statsim-compiler 0.1.0

StatSim compiler — parse .sm models and compile to PyMC, Stan, TFP.js, Z3

statsim-compiler

Compiler for the StatSim modeling language. Parses .sm models and compiles them to multiple probabilistic programming backends.

Install

pip install statsim-compiler

Usage

from statsim_compiler import parse, compile, model

# Parse and compile to PyMC code
blocks = parse('p ~ Beta(1, 1)\nx ~ Bin(10, p) | 7\nsample(5000)')
result = compile(blocks, 'pymc')
print(result['code'])        # PyMC source code
print(result['param_names']) # ['p']

# High-level: parse + compile + get a runner
runner = model('p ~ Beta(1, 1)\nx ~ Bin(10, p) | 7\nsample(5000)')
print(runner.code)           # generated PyMC code
trace = runner.run()         # run MCMC inference (requires pymc)

Targets

Target	Output
pymc	PyMC Python code
stan	Stan code
tfpjs	TensorFlow Probability (JavaScript)
z3	Z3 constraint solver (Python)

API

• parse(source) — parse .sm source, returns a BlockList
• compile(blocks, target) — compile to target, returns dict with code, data_names, param_names, inference_mode
• compile_script(blocks, target) — compile to raw code string
• model(source, target) — parse + compile + return a runner (PyMC runner for target='pymc')

Runner API (PyMC)

runner = model(source, 'pymc')
runner.code                    # generated source
runner.run(data, **kwargs)     # run inference, returns InferenceData
runner.build(data, **kwargs)   # get pm.Model for extension

Language

One line, one concept. Math notation:

p ~ Beta(1, 1)                    # prior
x ~ Bin(10, p) | 7                # observation
delta = p_B - p_A                 # deterministic
y[i] ~ N(mu, sigma), i=1:N       # plated
sample(5000)                      # directive

See the language reference for the full spec.

License

MIT