better support for finding dummy columns in accelerated failure rate …

…script

deckard/layers/afr.py

@@ -675,15 +675,17 @@ def clean_data_for_aft(
         subset = subset.drop(columns=list(dummy_dict.keys()))
         cleaned = pd.concat([subset, dummies], axis=1)
     else:
-        # Find non-numeric columns
-        non_numeric = subset.select_dtypes(exclude=[np.number]).columns
-        dummy_subset = subset[non_numeric]
-        dummies = pd.get_dummies(
-            dummy_subset,
-            columns=dummy_subset.columns,
-        )
-        subset = subset.drop(columns=dummy_subset.columns)
-        cleaned = pd.concat([subset, dummies], axis=1)
+        # Assume that some categorical variables exist and need to be one-hot encoded
+        cleaned = subset.copy()
+        dummy_cols = []
+        for col in cleaned.columns:
+            if cleaned[col].dtype == "object":
+                dummy_cols.append(col)
+        dummies = pd.get_dummies(cleaned[dummy_cols], prefix="", prefix_sep="")
+        cleaned = cleaned.drop(columns=dummy_cols)
+        cleaned = pd.concat([cleaned, dummies], axis=1)
+        cleaned = cleaned.astype(float)
+        cleaned = cleaned.dropna(axis=0, how="any")
     assert (
         target in cleaned.columns
     ), f"Target {target} not in dataftame with columns {cleaned.columns}"