NegLogProbLoss

class liesel.optim.NegLogProbLoss(model, split, validation_strategy='log_lik', scale=False)

Bases: LossMixin

Negative log-probability loss for Liesel models.

The training objective is the negative sum of the model log-likelihood and log-prior. During mini-batch optimization, likelihood terms are scaled through carry.batches.scaled_log_lik(...) so BatchManager can apply branch-specific scaling for multi-size observed data. Validation loss uses split.scaled_log_lik(...) for the same reason.

The model must use the standard sums of observed distribution factors and parameter priors. Weak observed variables and weak parameters contribute their likelihoods and priors like strong ones. Custom aggregate nodes or additional unclassified distribution factors require a custom Loss; supplying a manual split is not sufficient.

Parameters:

• model (Model) – Liesel model evaluated by the loss.

• split (PositionSplit | PositionSplitManager) – Train/validation/test split. Use PositionSplitManager for models with observed branches of different sample sizes.

• validation_strategy (Literal['log_lik', 'log_prob'], default: 'log_lik') – Validation objective. "log_lik" uses the scaled log-likelihood only. "log_prob" also includes the model log-prior.

• scale (bool, default: False) – If True, divide losses by the training sample size. For PositionSplitManager, the scalar is the sum of all branch-specific training sizes.

Examples

Construct a default loss for a simple observed model:

>>> import jax.numpy as jnp
>>> import liesel.model as lsl
>>> import tensorflow_probability.substrates.jax.distributions as tfd
>>> from liesel.optim import NegLogProbLoss, PositionSplit
>>> y = lsl.Var.new_obs(
...     jnp.arange(3.0),
...     lsl.Dist(tfd.Normal, loc=0.0, scale=1.0),
...     name="y",
... )
>>> model = lsl.Model([y])
>>> split = PositionSplit.from_model(model, position_keys=["y"])
>>> loss = NegLogProbLoss(model, split)
>>> loss.position([]) == {}
True
>>> repr(loss)
'NegLogProbLoss(validation_strategy=log_lik)'

Methods

loss_monitor(params, carry)	Computes validation loss.
loss_train(params, carry)	Computes full-data negative log posterior.
loss_train_batched(params, carry)	Computes mini-batch negative log posterior.
position(position_keys)	Extracts an initial optimizer position from the model.

Attributes

model	Liesel model evaluated by the loss.
scalar	Scalar multiplier applied to the negative log probability.
scale	If True, divide losses by the training sample size.
split	Train/validation/test split.
validation_strategy	Validation objective.