyro's MCMC it had very low acceptance rates and therefore slow convergence.
Put your code into vermicelly.py. Running it will compile your code into Python + Pyro and inject it into prob.py, which is just a template. The result is a.py, which actually runs the rendering (= inference).
Code for a simple model:
// anchors
OX_len ~ N(90.0, 1.0)
X := -OX_len * v(0, 1)
M ~ N(v(10, 55), v(2, 3))
// nose
n1 ~ N(v(0, 20), v(5, 5))
n2 ~ N(v(40, 0), v(4, 4))
n2s := MIRROR_X @ n2
n3 ~ N(v(30, -20), v(4, 4))
n3s := MIRROR_X @ n3
// lips
Mmb ~ N(0.2 * M, v(2, 2))
Mma ~ N(0.8 * PERP_LEFT @ M, v(5, 5))
ma := M + Mma
mas := M - Mma
mb := M + Mmb
mbs := M - MmbOther details you can find and/or modify directly in the code.
