nemos.callbacks.TrainingContext#

class nemos.callbacks.TrainingContext(model=None, solver=None, params=None, state=None, aux=None, pass_idx=None, batch_idx=None, n_passes=0, frozen=None)[source]#

Bases: object

Mutable context object passed to callbacks during training.

One instance is created per training run. Fields are updated in-place by the training loop before each callback invocation.

Parameters:

model (Any) – The model being trained (e.g. GLM instance). Set by stochastic_fit.
solver (Any) – The solver instance running the optimization.
params (Any) – Current model parameters. Exposed as a read/write property: the training loop assigns the actively optimized subtree (ctx.params = ...), and reading ctx.params recombines the frozen subtree (see frozen) so callbacks always see the complete parameters.
state (Any) – Current solver state.
aux (Any) – Auxiliary output from the last batch.
pass_idx (int | None) – Current pass index (0-based).
batch_idx (int | None) – Current batch index within the pass.
n_passes (int) – Total number of passes requested.
frozen (Any) – Parameter subtree held fixed during optimization (e.g. a zero intercept when fit_intercept=False). The solver optimizes only the active subtree; this is recombined with it so callbacks always see the complete parameters. None when nothing is frozen.

Attributes

`params`	Current parameters, with the frozen subtree recombined into the active one.
`should_stop`	Whether a callback has requested early stopping.
`stop_reason`	Reason for the stop request, if any.

__init__(model=None, solver=None, params=None, state=None, aux=None, pass_idx=None, batch_idx=None, n_passes=0, frozen=None)[source]#

Parameters:

model (Any)
solver (Any)
params (Any)
state (Any)
aux (Any)
pass_idx (int | None)
batch_idx (int | None)
n_passes (int)
frozen (Any)

Methods

`__init__`([model, solver, params, state, ...])
`request_stop`([reason])	Request early stopping of the training loop.
`to_summary`()	Create a post-fit summary from the current training context.

__repr__(N_CHAR_MAX=700)[source]#

Represent this context as a string.

Simple string representation, similar to that of a dataclass.

Return type:: str
Parameters:: N_CHAR_MAX (int)

property params: Any#: Current parameters, with the frozen subtree recombined into the active one.

request_stop(reason='')[source]#

Request early stopping of the training loop.

Parameters:: reason (str) – Human-readable reason for stopping.
Return type:: None

property should_stop: bool#: Whether a callback has requested early stopping.

property stop_reason: str#: Reason for the stop request, if any.

to_summary()[source]#

Create a post-fit summary from the current training context.

Return type:: StochasticFitSummary