Interface fix

hippynn/databases/database.py

@@ -191,7 +191,6 @@ def make_generator(self, split_type, evaluation_mode, batch_size=None, subsample
         if not self.splitting_completed:
             raise ValueError("Database has not yet been split.")
 
-
         if split_type not in self.splits:
             raise ValueError(f"Split {split_type} Invalid. Current splits:{list(self.splits.keys())}")
 
@@ -226,6 +225,51 @@ def make_generator(self, split_type, evaluation_mode, batch_size=None, subsample
         )
 
         return generator
+
+    def trim_all_arrays(self,index):
+        """
+        To be used in conjuction with remove_high_property
+        """
+        for key in self.arr_dict:
+            self.arr_dict[key] = self.arr_dict[key][index]
+
+    def remove_high_property(self,key,perAtom,species_key=None,cut=None,std_factor=10):
+        """
+        This function removes outlier data from the dataset
+        Must be called before splitting
+        "key": the property key in the dataset to check for high values
+        "perAtom": True if the property is defined per atom in axis 1, otherwise property is treated as full system
+        "std_factor": systems with values larger than this multiplier time the standard deviation of all data will be reomved. None to skip this step
+        "cut_factor": systems with values larger than this number are reomved. None to skip this step. This step is done first. 
+        """
+        if perAtom:
+            if species_key==None:
+                raise RuntimeError("species_key must be defined to trim a per atom quantity")
+            atom_ind = self.arr_dict[species_key] > 0
+        ndim = len(self.arr_dict[key].shape)
+        if cut!=None:
+            if perAtom:
+                Kmean = np.mean(self.arr_dict[key][atom_ind])
+            else:
+                Kmean = np.mean(self.arr_dict[key])
+            failArr = np.abs(self.arr_dict[key]-Kmean)>cut
+            #This does nothing with ndim=1
+            trimArr = np.sum(failArr,axis=tuple(range(1,ndim)))==0
+            self.trim_all_arrays(trimArr)
+
+        if std_factor!=None:
+            if perAtom:
+                atom_ind = self.arr_dict[species_key] > 0
+                Kmean = np.mean(self.arr_dict[key][atom_ind])
+                std_cut = np.std(self.arr_dict[key][atom_ind]) * std_factor
+            else: 
+                Kmean = np.mean(self.arr_dict[key])
+                std_cut = np.std(self.arr_dict[key]) * std_factor
+            failArr = np.abs(self.arr_dict[key]-Kmean)>std_cut
+            #This does nothing with ndim=1
+            trimArr = np.sum(failArr,axis=tuple(range(1,ndim)))==0
+            self.trim_all_arrays(trimArr)
+
 
 
 def compute_index_mask(indices, index_pool):

hippynn/interfaces/ase_interface/calculator.py

@@ -4,7 +4,6 @@
 import warnings
 import torch
 
-from ase.calculators import interface
 from ase.calculators.calculator import compare_atoms, PropertyNotImplementedError, Calculator # Calculator is required to allow HIPNN to be used with ASE Mixing Calculators
 
 from hippynn.graphs import find_relatives, find_unique_relative, get_subgraph, copy_subgraph, replace_node, GraphModule

hippynn/interfaces/lammps_interface/mliap_interface.py

@@ -41,7 +41,7 @@ def __init__(self, energy_node, element_types, ndescriptors=1,
         # Build the calculator
         self.rcutfac, self.species_set, self.graph = setup_LAMMPS_graph(energy_node)
         self.nparams = sum(p.nelement() for p in self.graph.parameters())
-        self.graph.to(torch.float64)
+        self.graph.to(torch.float32)
 
     def compute_gradients(self, data):
         pass
@@ -62,7 +62,7 @@ def compute_forces(self, data):
         z_vals = self.species_set[elems+1]
         pair_i = self.as_tensor(data.pair_i).type(torch.int64)
         pair_j = self.as_tensor(data.pair_j).type(torch.int64)
-        rij = self.as_tensor(data.rij).type(torch.float64)
+        rij = self.as_tensor(data.rij).type(torch.float32)
         nlocal = self.as_tensor(data.nlistatoms) 
 
         # note your sign for rij might need to be +1 or -1, depending on how your implementation works

hippynn/layers/targets.py

@@ -67,7 +67,7 @@ def forward(self, all_features, mol_index, n_molecules):
             total_hier = torch.zeros_like(total_energies)
             mol_hier = torch.zeros_like(total_energies)
             total_atom_hier = torch.zeros_like(total_atomen)
-            batch_hier = torch.zeros(1,dtype=total_energies.dtype,device=total_energies.dtype)
+            batch_hier = torch.zeros(1,dtype=total_energies.dtype,device=total_energies.device)
 
         return total_energies, total_atomen, partial_sums, total_hier, total_atom_hier, mol_hier, batch_hier