1 reporting.wandb.history

reporting.wandb.history

Load sampled histories and compute per-run dynamics without mutating inputs.

1.1 Classes

Name Description
MixedHistoryCoordinateKindsError Reject comparisons whose observations do not share one coordinate domain.

1.1.1 MixedHistoryCoordinateKindsError

reporting.wandb.history.MixedHistoryCoordinateKindsError(kinds)

Reject comparisons whose observations do not share one coordinate domain.

1.2 Functions

Name Description
build_dynamics_dataframe Compute per-run dynamics summaries for selected base metrics.
load_run_histories Fetch history dataframes for each run (by id).
prepare_history_long_dataframe Build a long-form dataframe for plotting metric curves.

1.2.1 build_dynamics_dataframe

reporting.wandb.history.build_dynamics_dataframe(
    runs,
    histories,
    *,
    base_metrics=None,
    prefer_x_keys=None,
    segment_frac=0.2,
)

Compute per-run dynamics summaries for selected base metrics.

Returns a new dataframe suitable for method chaining. x_kind records the exact coordinate domain (optimizer, logging, epoch, source row, or a named custom key); unlike kinds cannot enter one returned comparison.

1.2.2 load_run_histories

reporting.wandb.history.load_run_histories(
    runs,
    *,
    keys=None,
    max_rows=2000,
    replace_inf=True,
)

Fetch history dataframes for each run (by id).

1.2.3 prepare_history_long_dataframe

reporting.wandb.history.prepare_history_long_dataframe(
    histories,
    *,
    metric,
    prefer_x_keys=None,
    run_name_map=None,
)

Build a long-form dataframe for plotting metric curves.

1.2.3.1 Parameters

Name Type Description Default
histories Mapping[str, 'pd.DataFrame'] Mapping of run_id -> history dataframe. required
metric str Metric column name to extract. required
prefer_x_keys Sequence[str] | None Preferred x-axis keys (defaults to WANDB_STEP_KEYS). None
run_name_map Mapping[str, str] | None Optional map of run_id -> display name. None

1.2.3.2 Returns

Name Type Description
'pd.DataFrame' Long-form dataframe with run identity, coordinate kind, source-row
'pd.DataFrame' ordinal, selected coordinate in step, and metric value. Rows
'pd.DataFrame' retain acquired source order, including duplicate coordinates.

1.2.3.3 Raises

Name Type Description
ValueError If metric-bearing runs select unlike coordinate kinds.
Back to top