WIP

Author: iankchristie

environment.yml

@@ -12,6 +12,7 @@ dependencies:
     - pyogrio>=0.12.1,<0.13
     - pyarrow>=22.0.0,<23
     - proj-data>=1.24,<2
+    - scikit-learn>=1.4.0,<2
     - scipy>=1.14.0,<2
     - tqdm>=4.67.1,<5
     - exactextract>=0.2.2,<1

pixi.lock

@@ -43,6 +43,7 @@ dependencies = [
   "pydantic>=2.12.5,<3",
   "pyogrio>=0.12.1,<0.13",
   "rasterio>=1.4.4,<2",
+  "scikit-learn>=1.4.0,<2",
   "scipy>=1.14.0,<2",
   "tqdm>=4.67.1,<5",
 ]

pyproject.toml

@@ -3,10 +3,10 @@
 """
 from typing import List, Literal, Optional
 
-from pydantic import model_validator, FilePath, Field
+from pydantic import FilePath, Field
 from typing_extensions import Annotated
 
-from reVeal.config.config import BaseModelStrict, BaseGridConfig
+from reVeal.config.config import BaseGridConfig
 
 
 class LearnWeightsConfig(BaseGridConfig):

reVeal/config/learn_weights.py

@@ -5,9 +5,7 @@
 samples and auto-sampled background cells as unlabeled samples. Produces a
 score-weighted configuration with feature importances normalized as weights.
 """
-import json
 import logging
-from pathlib import Path
 
 import geopandas as gpd
 import numpy as np
@@ -24,7 +22,6 @@
 def prepare_pu_data(
     grid_df,
     labels_gdf,
-    attributes,
     background_samples=10000,
     test_size=0.2,
     validation_size=0.1,
@@ -42,8 +39,6 @@ def prepare_pu_data(
         Normalized grid with score attributes and a 'gid' column.
     labels_gdf : gpd.GeoDataFrame
         Point geometries representing positive label locations.
-    attributes : list of str
-        Column names to use as features.
     background_samples : int
         Number of background (unlabeled) cells to sample.
     test_size : float
@@ -153,10 +148,28 @@ def train_pu_model(grid_df, data_splits, attributes, n_estimators=500,
     p_tr = g[g["gid"].isin(data_splits["train_gids"])][attributes].to_numpy()
     u_tr = g[g["gid"].isin(data_splits["background_gids"])][attributes].to_numpy()
 
+    if len(p_tr) == 0:
+        raise ValueError(
+            "No positive training samples found. Ensure that the labels "
+            "spatially intersect the grid."
+        )
+    if len(u_tr) == 0:
+        raise ValueError(
+            "No background (unlabeled) training samples found. Ensure that "
+            "the grid has cells that do not intersect the labels."
+        )
+
     if class_prior is None:
         class_prior = len(p_tr) / (len(p_tr) + len(u_tr))
         logger.info(f"Auto-computed class_prior: {class_prior:.4f}")
 
+    if not (0 < class_prior < 1):
+        raise ValueError(
+            f"class_prior must be between 0 and 1 (exclusive), got {class_prior}. "
+            "This can happen when all grid cells are labeled as positive or "
+            "no positive samples are found."
+        )
+
     # Train model
     model = PUExtraTrees(
         n_estimators=n_estimators,
@@ -185,7 +198,8 @@ def train_pu_model(grid_df, data_splits, attributes, n_estimators=500,
             np.zeros(len(u_test), dtype=int),
         ))
         y_pred = (model.predict(x_test) == 1).astype(int)
-        metrics["auc"] = float(roc_auc_score(y_test, y_pred))
+        y_score = model.predict_proba(x_test)
+        metrics["auc"] = float(roc_auc_score(y_test, y_score))
         tpr = y_pred[y_test == 1].sum() / y_test.sum()
         metrics["tpr"] = float(tpr)
         logger.info(f"Test metrics: AUC={metrics['auc']:.4f}, TPR={metrics['tpr']:.4f}")
@@ -254,7 +268,8 @@ def _compute_metric(grid_df, data_splits, attributes, class_prior, n_estimators,
     y_pred = (model.predict(x_val) == 1).astype(int)
 
     if metric == "auc":
-        return float(roc_auc_score(y_val, y_pred))
+        y_score = model.predict_proba(x_val)
+        return float(roc_auc_score(y_val, y_score))
     elif metric == "tpr":
         return float(y_pred[y_val == 1].sum() / y_val.sum())
     else:
@@ -442,7 +457,6 @@ def run_learn_weights(config):
     data_splits = prepare_pu_data(
         grid_df=grid_df,
         labels_gdf=labels_gdf,
-        attributes=attributes,
         background_samples=config.background_samples,
         test_size=config.test_size,
         validation_size=config.validation_size,

reVeal/learn_weights.py

@@ -2,7 +2,6 @@
 """
 Tests for learn_weights module.
 """
-import json
 import numpy as np
 import geopandas as gpd
 import pytest
@@ -54,11 +53,9 @@ class TestPrepareData:
 
     def test_basic_splits(self, synthetic_grid, synthetic_labels):
         """Test that data splits are created with correct sizes."""
-        attributes = ["feature_a_score", "feature_b_score", "feature_c_score"]
         splits = prepare_pu_data(
             grid_df=synthetic_grid,
             labels_gdf=synthetic_labels,
-            attributes=attributes,
             background_samples=50,
             test_size=0.2,
             validation_size=0.1,
@@ -77,11 +74,9 @@ def test_basic_splits(self, synthetic_grid, synthetic_labels):
 
     def test_background_size_capped(self, synthetic_grid, synthetic_labels):
         """Test that background sampling respects available cells."""
-        attributes = ["feature_a_score", "feature_b_score", "feature_c_score"]
         splits = prepare_pu_data(
             grid_df=synthetic_grid,
             labels_gdf=synthetic_labels,
-            attributes=attributes,
             background_samples=1000,  # More than available
             random_state=42,
         )
@@ -98,7 +93,6 @@ def test_train_produces_importances(self, synthetic_grid, synthetic_labels):
         splits = prepare_pu_data(
             grid_df=synthetic_grid,
             labels_gdf=synthetic_labels,
-            attributes=attributes,
             background_samples=50,
             random_state=42,
         )
@@ -115,6 +109,65 @@ def test_train_produces_importances(self, synthetic_grid, synthetic_labels):
         assert len(results["feature_importances"]) == 3
         assert results["model"] is not None
 
+    def test_empty_positives_raises(self, synthetic_grid):
+        """Test that empty positive training set raises ValueError."""
+        attributes = ["feature_a_score", "feature_b_score", "feature_c_score"]
+        splits = {
+            "train_gids": np.array([]),
+            "validation_gids": np.array([]),
+            "test_gids": np.array([]),
+            "background_gids": np.array([0, 1, 2, 3, 4]),
+            "background_test_gids": np.array([]),
+            "background_val_gids": np.array([]),
+        }
+        with pytest.raises(ValueError, match="No positive training samples"):
+            train_pu_model(
+                grid_df=synthetic_grid,
+                data_splits=splits,
+                attributes=attributes,
+                n_estimators=10,
+                random_state=42,
+            )
+
+    def test_empty_background_raises(self, synthetic_grid):
+        """Test that empty background training set raises ValueError."""
+        attributes = ["feature_a_score", "feature_b_score", "feature_c_score"]
+        splits = {
+            "train_gids": np.array([0, 1, 2, 3, 4]),
+            "validation_gids": np.array([]),
+            "test_gids": np.array([]),
+            "background_gids": np.array([]),
+            "background_test_gids": np.array([]),
+            "background_val_gids": np.array([]),
+        }
+        with pytest.raises(ValueError, match="No background"):
+            train_pu_model(
+                grid_df=synthetic_grid,
+                data_splits=splits,
+                attributes=attributes,
+                n_estimators=10,
+                random_state=42,
+            )
+
+    def test_invalid_class_prior_raises(self, synthetic_grid, synthetic_labels):
+        """Test that class_prior outside (0,1) raises ValueError."""
+        attributes = ["feature_a_score", "feature_b_score", "feature_c_score"]
+        splits = prepare_pu_data(
+            grid_df=synthetic_grid,
+            labels_gdf=synthetic_labels,
+            background_samples=50,
+            random_state=42,
+        )
+        with pytest.raises(ValueError, match="class_prior must be between"):
+            train_pu_model(
+                grid_df=synthetic_grid,
+                data_splits=splits,
+                attributes=attributes,
+                n_estimators=10,
+                class_prior=1.0,
+                random_state=42,
+            )
+
 
 class TestImportancesToWeights:
     """Tests for importances_to_weights."""
@@ -195,7 +248,6 @@ def test_full_pipeline(self, synthetic_grid, synthetic_labels):
         splits = prepare_pu_data(
             grid_df=synthetic_grid,
             labels_gdf=synthetic_labels,
-            attributes=attributes,
             background_samples=50,
             random_state=42,
         )
@@ -235,7 +287,6 @@ def test_returns_best_prior(self, synthetic_grid, synthetic_labels):
         splits = prepare_pu_data(
             grid_df=synthetic_grid,
             labels_gdf=synthetic_labels,
-            attributes=attributes,
             background_samples=50,
             random_state=42,
         )
@@ -262,7 +313,6 @@ def test_tpr_metric(self, synthetic_grid, synthetic_labels):
         splits = prepare_pu_data(
             grid_df=synthetic_grid,
             labels_gdf=synthetic_labels,
-            attributes=attributes,
             background_samples=50,
             random_state=42,
         )
@@ -286,7 +336,6 @@ def test_tuned_prior_used_in_training(self, synthetic_grid, synthetic_labels):
         splits = prepare_pu_data(
             grid_df=synthetic_grid,
             labels_gdf=synthetic_labels,
-            attributes=attributes,
             background_samples=50,
             random_state=42,
         )