Skip to main content

A user-centered Python package for differentiable probabilistic inference

Project description

Brancher: A user-centered Python package for differentiable probabilistic inference

Brancher allows to design and train differentiable Bayesian models using stochastic variational inference. Brancher is based on the deep learning framework PyTorch.

Building probabilistic models

Probabilistic models are defined symbolically. Random variables can be created as follows:

a = NormalVariable(loc = 0., scale = 1., name = 'a')
b = NormalVariable(loc = 0., scale = 1., name = 'b')

It is possible to chain together random variables by using arithmetic and mathematical functions:

c = NormalVariable(loc = a**2 + BF.sin(b), 
                   scale = BF.exp(b), 
                   name = 'a')

In this way, it is possible to create arbitrarely complex probabilistic models. It is also possible to use all the deep learning tools of PyTorch in order to define probabilistic models with deep neural networks.

Example: Autoregressive modeling

Probabilistic model

Probabilistic models are defined symbolically:

T = 20
driving_noise = 1.
measure_noise = 0.3
x0 = NormalVariable(0., driving_noise, 'x0')
y0 = NormalVariable(x0, measure_noise, 'x0')
b = LogitNormalVariable(0.5, 1., 'b')

x = [x0]
y = [y0]
x_names = ["x0"]
y_names = ["y0"]
for t in range(1,T):
    x.append(NormalVariable(b*x[t-1], driving_noise, x_names[t]))
    y.append(NormalVariable(x[t], measure_noise, y_names[t]))
AR_model = ProbabilisticModel(x + y)

Observe data

Once the probabilistic model is define, we can decide which variable is observed:

[yt.observe(data[yt][:, 0, :]) for yt in y]

Autoregressive variational distribution

The variational distribution can have an arbitrary structure:

Qb = LogitNormalVariable(0.5, 0.5, "b", learnable=True)
logit_b_post = DeterministicVariable(0., 'logit_b_post', learnable=True)
Qx = [NormalVariable(0., 1., 'x0', learnable=True)]
Qx_mean = [DeterministicVariable(0., 'x0_mean', learnable=True)]
for t in range(1, T):
    Qx_mean.append(DeterministicVariable(0., x_names[t] + "_mean", learnable=True))
    Qx.append(NormalVariable(BF.sigmoid(logit_b_post)*Qx[t-1] + Qx_mean[t], 1., x_names[t], learnable=True))
variational_posterior = ProbabilisticModel([Qb] + Qx)


Now that the models are specified we can perform approximate inference using stochastic gradient descent:


Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

brancher-0.3.5.tar.gz (37.9 kB view hashes)

Uploaded source

Built Distribution

brancher-0.3.5-py3-none-any.whl (44.2 kB view hashes)

Uploaded py3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page