/PyCausal

Python package for creating and sampling from structural causal models

Primary LanguagePythonMIT LicenseMIT

PyCausal - Causal Inference and Reasoning in Python

Package for defining large scale Structural Causal Models(SCMs), interventions and sampling from them.

Example

alt text

where $ N_Z, N_Y, N_X $ is the standard normal.

Code

from pycausal import *

model = SCM("Simple Causal Graph")

X = Variable("X", stats.norm(loc=0,scale=1))
Z = Variable("Z", stats.beta(0.5,0.5))

Ny = HiddenVariable("Ny", stats.norm(loc=0,scale=1))

Y = Ny * Z + exp( X**2 ) << "Y"

model.draw()

with the corresponding graphical causal model,

alt text

model.sample(2)

or

(~model)(2)

output:

{'Z': array([0.99181398, 0.02439115]), 
 'X': array([-0.07538367,  1.69771261]), 
 'Y': array([ 2.64181855, 17.87651557]) }

model.intervention({ X: 0 },2)

or

imodel = model&{ X: 0 }
imodel(2)

output:

{'X': array([0, 0]), 
 'Z': array([0.34692997, 0.16893219]),
 'Y': array([1.42016021, 0.86607793]) }

We can stack interventions aswell

imodel = model&{ X: 0, Ny: 2}
imodel(2)

is the same as

imodel = model&{X: 0}&{Ny: 2}
imodel(2)

or

imodel = model&{X: 0}
jmodel = imodel&{Ny: 2}
jmodel(2)

We can also sample specific variables instead of the full model.

Y.sample(2)

equivalently, we can write

( ~Y )(2)

output:

array([ 2.64181855, 17.87651557])

Or do independence tests(based on samples or graphical).

Y.independent_of(Ny, significance=0.05)

equivalently, we can write

Y | Ny

output:

False

We can actually define custom operations using func, and define matrix variables and assignments.

model = SCM("Matrix assignments model")


new_op = func( jax.nn.softmax, name="softmax")

X = Variable("X", stats.uniform(-2,5), shape=[2,2])
Ny = HiddenVariable("Ny",stats.beta(0.4,0.1), shape=[1,1])

y2 = -(sin(X)*2)@np.ones(shape=[2,2])
y3 = new_op(y2)

Y = reduce_sum(y3 + Ny, axis=[-1,-2], keepdims=False) << "Y"

install :~

python3 setup.py sdist bdist_wheel

pip install .