paddlets.models.forecasting.dl.distributions.likelihood

class Likelihood(mode: str = 'distribution')[source]

Bases: ABC

Abstract class for a distributional regression model.

Parameters: mode (str) – The default value is “distribution” for probability distributional regression, for quantile regression, set “quantiles”.

abstract output_to_params(model_output: Tensor) → Tensor[source]

Rescale model output to distribution params, to be implemented in subclasses.

Parameters: model_output (paddle.Tensor) – The output of model.
Returns: The distribution parameters respect to subclass’ distribution.
Return type: paddle.Tensor

abstract params_to_distr(distr_params: Tensor) → Distribution[source]

Construct a distribution class by distribution parameters, to be implemented in subclasses.

Parameters: distr_params (paddle.Tensor) – The parameters of distribution.
Returns: The distribution instance defined in paddle.
Return type: Distribution

sample(model_output: Tensor, num_samples: int = 1) → Tensor[source]

Samples a prediction from the model output：

1> output to distribution parameters;

2> distribution parameters to distribution;

3> sample by distribution

Parameters

model_output (paddle.Tensor) – The output of model.
num_samples (int) – The number of samples to be sampled.

Returns

The samples of distribution.

Return type

paddle.Tensor

abstract get_mean(distr_params: Tensor) → Tensor[source]

Compute mean by distribution params, to be implemented in subclasses.

Parameters: distr_params – The params of distribution.
Returns: The mean of distribution.
Return type: paddle.Tensor

abstract property num_params: int

Returns the number of parameters that define the probability distribution for one single target value.

Returns: The number of parameters.
Return type: int

loss(distr_params: Tensor, target: Tensor) → Tensor[source]

Compute NLL loss by predicted distrbution parameters and ground truth target.

This is the basic way to compute the NLL loss. It can be overwritten by likelihood for which paddle proposes a numerically better NLL loss.

Parameters

distr_params – The parameters of distribution.
target – The ground truth of the sample.

Returns

The loss computed by distribution parameters and ground truth.

Return type

paddle.Tensor

class GaussianLikelihood[source]

Bases: Likelihood

Univariate Gaussian distribution.

output_to_params(model_output: Tensor) → Tensor[source]

Use softplus to rescale sigma parameter as it should be positive.

Parameters: model_output (paddle.Tensor) – The output of model.
Returns: The Gaussian distribution parameters(mu and sigma).
Return type: paddle.Tensor

params_to_distr(distr_params: Tensor) → Distribution[source]

Construct Normal instance by parameters: mu and sigma.

Parameters: distr_params (paddle.Tensor) – Tensor of mu and sigma.
Returns: The Gaussian instance defined in paddle.
Return type: Distribution

get_mean(distr_params: Tensor) → Tensor[source]

Return mean of the distribution.

Parameters: distr_params – Tensor of parameters mu and sigma.
Returns: Mean of Gaussian distribution.
Return type: paddle.Tensor

property num_params: int

For Gaussian, the number of parameters is 2.

Returns: The number of parameters of Gaussian distribution.
Return type: int