Load results in process.

benchmarks/results/process.py

@@ -27,9 +27,18 @@ def process():
     filter_params = ["tol", "seed", "max_epochs", "lr", "batch_size"]
 
     data = {}
+
+    # Load existing data
+    try:
+        with open("results/data.json", "r") as f:
+            data = json.load(f)
+    except FileNotFoundError:
+        pass
+
     benchmarks = sorted(df["benchmark"].unique())
     for bm in benchmarks:
-        data[bm] = []
+        if bm not in data:
+            data[bm] = []
 
         df_benchmark = df[df["benchmark"] == bm]
         operators = sorted(df_benchmark["operator"].unique())
@@ -40,6 +49,22 @@ def process():
             # Get the best operator run for this benchmark
             best = operator_data.sort_values("loss/test").iloc[0].to_dict()
 
+            # Check if the operator already exists with a better loss
+            existing_operators = [v["operator"] for v in data[bm]]
+            if operator in existing_operators:
+                old_best_value = float(
+                    [v["loss/test"] for v in data[bm] if v["operator"] == operator][0]
+                )
+                if best["loss/test"] >= old_best_value:
+                    print(
+                        f"Skipping {bm} {operator} because it has a better loss already: "
+                        f"old={old_best_value:.4e}  new={best['loss/test']:.4e}"
+                    )
+                    continue
+                else:
+                    # Remove the old operator
+                    data[bm] = [v for v in data[bm] if v["operator"] != operator]
+
             # Filter out parameters
             best["params"] = {
                 k: v for k, v in best["params"].items() if k not in filter_params