The API documentation

class tensorprob.model.Model(name=None)

The model class is the primary interface of TensorProb. It allows you to declare random variables, describe the (directed) probabalistic relationships between them, provide observations for some of them, and perform inference on the unobserved (latent) variables.

Models are agnostic as to whether you want to follow frequentist or bayesian paradigms of inference. They allow you to find the maximum likelihood or maximum a posteriori estimate for you model given the data using the .fit method, or to sample from the likelihood/posterior using MCMC techniques (See the .mcmc method).

Random variables can only be instantiated inside the with context of a model, and each model can only have a single with block.

Inside the with context of the model, you can define variables and their relationships by telling a “generative story”. For example, defining a new variable X with X ~ Normal(0, 1) is written as X = Normal(0, 1). Random variables can then be plugged in as the conditional parameters of other distributions.

After the .initialize method is called, the model has a state for each latent variable, which is used for the initial parameters in the .fit and .mcmc methods, as well as when using the .pdf method.

Parameters:name (string, default None) – An optional name for this model. This is currently not used, but should be useful when working with multiple models simultaneously in the future.

Examples

>>> with Model() as model:
...     n = Parameter(lower=0)
...     N = Poisson(n)
... model.observed(N)
... model.initialize({ n: 10 })
... model.fit([20])

assign(assign_dict)

Set the state of specific unobserved (latent) variables to the specified values.

Parameters:assign_dict (dict) – A dictionary from random variables to values. This has to specify a value for a subset of the unobserved (latent) variables of the model.

fit(*args, **kwargs)

Perform a maximum likelihood or maximum a posteriori estimate using one of the available function optimization backends.

Parameters:

• args (lists or ndarrays) – The datasets from which we want to infer the values of unobserved (latent) variables. The arguments don’t need to have the same shape.
• use_gradient (bool) – Whether the optimizer should use gradients derived using TensorFlow. Some optimizers may not be able to use gradient information, in which case this argument is ignored.
• optimizer (subclass of BaseOptimizer) – The optimization backend to use. See the optimizers module for which optimizers are available.

initialize(assign_dict)

Allows you to specify the initial state of the unobserved (latent) variables.

Can only be called after observed variables have been declared with .observed.

Parameters:assign_dict (dict) – A dictionary from random variables to values. This has to specify a value for all unobserved (latent) variables of the model.

mcmc(*args, **kwargs)

Perform MCMC sampling of the possible values of unobserved (latent) variables using one of the available sampling backends.

Parameters:

• args (lists or ndarrays) – The datasets from which we want to infer the values of unobserved (latent) variables. The arguments don’t need to have the same shape.
• sampler (subclass of BaseSampler) – The sampling backend to use. See the samplers module for which samplers are available.

nll(*args)

The negative log-likelihood for all passed datasets.

Parameters:args (lists or ndarrays) – The datasets for which we want to know the value of the negative log-likelihood density function. The arguments don’t need to have the same shape.

observed(*args)

Declares the random variables in args as observed, which means that data is available for them.

The order in which variables are used here defines the order in which they will have to be passed in later when using methods like .fit or .mcmc. All variables in the model that are not declared as observed are automatically declared as latent and become the subject of inference.

.observed can only be called once per Model and is a requirement for calling .initialize.

Parameters:

• *args –
• random variables for which data is available. (The) –

pdf(*args_in)

The probability density function for observing a single entry of each random variable that has been declared as observed.

This allows you to easily plot the probability density function.

Parameters:args (lists or ndarrays) – The entries for which we want to know the values of the probability density function. All arguments must have the same shape.

Examples

>>> import numpy as np
>>> import matplotlib.pyplot as plt
>>> xs = np.linspace(-1, 1, 200)
>>> plt.plot(xs, model.pdf(xs))

state

The current state of every unobserved (latent) variable of the model. This is a dict from random variables to values.

Example

>>> # Assume we have a random variable X with value 42
>>> model.state[X]
42