add --export-raw-predictions

4 files changed, 31 insertions, 6 deletions

diff --git a/bin/analyze-archive.py b/bin/analyze-archive.py
index a725833..4aab326 100755
--- a/bin/analyze-archive.py
+++ b/bin/analyze-archive.py
@@ -910,7 +910,12 @@ if __name__ == "__main__":
             lambda m: m.get_fitted()[0], xv_count
         )
     else:
-        analytic_quality = model.assess(param_model)
+        if args.export_raw_predictions:
+            analytic_quality, raw_results = model.assess(param_model, return_raw=True)
+            with open(args.export_raw_predictions, "w") as f:
+                json.dump(raw_results, f, cls=NpEncoder)
+        else:
+            analytic_quality = model.assess(param_model)
         xv_analytic_models = [model]
 
     if "tex" in show_models or "tex" in show_quality:
diff --git a/bin/analyze-kconfig.py b/bin/analyze-kconfig.py
index 1d03839..aab71e8 100755
--- a/bin/analyze-kconfig.py
+++ b/bin/analyze-kconfig.py
@@ -258,7 +258,12 @@ def main():
         )
     else:
         static_quality = model.assess(static_model)
-        analytic_quality = model.assess(param_model)
+        if args.export_raw_predictions:
+            analytic_quality, raw_results = model.assess(param_model, return_raw=True)
+            with open(args.export_raw_predictions, "w") as f:
+                json.dump(raw_results, f, cls=dfatool.utils.NpEncoder)
+        else:
+            analytic_quality = model.assess(param_model)
         xv_analytic_models = [model]
         if lut_model:
             lut_quality = model.assess(lut_model)
diff --git a/lib/cli.py b/lib/cli.py
index 3af8cc1..f988421 100644
--- a/lib/cli.py
+++ b/lib/cli.py
@@ -184,6 +184,12 @@ def add_standard_arguments(parser):
         help="Export raw cross-validation results to FILE for later analysis (e.g. to compare different modeling approaches by means of a t-test)",
     )
     parser.add_argument(
+        "--export-raw-predictions",
+        metavar="FILE",
+        type=str,
+        help="Export raw model error data (i.e., ground truth vs. model output) to FILE for later analysis (e.g. to compare different modeling approaches by means of a t-test)",
+    )
+    parser.add_argument(
         "--info",
         action="store_true",
         help="Show benchmark information (number of measurements, parameter values, ...)",
diff --git a/lib/model.py b/lib/model.py
index 227a323..558f049 100644
--- a/lib/model.py
+++ b/lib/model.py
@@ -399,7 +399,7 @@ class AnalyticModel:
 
         return model_getter, info_getter
 
-    def assess(self, model_function, ref=None):
+    def assess(self, model_function, ref=None, return_raw=False):
         """
         Calculate MAE, SMAPE, etc. of model_function for each by_name entry.
 
@@ -412,11 +412,13 @@ class AnalyticModel:
         exclusive (e.g. by performing cross validation). Otherwise,
         overfitting cannot be detected.
         """
-        detailed_results = {}
+        detailed_results = dict()
+        raw_results = dict()
         if ref is None:
             ref = self.by_name
         for name, elem in sorted(ref.items()):
-            detailed_results[name] = {}
+            detailed_results[name] = dict()
+            raw_results[name] = dict()
             for attribute in elem["attributes"]:
                 predicted_data = np.array(
                     list(
@@ -430,7 +432,14 @@ class AnalyticModel:
                 )
                 measures = regression_measures(predicted_data, elem[attribute])
                 detailed_results[name][attribute] = measures
-
+                if return_raw:
+                    raw_results[name][attribute] = {
+                        "ground_truth": list(elem[attribute]),
+                        "model_output": list(predicted_data),
+                    }
+
+        if return_raw:
+            return detailed_results, raw_results
         return detailed_results
 
     def build_dtree(