Add PotentialMetrics

pyiron_potentialfit/atomistics/job/trainingcontainer.py

@@ -353,7 +353,7 @@ def train_test_split(self, train_size, seed=None):
         brk = int(len(self) * train_size)
         if brk in (0, 1):
             raise ValueError(
-                    f"container not large enough to realize this split, only multiples of {1/len(self)} possible!"
+                f"container not large enough to realize this split, only multiples of {1/len(self)} possible!"
             )
 
         # somewhat inefficient, but probably good enough for normal training set sizes
@@ -362,8 +362,8 @@ def train_test_split(self, train_size, seed=None):
         train_idx = idx[:brk]
         test_idx = idx[brk:]
         return (
-                self.sample(lambda f, i: i in train_idx),
-                self.sample(lambda f, i: i in test_idx),
+            self.sample(lambda f, i: i in train_idx),
+            self.sample(lambda f, i: i in test_idx),
         )
 
 
@@ -633,9 +633,15 @@ def iter(self, *arrays, wrap_atoms=True):
         """
         yield from self._container.iter(*arrays, wrap_atoms=wrap_atoms)
 
-    def train_test_split(self, train_size: float, seed=None,
-                         run: bool = True,
-                         train_name: str = None, test_name: str = None, project=None):
+    def train_test_split(
+        self,
+        train_size: float,
+        seed=None,
+        run: bool = True,
+        train_name: str = None,
+        test_name: str = None,
+        project=None,
+    ):
         """
         Split into two random sub sets for training and testing.
 

pyiron_potentialfit/ml/potentialfit.py

@@ -88,9 +88,20 @@ def predicted_data(self) -> FlattenedStorage:
     def plot(self):
         """
         Plots correlation and (training) error histograms.
+
+        See :class:`.PotentialPlots`.
         """
         return PotentialPlots(self.training_data, self.predicted_data)
 
+    @property
+    def metrics(self):
+        """
+        Calculate error (training) metrics.
+
+        See :class:`.PotentialMetrics`.
+        """
+        return PotentialMetrics(self.training_data, self.predicted_data)
+
     @abc.abstractmethod
     def get_lammps_potential(self) -> pd.DataFrame:
         """
@@ -291,3 +302,49 @@ def force_angle_histogram(
             "Angular Deviation of Force [" + ["rad", "deg"][angle_in_degrees] + "]"
         )
         plt.ylabel("Count")
+
+
+class PotentialMetrics:
+    """
+    Calculates various error metrics on training and test data.
+    """
+
+    __slots__ = ("_true_data", "_predicted_data")
+
+    def __init__(self, true_data: TrainingStorage, predicted_data: FlattenedStorage):
+        self._true_data = true_data
+        self._predicted_data = predicted_data
+
+    def _rmse(self, a, b):
+        return np.sqrt(np.mean((a - b) ** 2))
+
+    def _mae(self, a, b):
+        return np.mean(np.abs(a - b))
+
+    @property
+    def energy_rmse(self):
+        N = self._true_data["length"]
+        return self._rmse(
+            self._true_data["energy"] / N, self._predicted_data["energy"] / N
+        )
+
+    @property
+    def energy_mae(self):
+        N = self._true_data["length"]
+        return self._mae(
+            self._true_data["energy"] / N, self._predicted_data["energy"] / N
+        )
+
+    @property
+    def force_rmse(self):
+        return self._rmse(
+            np.linalg.norm(self._true_data["forces"], axis=-1),
+            np.linalg.norm(self._predicted_data["forces"], axis=-1),
+        )
+
+    @property
+    def force_mae(self):
+        return self._mae(
+            np.linalg.norm(self._true_data["forces"], axis=-1),
+            np.linalg.norm(self._predicted_data["forces"], axis=-1),
+        )

pyiron_potentialfit/spgfit/learn.py

@@ -254,7 +254,7 @@ def energy_mae(j):
     N = inpt.get_array("length")
     train = np.squeeze(inpt.get_array("energy")) / N
     pred = np.squeeze(j["output/training_efs"].to_object().get_array("energy")) / N
-    return np.abs(train-pred).mean()
+    return np.abs(train - pred).mean()
 
 
 def energy_max(j):

pyiron_potentialfit/spgfit/structures.py

@@ -857,9 +857,10 @@ def main():
         help="Retry the current step from scratch",
     )
     parser.add_argument(
-        "--export", type=str,
+        "--export",
+        type=str,
         help="Optionally specify a directory where to dump POSCAR files with the generated structures after everything "
-             "is finished."
+        "is finished.",
     )
     args = parser.parse_args()
 
@@ -892,7 +893,11 @@ def main():
             dir_path = os.path.join(args.export, cont.name)
             os.makedirs(dir_path, exist_ok=True)
             for i, s in enumerate(cont.iter_structures()):
-                s.write(os.path.join(dir_path, cont._container["identifier", i]) + ".POSCAR", format="vasp")
+                s.write(
+                    os.path.join(dir_path, cont._container["identifier", i])
+                    + ".POSCAR",
+                    format="vasp",
+                )
 
 
 if __name__ == "__main__":