upgini 1.2.159a4156.dev2__tar.gz → 1.2.159a4156.dev3__tar.gz

{upgini-1.2.159a4156.dev2 → upgini-1.2.159a4156.dev3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: upgini
-Version: 1.2.159a4156.dev2
+Version: 1.2.159a4156.dev3
 Summary: Intelligent data search & enrichment for Machine Learning
 Project-URL: Bug Reports, https://github.com/upgini/upgini/issues
 Project-URL: Homepage, https://upgini.com/

upgini-1.2.159a4156.dev3/src/upgini/__about__.py

@@ -0,0 +1 @@
+__version__ = "1.2.159.a4156.dev3"

{upgini-1.2.159a4156.dev2 → upgini-1.2.159a4156.dev3}/src/upgini/features_enricher.py

@@ -2017,6 +2017,18 @@ class FeaturesEnricher(TransformerMixin):
         resolved, _ = self._resolve_client_features_in_sampled(etalon_columns, columns_renaming, available_columns)
         return resolved
 
+    @staticmethod
+    def _column_name_aliases(names: list[str] | set[str] | tuple[str, ...], columns_renaming: dict[str, str]) -> set[str]:
+        """Return names in both hashed and original forms for exclusion matching."""
+        hashed_to_original = dict(columns_renaming)
+        original_to_hashed = {original: hashed for hashed, original in columns_renaming.items()}
+        aliases: set[str] = set()
+        for name in names:
+            aliases.add(name)
+            aliases.add(hashed_to_original.get(name, name))
+            aliases.add(original_to_hashed.get(name, name))
+        return aliases
+
     def _get_cached_enriched_data(
         self,
         X: pd.DataFrame | pd.Series | np.ndarray | None = None,
@@ -2066,15 +2078,18 @@ class FeaturesEnricher(TransformerMixin):
         file_meta = self._search_task.get_file_metadata(self._get_trace_id())
         fit_dropped_features = list(self.fit_dropped_features or file_meta.droppedColumns or [])
         renamed_to_original = dict(columns_renaming)
-        original_to_renamed = {original: hashed for hashed, original in columns_renaming.items()}
-        excluding_search_keys_original = [renamed_to_original.get(sk, sk) for sk in excluding_search_keys]
 
-        excluded_client_columns = (
-            excluding_search_keys_original
+        # Exclude search keys / system cols in both hashed and original name forms.
+        excluded_client_columns = self._column_name_aliases(
+            list(excluding_search_keys)
             + fit_dropped_features
-            + [DateTimeConverter.DATETIME_COL, SYSTEM_RECORD_ID, ENTITY_SYSTEM_RECORD_ID]
+            + [DateTimeConverter.DATETIME_COL, SYSTEM_RECORD_ID, ENTITY_SYSTEM_RECORD_ID],
+            columns_renaming,
+        )
+        # Baseline should not use search-key-derived generated features (e.g. datetime cyclical).
+        excluded_baseline_columns = excluded_client_columns | self._column_name_aliases(
+            generated_features, columns_renaming
         )
-        excluded_client_columns_renamed = {original_to_renamed.get(c, c) for c in excluded_client_columns}
 
         # Client columns for enriched metrics (respects select_features).
         client_features = [
@@ -2104,11 +2119,10 @@ class FeaturesEnricher(TransformerMixin):
             if columns_renaming.get(c, c) in enriched_X_sorted.columns
         ]
 
-        # Baseline always uses all etalon columns available in the sample (ignores select_features).
+        etalon_columns_in_sampled = self._get_etalon_columns_renamed(columns_renaming, X_sorted.columns)
+        # Baseline: all etalon columns except search keys and generated (cyclical) features.
         baseline_client_features_in_sampled = [
-            c
-            for c in self._get_etalon_columns_renamed(columns_renaming, X_sorted.columns)
-            if c not in excluded_client_columns_renamed
+            c for c in etalon_columns_in_sampled if c not in excluded_baseline_columns
         ]
         self.logger.info(f"Baseline etalon columns for metrics: {baseline_client_features_in_sampled}")
         if self.fit_select_features:
@@ -2116,7 +2130,10 @@ class FeaturesEnricher(TransformerMixin):
                 client_features, columns_renaming, X_sorted.columns
             )
         else:
-            enriched_client_features_in_sampled = baseline_client_features_in_sampled
+            # Enriched may still use generated features; baseline does not.
+            enriched_client_features_in_sampled = [
+                c for c in etalon_columns_in_sampled if c not in excluded_client_columns
+            ]
             missing_client_features = set()
         if missing_client_features:
             self.logger.warning(

{upgini-1.2.159a4156.dev2 → upgini-1.2.159a4156.dev3}/src/upgini/utils/sample_utils.py

@@ -9,7 +9,6 @@ import pandas as pd
 from upgini.metadata import (
     EVAL_SET_INDEX,
     SYSTEM_RECORD_ID,
-    TARGET,
     CVType,
     ModelTaskType,
 )
@@ -106,49 +105,49 @@ def sample(
             **kwargs,
         )
 
-    # Combine balanced train with eval sets
-    if len(eval_sets_dfs) > 0:
-        df = pd.concat([train_df] + eval_sets_dfs, ignore_index=False)
-    else:
-        df = train_df
+    # Downsample train and each eval set independently (same cap as former OOT path)
+    parts: List[tuple[str, pd.DataFrame]] = [("Train", train_df)]
+    for eval_df in eval_sets_dfs:
+        eval_idx = eval_df[EVAL_SET_INDEX].iloc[0]
+        parts.append((f"Eval set {int(eval_idx)}", eval_df))
+
+    sampled_parts = [
+        _threshold_sample(
+            part_df,
+            fit_sample_threshold,
+            fit_sample_rows,
+            random_state,
+            logger,
+            part_name,
+        )
+        for part_name, part_df in parts
+    ]
+    df = pd.concat(sampled_parts, ignore_index=False) if len(sampled_parts) > 1 else sampled_parts[0]
+
+    if logger is not None:
+        logger.info(f"Dataset size after downsampling: {len(df)}")
+
+    return df
 
-    # separate OOT
-    oot_dfs = []
-    other_dfs = []
-    if EVAL_SET_INDEX in df.columns:
-        for eval_set_index in df[EVAL_SET_INDEX].unique():
-            eval_df = df[df[EVAL_SET_INDEX] == eval_set_index]
-            if TARGET in eval_df.columns and eval_df[TARGET].isna().all():
-                oot_dfs.append(eval_df)
-            else:
-                other_dfs.append(eval_df)
-    if len(oot_dfs) > 0:
-        oot_df = pd.concat(oot_dfs, ignore_index=False)
-        df = pd.concat(other_dfs, ignore_index=False)
-    else:
-        oot_df = None
 
+def _threshold_sample(
+    df: pd.DataFrame,
+    fit_sample_threshold: int,
+    fit_sample_rows: int,
+    random_state: int,
+    logger: Optional[logging.Logger],
+    part_name: str,
+) -> pd.DataFrame:
     num_samples = _num_samples(df)
     if num_samples > fit_sample_threshold:
-        logger.info(
-            f"Etalon has size {num_samples} more than threshold {fit_sample_threshold} "
-            f"and will be downsampled to {fit_sample_rows}"
-        )
-        df = df.sample(n=fit_sample_rows, random_state=random_state)
-        logger.info(f"Shape after threshold resampling: {df.shape}")
-
-    if oot_df is not None:
-        num_samples_oot = _num_samples(oot_df)
-        if num_samples_oot > fit_sample_threshold:
+        if logger is not None:
             logger.info(
-                f"OOT has size {num_samples_oot} more than threshold {fit_sample_threshold} "
+                f"{part_name} has size {num_samples} more than threshold {fit_sample_threshold} "
                 f"and will be downsampled to {fit_sample_rows}"
             )
-            oot_df = oot_df.sample(n=fit_sample_rows, random_state=random_state)
-        df = pd.concat([df, oot_df], ignore_index=False)
-
-    logger.info(f"Dataset size after downsampling: {len(df)}")
-
+        df = df.sample(n=fit_sample_rows, random_state=random_state)
+        if logger is not None:
+            logger.info(f"Shape after threshold resampling ({part_name}): {df.shape}")
     return df
 
 

upgini-1.2.159a4156.dev2/src/upgini/__about__.py

@@ -1 +0,0 @@
-__version__ = "1.2.159.a4156.dev2"