ViX

ViX is a framework setup to run probabilistic programs on the edge. Currently, there are two ways to use this framework:

a) PyVix: A Python framework which generates c++ code that can be run both on the cpu and the arduino.

b) Cpp interface: The header files can directly be used to write the probabilistic models as well, as they are in Benchmarks/

PyViX Syntax and Example

Workflow summary: PyVix is a python interface that generates the C++17 and Arduino code (compatible with ARM processors) with the appropriate interval analysis of the distributions, evidence lower bound and scores.

Linear Regression Model

    mu_0, sigma_0 = Variable(), Variable(init = 2, range=[0.1,5])
    a0 = Uniform(0, 30, name = 'a0',learnable_params=[mu_0, sigma_0]) # Each name should be unique
    data = Data(data_X, name = 'data')
    Y_obs_data = Data(data_Y, name = 'Y_obs_data')
    Y = Observed(dist = Normal(a0 * data , 1, name = 'Y'), data = Y_obs_data, name = 'Y')
    M = Model([Y, a0, data], params = parameters)

Defining learnable parameters

mu_0, sigma_0 = Variable(), Variable(init = 2, range=[0.1,5])

First, we define the parameters of the posterior using Variable() or Constant(). ViX only supports unimodal gaussian as a posterior so the parameters in this case are mean and variance. We can also provide an initial value for each parameter and a range if it suits the application.

Defining Prior

a0 = Uniform(0, 30, name = 'a0',learnable_params=[mu_0, sigma_0])

Currently, ViX supports Uniform and Normal distributions as priors.

Input Data

data = Data(data_X, name = 'data')
Y_obs_data = Data(data_Y, name = 'Y_obs_data')

Input and Observed data is wrapped around the Data() object.

Observed Variable

Y = Observed(dist = Normal(a0 * data , 1, name = 'Y'), data = Y_obs_data, name = 'Y')

The observed variable is defined using Observed() object.

Build Model and Generate Code

M = Model([Y, a0, data], params = parameters)
gen_code(M, file_name = "test_1v", type = "fixed") #type = fixed/float/double

The priors, data objects and observed variable need to be passed as a list to Model(). Model() object is passed to gen_code. When type is set to "fixed" the PyVix runs the interval analysis to generate a fixed point configuration that supports the program.

Warnings

The x86 version of the code will print warnings about potential overflows.

If the analysis generates a fixed point configuration not supported by the fixed point library then it will generate a warning and resort to a default value for the fractional bits of 12 and estimate K and I accordingly.

Notes:

Benchmarks/run_benchmarks.py can be used to easily run all experiments, and generate the arduino code from the cpp files.