feat: basis transfer from small basis to larger ones by finetuning on the the eigenvalues.

dptb/nnops/trainer.py

@@ -38,11 +38,17 @@ def __init__(
         # get the task from train_datasets label
         self.task = None
         if self.train_datasets.get_Hamiltonian:
-            self.task = "hamiltonians"
+            if self.train_datasets.get_eigenvalues:
+                 self.task = "hamil_eigvals"
+            else:
+                self.task = "hamiltonians"
         elif self.train_datasets.get_DM:
             self.task = "DM"
-        else:
+        elif self.train_datasets.get_eigenvalues:
             self.task = "eigenvalues"
+        else:
+            self.task = "custom"
+            log.warning("The train data set does not have get_Hamiltonian, get_DM or get_eigenvalues set to True. please make sure this is intended.")
 
         self.use_reference = False
         if reference_datasets is not None:

dptb/utils/argcheck.py

@@ -854,8 +854,8 @@ def loss_options():
     ]
 
     eig_ham = [
-        Argument("coeff_ham", float, optional=True, default=1., doc="The coefficient of the hamiltonian penalty. Default: 1"),
-        Argument("coeff_ovp", float, optional=True, default=1., doc="The coefficient of the hamiltonian penalty. Default: 1"),
+        Argument("coeff_ham", [int, float], optional=True, default=1., doc="The coefficient of the hamiltonian penalty. Default: 1"),
+        Argument("coeff_ovp", [int, float], optional=True, default=1., doc="The coefficient of the overlap penalty. Default: 1"),
     ]
 
     skints = [
@@ -1782,4 +1782,4 @@ def normalize_skf2nnsk(data):
     base.check_value(data, strict=True)
 
     return data
-
+