update stress training

when extracting data, stress_new = stress_old*1000

.spyproject/config/backups/workspace.ini.bak

@@ -4,7 +4,7 @@ save_data_on_exit = True
 save_history = True
 save_non_project_files = False
 project_type = 'empty-project-type'
-recent_files = []
+recent_files = ['C:\\Users\\ZHANG Jun\\.spyder-py3\\temp.py', 'agat\\model\\model.py', 'Change_log.md', 'agat\\test\\POSCAR', 'agat\\test\\test_stress_model.py']
 
 [main]
 version = 0.2.0

.spyproject/config/workspace.ini

@@ -4,7 +4,7 @@ save_data_on_exit = True
 save_history = True
 save_non_project_files = False
 project_type = 'empty-project-type'
-recent_files = ['C:\\Users\\ZHANG Jun\\.spyder-py3\\temp.py', 'agat\\test\\test_stress_model.py', 'agat\\model\\model.py', 'Change_log.md']
+recent_files = ['agat\\test\\dataloader_test.py', 'agat\\data\\build_dataset.py']
 
 [main]
 version = 0.2.0

Change_log.md

@@ -1,4 +1,6 @@
 # [main](https://github.com/jzhang-github/AGAT/tree/main)
+- when extracting data, the true stress is amplified 1000 to avoid the pytorch accuracy.
+    - Stresses of many snapshots are lower than 1e-6, which can be problematic when training with torch.float32
 
 # [v9.0.0](https://github.com/jzhang-github/AGAT/tree/v9.0.0)
 **Note: AGAT after this version (included) cannot load the well-trained model before.** If you need to do so, please use v8.0.5: https://pypi.org/project/agat/8.0.5/

agat/__init__.py

@@ -5,7 +5,7 @@
 @author: ZHANG Jun
 """
 
-__version__ = '9.0.0'
+__version__ = '9.0.1'
 
 import os
 # os.environ['DGLBACKEND']="pytorch"
@@ -14,6 +14,7 @@
 from .data.build_dataset import BuildDatabase
 from .data.build_dataset import CrystalGraph
 from .model.fit import Fit
+from .model.model import PotentialModel
 from .app.cata.high_throughput_predict import HtAds
 
 del os

agat/app/app.py

@@ -111,11 +111,14 @@ def calculate(self, atoms=None, properties=None, system_changes=['positions', 'n
         graph, info = self.cg.get_graph(atoms)
         graph = graph.to(self.device)
 
+        # Stresses of many snapshots are lower than 1e-6, which can be
+        # problematic when training with torch.float32
         with torch.no_grad():
-            energy_pred, force_pred, stress_pred = self.model.forward(graph)
+            energy_pred, force_pred, stress1000_pred = self.model.forward(graph)
 
         self.results = {'energy': energy_pred[0].item() * len(atoms),
-                        'forces': force_pred.cpu().numpy()}
+                        'forces': force_pred.cpu().numpy(),
+                        'stress': stress1000_pred.cpu().numpy() / 1000}
 
 # class GatCalculator(Calculator):
 #     """ Calculator with ASE module. Pymatgen is also needed.
@@ -186,4 +189,3 @@ def calculate(self, atoms=None, properties=None, system_changes=['positions', 'n
 # debug
 if __name__ == '__main__':
     pass
-

agat/data/build_dataset.py

@@ -277,8 +277,13 @@ def get_graph_from_ase(self, fname): # ase_atoms or file name
             energy_true = torch.tensor(np.load(fname+'_energy.npy'), dtype=self.dtype)
             graph_info['energy_true'] = energy_true
         if self.data_config['build_properties']['stress']:
-            stress_true = torch.tensor(np.load(fname+'_stress.npy'), dtype=self.dtype)
-            graph_info['stress_true'] = stress_true
+            # Stresses of many snapshots are lower than 1e-6, which can be
+            # problematic when training with torch.float32
+            stress_true = np.load(fname+'_stress.npy')
+            # graph_info['stress_true'] = torch.tensor(stress_true,
+            #                                          dtype=self.dtype)
+            graph_info['stress_true1000'] = torch.tensor(stress_true * 1000,
+                                                         dtype=self.dtype)
         if self.data_config['build_properties']['cell']:
             cell_true = torch.tensor(ase_atoms.cell.array, dtype=self.dtype)
             graph_info['cell_true'] = cell_true
@@ -354,8 +359,13 @@ def get_graph_from_pymatgen(self, crystal_fname):
             energy_true = torch.tensor(np.load(crystal_fname+'_energy.npy'), dtype=self.dtype)
             graph_info['energy_true'] = torch.tensor((energy_true), dtype=self.dtype)
         if self.data_config['build_properties']['stress']:
-            stress_true = torch.tensor(np.load(crystal_fname+'_stress.npy'), dtype=self.dtype)
-            graph_info['stress_true'] = torch.tensor((stress_true), dtype=self.dtype)
+            # Stresses of many snapshots are lower than 1e-6, which can be
+            # problematic when training with torch.float32
+            stress_true = np.load(crystal_fname+'_stress.npy')
+            # graph_info['stress_true'] = torch.tensor(stress_true,
+            #                                          dtype=self.dtype)
+            graph_info['stress_true1000'] = torch.tensor(stress_true * 1000,
+                                                         dtype=self.dtype)
         if self.data_config['build_properties']['cell']:
             cell_true = torch.tensor(mycrystal.lattice.matrix, dtype=self.dtype)
             graph_info['cell_true'] = torch.tensor((cell_true), dtype=self.dtype)
@@ -827,4 +837,3 @@ def select_graphs_random(fname: str, num: int):
 if __name__ == '__main__':
     ad = BuildDatabase(mode_of_NN='pymatgen_dist', num_of_cores=16)
     ad.build()
-

agat/default_parameters.py

@@ -81,8 +81,8 @@
     'negative_slope': 0.2,
     'criterion': nn.MSELoss(),
     'a': 1.0,
-    'b': 50.0,
-    'c': 1000.0,
+    'b': 1.0,
+    'c': 1.0,
     # 'optimizer': 'adam',
     'learning_rate': 0.0001,
     'weight_decay': 0.0, # weight decay (L2 penalty)

agat/model/__init__.py

@@ -7,4 +7,5 @@
 
 from .layer import Layer
 from .model import PotentialModel, CrystalPropertyModel, AtomicPropertyModel, AtomicVectorModel
-from .fit import Fit
+from .fit import Fit
+from agat.lib.model_lib import save_model, load_model, save_state_dict, load_state_dict

agat/model/fit.py

@@ -210,7 +210,7 @@ def fit(self, **train_config):
             for i, (graph, props) in enumerate(self.train_loader):
                 energy_true = props['energy_true']
                 force_true = graph.ndata['forces_true']
-                stress_true = props['stress_true']
+                stress_true = props['stress_true1000']
                 optimizer.zero_grad()
                 energy_pred, force_pred, stress_pred = model.forward(graph)
                 energy_loss = criterion(energy_pred, energy_true)
@@ -241,7 +241,7 @@ def fit(self, **train_config):
                 for i, (graph, props) in enumerate(self.val_loader):
                     energy_true_all.append(props['energy_true'])
                     force_true = graph.ndata['forces_true']
-                    stress_true_all.append(props['stress_true'])
+                    stress_true_all.append(props['stress_true1000'])
                     energy_pred, force_pred, stress_pred = model.forward(graph)
                     energy_pred_all.append(energy_pred)
                     if self._has_adsorbate:
@@ -326,7 +326,7 @@ def fit(self, **train_config):
             for i, (graph, props) in enumerate(self.test_loader):
                 energy_true_all.append(props['energy_true'])
                 force_true_all.append(graph.ndata['forces_true'])
-                stress_true_all.append(props['stress_true'])
+                stress_true_all.append(props['stress_true1000'])
                 energy_pred, force_pred, stress_pred = model.forward(graph)
                 energy_pred_all.append(energy_pred)
                 force_pred_all.append(force_pred)
@@ -356,11 +356,11 @@ def fit(self, **train_config):
     Epoch      : {epoch}
     Energy loss: {energy_loss.item()}
     Force_Loss : {force_loss.item()}
-    Stress_Loss: {stress_loss.item()}
+    Stress_Loss: {stress_loss.item()} (units: 1e-3 ASE stress)
     Total_Loss : {total_loss.item()}
     Energy_MAE : {energy_mae.item()}
     Force_MAE  : {force_mae.item()}
-    Stress_MAE : {stress_mae.item()}
+    Stress_MAE : {stress_mae.item()} (units: 1e-3 ASE stress)
     Energy_R   : {energy_r.item()}
     Force_R    : {force_r.item()}
     Stress_R   : {stress_r.item()}

agat/test/export_cell_stress.py

@@ -12,7 +12,7 @@
 os.environ['KMP_DUPLICATE_LIB_OK']='TRUE'
 graph_list, props = load_graphs('all_graphs_generation_0.bin')
 
-stresses = props['stress_true'].numpy()
+stresses = props['stress_true'].numpy() # props['stress_true1000']
 
 
 plt.hist(stresses[:,0], bins=100, range=[-0.01, 0.01])