OptimNaNDebugInfo

class liesel.optim.OptimNaNDebugInfo(kind, epoch, batch, obs_batch, last_non_nan_position, nan_position, loss, optimizer_index, optimizer_identifier, optimizer_position_keys, reproduction_position, reproduction_carry)

Bases: object

Reproduction data for the first NaN captured by OptimEngine.

Use reproduce_step() for NaNs introduced by an optimizer update and reproduce_loss() for NaNs returned by an optimizer’s pre-update loss or by the explicit batched loss evaluation when no optimizer is active.

The reported optimizer is where NaN was first detected. An earlier update may have produced finite values outside the model’s valid parameter domain. Inspect last_non_nan_position: it is free of NaNs, but is not necessarily a valid model state.

Methods

optimizer(engine)	Returns the optimizer associated with the captured NaN, if known.
reproduce_loss(engine)	Reproduces the captured pre-update loss.
reproduce_step(engine)	Re-runs the optimizer step that introduced a NaN position.

Attributes

position	Position most directly associated with the captured NaN event.
kind	Stage at which the first NaN was detected.
epoch	Epoch index at which the NaN was detected.
batch	Batch index at which the NaN was detected.
obs_batch	Observed batch used when the NaN was detected.
last_non_nan_position	Last position free of NaNs; it need not be a valid model state.
nan_position	Position containing the detected NaN, if available.
loss	Loss value associated with the captured event, if available.
optimizer_index	Index of the optimizer at detection, if known.
optimizer_identifier	Identifier of the optimizer at detection, if known.
optimizer_position_keys	Position keys handled by the optimizer at detection, if known.
reproduction_position	Parameter position used to reproduce the captured event.
reproduction_carry	Optimizer carry used to reproduce the captured event.