simppl is not your typical probabilistic programming language.
It does not include the latest and greatest inference methods.
In fact, it only has exhaustive/enumerative inference.
It does not have your favorite distribution; it only includes a couple of very simplified, discretized distribution.
It doesn't use that new awesome computational graph/deep learning framework, it has minimal dependencies beyond the python data stack.
Because of this, models will be dead simple and inference should hopefully be fast and straightforward.
Rather than a library you should use for production, it is a sandbox for quirky Bayesian inference ideas and for starting conversations around the theme of "Bayesian inference for users with deadlines" (with apologies to Django).
Read the blogpost here.
With pip:
pip install git+git://github.com/dimenwarper/simppl.git@master
A simple model illustrating that simppl flavor:
from simppl.distributions import SomeValue
from simppl.utils import capture_locals
from simppl.inference import Exhaustive
import numpy as np
def sigmoid_response(toxicity, log_concentration):
return 1 / (1 + np.exp(toxicity * log_concentration))
log_concentration = 1
true_toxicity = 0.6
inhibitions = (sigmoid_response(true_toxicity, log_concentration) + np.random.randn(2) * 0.001).reshape(-1, 1)
def cytotoxicity_model(inibitions=inhibitions):
toxicity = SomeValue('toxicity', between=[0, 2], around=[0, 1.5], mostly=1.3)
inhibition = SomeValue(
'inibition',
between=[0, 1],
mostly=sigmoid_response(toxicity, log_concentration),
observations=inhibitions
)
capture_locals()
return inhibition
env = Exhaustive(cytotoxicity_model)
(env
.executions
.assign(toxicity=lambda df: np.round(df['toxicity'], 2))
.groupby('toxicity')['_probability_']
.sum()
.plot.bar()
);This and other examples are in the showcase.ipynb.
️