sgdrf.model

Implementation of SGDRF model.

Module Contents

class sgdrf.model.SGDRF(config: sgdrf.sgdrf_config.SGDRFConfig)

Streaming Gaussian Dirichlet Random Field model.

Parameters:

config (SGDRFConfig) – SGDRF config object

xu

Sparse inducing points

Type:

torch.Tensor

dims

The number of spatiotemporal dimensions for this model

Type:

int

V

The number of observation types

Type:

int

K

The number of latent Gaussian processes

Type:

int

M

The number of inducing points

Type:

int

latent_shape

The Pyro shape of the latent Gaussian processes

Type:

tuple[int]

max_obs

The maximum number of possible simultaneous categorical observations

Type:

int

device

Pytorch device (e.g. torch.device(‘cuda’))

Type:

torch.device

dir_p

The Dirichlet hyperparameters for each entry in the word-topic matrix

Type:

torch.Tensor

jitter

Small jitter to add to covariance matrix diagonal

Type:

float

zero_loc

Tensor of all-zeros matching the inducing points

Type:

torch.Tensor

uloc

Inducing point mean variational parameter

Type:

pyro.nn.module.PyroParam

uscaletril

Inducing point lower triangular covariance Cholesky decomposition variational parameter

Type:

pyro.nn.module.PyroParam

word_topic_probs

Maximum a posteriori word-topic matrix variational parameter

Type:

pyro.nn.module.PyroParam

kernel

Gaussian process kernel

Type:

pyro.contrib.gp.kernels.Kernel

whiten

Whether the Gaussian process covariance matrix is whitened

Type:

bool

subsampler

Subsampling algorithm

Type:

Subsampler

num_particles

Number of parallel posterior latent samples to draw

Type:

int

objective

The objective function used during training

Type:

pyro.infer.ELBO

xs

All the locations of past observations

Type:

torch.Tensor

ws

All the past observations

Type:

torch.Tensor

optimizer

Stochastic gradient descent optimizer

Type:

pyro.optim.PyroOptim

svi

Stochastic variational inference helper object

Type:

pyro.infer.SVI

fail_on_nan_loss

Whether to raise an exception if training loss is NaN

Type:

bool

n_xs

The number of past observation locations

Type:

bool

Implementation of SGDRF model.

static check_inputs(xs: torch.Tensor | None = None, ws: torch.Tensor | None = None)

Check that input location and observation sizes agree in shape.

Parameters:

• xs (Optional[torch.Tensor], optional) – The input locations, by default None

• ws (Optional[torch.Tensor], optional) – The input observations, by default None

forward(xs: torch.Tensor) → torch.Tensor

Produce word probabilities at a given set of locations.

Parameters:

xs (torch.Tensor) – Locations to generate observation probabilities at

Returns:

Observation probabilities

Return type:

torch.Tensor

guide(xs: torch.Tensor, ws: torch.Tensor, subsample: torch.Tensor)

Run the stochastic variational inference approximate posterior.

Parameters:

• xs (torch.Tensor) – Locations of all observations

• ws (torch.Tensor) – Observed categorical data

• subsample (torch.Tensor) – Indices of past observations to use in this training step

model(xs: torch.Tensor, ws: torch.Tensor, subsample: torch.Tensor) → torch.Tensor

Run the stochastic variational inference prior and likelihood model.

Parameters:

• xs (torch.Tensor) – Locations of all observations

• ws (torch.Tensor) – Observed categorical data

• subsample (torch.Tensor) – Indices of past observations to use in this training step

Returns:

The observations

Return type:

torch.Tensor

process_inputs(xs: torch.Tensor | None = None, ws: torch.Tensor | None = None)

Ingest new observations.

Parameters:

• xs (Optional[torch.Tensor], optional) – New locations, by default None

• ws (Optional[torch.Tensor], optional) – New observations, by default None

step() → float

Take a single training step.

Returns:

Training loss for this step

Return type:

float

Raises:

ValueError – If self.fail_on_nan_loss is True, raise an error if loss is NaN

topic_prob(xs: torch.Tensor | None = None) → torch.Tensor

Infer or predict topic probabilities.

Parameters:

xs (Optional[torch.Tensor], optional) – The points to predict at, by default self.xs

Returns:

Topic probabilities at each point

Return type:

torch.Tensor

word_prob(xs: torch.Tensor | None = None) → torch.Tensor

Infer or predict word probabilities.

Parameters:

xs (Optional[torch.Tensor], optional) – The points to predict at, by default self.xs

Returns:

Word probabilities at each point

Return type:

torch.Tensor

word_topic_prob() → pyro.nn.module.PyroParam

Get the inferred word-topic probability matrix.

Returns:

The inferred word-topic probability matrix (MAP estimate)

Return type:

pyro.nn.module.PyroParam

sgdrf.model.EPSILON = 0.01

sgdrf.model.TORCH_DEVICE_CPU