[WIP] Parallel decoupled Kalman filter training

examples/input.nn.recommended

@@ -35,6 +35,7 @@ parallel_mode                   0              # Training parallelization used (
 jacobian_mode                   1              # Jacobian computation mode (0 = Summation to single gradient, 1 = Per-task summed gradient, 2 = Full Jacobian).
 update_strategy                 0              # Update strategy (0 = Combined, 1 = Per-element).
 selection_mode                  2              # Update candidate selection mode (0 = Random, 1 = Sort, 2 = Threshold).
+decoupling_type                 0              # Kalman filter decoupling type (0 = GEKF, 1 = ED-GEKF, 2 = Layer, 3 = NDEKF, 4 = FDEKF).
 task_batch_size_energy          1              # Number of energy update candidates prepared per task for each update (0 = Entire training set).
 task_batch_size_force           1              # Number of force update candidates prepared per task for each update (0 = Entire training set).
 memorize_symfunc_results                       # Keep symmetry function results in memory.

examples/nnp-train/Cu2S_PBE/input.nn

@@ -44,6 +44,7 @@ parallel_mode                   0              # Training parallelization used (
 jacobian_mode                   1              # Jacobian computation mode (0 = Summation to single gradient, 1 = Per-task summed gradient, 2 = Full Jacobian).
 update_strategy                 0              # Update strategy (0 = Combined, 1 = Per-element).
 selection_mode                  2              # Update candidate selection mode (0 = Random, 1 = Sort, 2 = Threshold).
+decoupling_type                 0              # Kalman filter decoupling type (0 = GEKF, 1 = ED-GEKF, 2 = Layer, 3 = NDEKF, 4 = FDEKF).
 task_batch_size_energy          1              # Number of energy update candidates prepared per task for each update (0 = Entire training set).
 task_batch_size_force           1              # Number of force update candidates prepared per task for each update (0 = Entire training set).
 memorize_symfunc_results                       # Keep symmetry function results in memory.

examples/nnp-train/H2O_RPBE-D3/input.nn

@@ -44,6 +44,7 @@ parallel_mode                   0              # Training parallelization used (
 jacobian_mode                   1              # Jacobian computation mode (0 = Summation to single gradient, 1 = Per-task summed gradient, 2 = Full Jacobian).
 update_strategy                 0              # Update strategy (0 = Combined, 1 = Per-element).
 selection_mode                  2              # Update candidate selection mode (0 = Random, 1 = Sort, 2 = Threshold).
+decoupling_type                 0              # Kalman filter decoupling type (0 = GEKF, 1 = ED-GEKF, 2 = Layer, 3 = NDEKF, 4 = FDEKF).
 task_batch_size_energy          1              # Number of energy update candidates prepared per task for each update (0 = Entire training set).
 task_batch_size_force           1              # Number of force update candidates prepared per task for each update (0 = Entire training set).
 memorize_symfunc_results                       # Keep symmetry function results in memory.

examples/nnp-train/LJ/input.nn

@@ -34,6 +34,7 @@ parallel_mode                   0              # Training parallelization used (
 jacobian_mode                   1              # Jacobian computation mode (0 = Summation to single gradient, 1 = Per-task summed gradient, 2 = Full Jacobian).
 update_strategy                 0              # Update strategy (0 = Combined, 1 = Per-element).
 selection_mode                  2              # Update candidate selection mode (0 = Random, 1 = Sort, 2 = Threshold).
+decoupling_type                 0              # Kalman filter decoupling type (0 = GEKF, 1 = ED-GEKF, 2 = Layer, 3 = NDEKF, 4 = FDEKF).
 task_batch_size_energy          1              # Number of energy update candidates prepared per task for each update (0 = Entire training set).
 task_batch_size_force           1              # Number of force update candidates prepared per task for each update (0 = Entire training set).
 memorize_symfunc_results                       # Keep symmetry function results in memory.

src/libnnp/Mode.cpp

@@ -884,7 +884,9 @@ void Mode::setupNeuralNetworkWeights(string const& fileNameFormat)
                 weights.push_back(atof(splitLine.at(0).c_str()));
             }
         }
-        it->neuralNetwork->setConnections(&(weights.front()));
+        // Attention: need alternative (old) ordering scheme here for
+        // backward compatibility!
+        it->neuralNetwork->setConnectionsAO(&(weights.front()));
         file.close();
     }