add force l2 norm scaled error and max error to utils

finetuna/utils.py

@@ -11,6 +11,7 @@
 from finetuna.calcs import DeltaCalc
 import numpy as np
 from numpy.linalg import norm
+from ase.db import connect
 
 
 def convert_to_singlepoint(images):
@@ -345,6 +346,14 @@ def add_hookean_constraint(
     image.set_constraint(cons)
 
 
+def get_online_asedb_as_parent_images(db_path):
+    images = []
+    with connect(db_path) as queried_db:
+        for row in queried_db.select(check=True):
+            images.append(asedb_row_to_atoms(row))
+    return images
+
+
 def force_l2_norm_err(forces0, forces1):
     diff_array = forces0 - forces1
     l2_norm_vector = norm(diff_array, axis=1)
@@ -367,3 +376,52 @@ def force_cos_sim(forces0, forces1):
     cos_sim = dot_array / (magnitudes0 * magnitudes1)
     mean_cos_sim = np.mean(cos_sim)
     return mean_cos_sim
+
+
+def force_l2_norm_scaled_error(forces0, forces1):
+    # forces0 is ground truth
+    diff_array = forces0 - forces1
+    l2_norm_vector = norm(diff_array, axis=1)
+    forces0_norm_vector = norm(forces0, axis=1)
+    scaled_l2_norm_vector = l2_norm_vector / forces0_norm_vector
+    scaled_mae = np.mean(scaled_l2_norm_vector)
+    return (
+        np.max(scaled_l2_norm_vector),
+        scaled_mae,
+        np.mean(l2_norm_vector),
+        np.mean(forces0_norm_vector),
+        np.mean(norm(forces1, axis=1)),
+    )
+
+
+def force_l2_norm_max_err(forces0, forces1):
+    diff_array = forces0 - forces1
+    l2_norm_vector = norm(diff_array, axis=1)
+    mae = np.max(l2_norm_vector)
+    return mae
+
+
+if __name__ == "__main__":
+    from ase.io import Trajectory
+
+    i = 0
+    images = []
+    ml_images = []
+    re_images = []
+    db_path = "/home/jovyan/shared-scratch/joe/jobs/finetuna_2_tests/all_sudheesh_data/5_mae_threshold_005_reduce/Sudheesh_mof_zeolite/Cu_zeolites/MOR_T1_T4_96_141_1/finetuna/oal_queried_images.db"
+    with connect(db_path) as queried_db:
+        for row in queried_db.select(check=True):
+            i += 1
+            images.append(asedb_row_to_atoms(row, calc="parent"))
+            ml_images.append(asedb_row_to_atoms(row, calc="ml"))
+            re_images.append(asedb_row_to_atoms(row, calc="retrained"))
+            print(
+                (
+                    i,
+                    force_l2_norm_scaled_error(
+                        images[-1].get_forces(), ml_images[-1].get_forces()
+                    ),
+                )
+            )
+            if i > 100000:
+                break