Functions to run experiments with parametrize inputs to the neural ne…

…twork models

benchmarks/roman_pots/train_dense_neural_network.py

@@ -46,7 +46,9 @@ def standardize(x):
   return standardized_tensor, mean, std
 
 def train_model(input_tensor, target_tensor, model, num_epochs, learning_rate):
-
+  # Send model to device
+  model=model.to(device)
+
   # Define the loss function and optimizer
   criterion = torch.nn.HuberLoss(reduction='mean', delta=1.0)
   optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
@@ -88,4 +90,74 @@ def train_model(input_tensor, target_tensor, model, num_epochs, learning_rate):
 
   return model
 
+def run_experiment(input_files, target_files, hyperparameters):
+
+  # Load input and target training data in tensors
+  training_RP_pos = pd.DataFrame()
+  training_MC_mom = pd.DataFrame()
+
+  for i in range(1,num_training_inputs+1):
+    temp_training_RP_pos = pd.read_csv(hyperparameters.input_files+str(i)+'.txt', delimiter='\t', header=None)
+    training_RP_pos = pd.concat([training_RP_pos, temp_training_RP_pos], ignore_index=True)
+    temp_training_MC_mom = pd.read_csv(hyperparameters.target_files+str(i)+'.txt', delimiter='\t', header=None)
+    training_MC_mom = pd.concat([training_MC_mom, temp_training_MC_mom], ignore_index=True)
+
+  training_RP_pos_tensor = torch.tensor(training_RP_pos.values, dtype=torch.float32)
+  training_MC_mom_tensor = torch.tensor(training_MC_mom.values, dtype=torch.float32)
+
+  # Standardize training data
+  source_pz = training_RP_pos_tensor
+  scaled_source_pz, mean_source_pz, std_source_pz = standardize(source_pz)
+  target_pz = training_MC_mom_tensor[:,2].unsqueeze(1)
+
+  source_py = torch.cat((training_RP_pos_tensor[:,2:4], training_MC_mom_tensor[:,2].unsqueeze(1)), 1)
+  scaled_source_py, mean_source_py, std_source_py = standardize(source_py)
+  target_py = training_MC_mom_tensor[:,1].unsqueeze(1)
+
+  source_px = torch.cat((training_RP_pos_tensor[:,0:2], training_MC_mom_tensor[:,2].unsqueeze(1)), 1)
+  scaled_source_px, mean_source_px, std_source_px = standardize(source_px)
+  target_px = training_MC_mom_tensor[:,0].unsqueeze(1)
+
+  # Initialize models
+  initial_model_pz = NeuralNet(size_input=hyperparameters.size_input_pz,
+                               size_output=hyperparameters.size_output_pz, 
+                               n_layers=hyperparameters.n_layers_pz,
+                               size_first_hidden_layer=hyperparameters.size_first_hidden_layer_pz,
+                               multiplier=hyperparameters.muliplier_pz,
+                               leak_rate=hyperparameters.leak_rate_pz)
+  initial_model_py = NeuralNet(size_input=hyperparameters.size_input_py,
+                               size_output=hyperparameters.size_output_py, 
+                               n_layers=hyperparameters.n_layers_py,
+                               size_first_hidden_layer=hyperparameters.size_first_hidden_layer_py,
+                               multiplier=hyperparameters.muliplier_py,
+                               leak_rate=hyperparameters.leak_rate_py)
+  initial_model_px = NeuralNet(size_input=hyperparameters.size_input_px,
+                               size_output=hyperparameters.size_output_px, 
+                               n_layers=hyperparameters.n_layers_px,
+                               size_first_hidden_layer=hyperparameters.size_first_hidden_layer_px,
+                               multiplier=hyperparameters.muliplier_px,
+                               leak_rate=hyperparameters.leak_rate_px)
+
+  # Train models
+  model_pz = train_model(scaled_source_pz, target_pz, initial_model_pz, num_epochs=hyperparameters.num_epochs_pz, learning_rate=hyperparameters.learning_rate_pz)
+  model_py = train_model(scaled_source_py, target_py, initial_model_py, num_epochs=hyperparameters.num_epochs_py, learning_rate=hyperparameters.learning_rate_py)
+  model_px = train_model(scaled_source_px, target_px, initial_model_px, num_epochs=hyperparameters.num_epochs_px, learning_rate=hyperparameters.learning_rate_px)
+
+  # Save models
+  torch.jit.script(model_pz).save('model_pz.pt')
+  torch.jit.script(model_py).save('model_py.pt')
+  torch.jit.script(model_px).save('model_px.pt')
+
+
+if __name__ == "__main__":
+  parser = argparse.ArgumentParser(fromfile_prefix_chars='@')
+  hyperparameters_list = ['--input_files', '--target_files', '--num_training_inputs', 
+                   '--num_epochs_pz', '--learning_rate_pz', '--size_input_pz', '--size_output_pz', '--n_layers_pz', '--size_first_hidden_layer_pz', '--multiplier_pz', 'leak_rate_pz',
+                   '--num_epochs_py', '--learning_rate_py', '--size_input_py', '--size_output_py', '--n_layers_py', '--size_first_hidden_layer_py', '--multiplier_py', 'leak_rate_py',
+                   '--num_epochs_px', '--learning_rate_px', '--size_input_px', '--size_output_px', '--n_layers_px', '--size_first_hidden_layer_px', '--multiplier_px', 'leak_rate_px']
+  for hyperparameter in hyperparameters_list:
+    parser.add_arguments(hyperparameter)
+  hyperparameters = parser.parse_args(['@'+str(sys.argv[1])])
+  run_experiment(hyperparameters)
+