WandbLoggerCallback

Log metrics + benchmark predictions to Weights & Biases.

One W&B run per arm: opened in on_run_start (so it exists before the first step) and closed in on_run_end. wandb is imported lazily - if it isn't installed the callback warns once and every hook no-ops (training is never affected; :func:dispatch also isolates errors). The run URL is surfaced on ctx.extras['wandb_url'] so a downstream logger (evsys_logger) can record it on the dashboard run.

Attributes

attributenamestr

= name

attributeConfigtype

= WandbLoggerConfig

attributeproject

= project

attributeentity

= entity

attributemode

= mode

attributelog_every

= max(1, int(log_every))

attributemax_pred_rows

= max(0, int(max_pred_rows))

Functions

func__init__(self, *, project=None, entity=None, name=None, mode='online', log_every=1, max_pred_rows=100) -> None

Source Code

paramself

paramprojectstr | None

= None

paramentitystr | None

= None

paramnamestr | None

= None

parammodestr

= 'online'

paramlog_everyint

= 1

parammax_pred_rowsint

= 100

Returns

None

func_lazy_wandb(self) -> Any

Source Code

paramself

Returns

typing.Any

funcon_run_start(self, ctx) -> None

Source Code

paramself

paramctxLogContext

Returns

None

funcon_step_end(self, state, step_idx, batch, metrics) -> None

Source Code

paramself

paramstateLoopState

paramstep_idx

parambatch

parammetrics

Returns

None

funcon_eval(self, state, step_idx, eval_name, metrics) -> None

Source Code

paramself

paramstateLoopState

paramstep_idx

parameval_name

parammetrics

Returns

None

funcon_benchmark_eval(self, ctx, eval_result, predictions, *, step=None) -> None

Source Code

paramself

paramctx

parameval_result

parampredictions

paramstep

= None

Returns

None

funcon_run_end(self, ctx, run_result, arm) -> None

Source Code

paramself

paramctx

paramrun_result

paramarm

Returns

None