We talk offline about this but I wanted to create an issue so that we dont forget this one.
Currently, for rare cases we have tabular predictions available but not metrics (such as runtime).
from autogluon_zeroshot.repository.evaluation_repository import load
repo = load(version="BAG_D244_F10_C608_FULL")
config = "RandomForest_r6_BAG_L1"
tid = 3483
fold = 0
# prediction available in tabular predictions...
print(repo._tabular_predictions.predict(tid, 0, models=[config]))
# ... but not in results
dd = repo._zeroshot_context.df_results_by_dataset_vs_automl
print(len(dd[(dd.framework == config) & (dd.dataset == f"{tid}_{fold}")]))
# print missing tasks
xx = dd.pivot_table(index="framework", columns="dataset", values="score_val").loc["RandomForest_r6_BAG_L1"]
print(xx[xx.isna()])
We could backfill metrics (ideal), in the meantime we could also impute metrics or remove the partial tasks, currently the following tasks are not complete for "RandomForest_r6_BAG_L1:
3483_0 NaN
3583_0 NaN
359932_0 NaN
359932_8 NaN
359933_8 NaN
359944_1 NaN
359944_8 NaN
359944_9 NaN
359946_0 NaN
359946_2 NaN
359946_5 NaN
58_9 NaN
We talk offline about this but I wanted to create an issue so that we dont forget this one.
Currently, for rare cases we have tabular predictions available but not metrics (such as runtime).
We could backfill metrics (ideal), in the meantime we could also impute metrics or remove the partial tasks, currently the following tasks are not complete for "RandomForest_r6_BAG_L1: