use_patience

train.py

@@ -55,12 +55,14 @@
 
 # Number of trials to run.
 # Each trial generates a new neural network topology with new hyper parameters and trains it.
-NUMBER_OF_STUDY_TRIALS = 150
+NUMBER_OF_STUDY_TRIALS = 100
 
 # Hyper parameter search bounds
 NUM_EPOCHS = 2000
+# Patience is an custom terminator that stops a training if no improvement has happened in this many epochs
+USE_PATIENCE = False
 MIN_PATIENCE = 25
-MAX_PATIENCE = 300  # if no improvement in this many epochs, stop early
+MAX_PATIENCE = 300
 MIN_NUMBER_OF_EPOCHS = 50
 MIN_HIDDEN_LAYERS = 3
 MAX_HIDDEN_LAYERS = 9
@@ -618,7 +620,7 @@ def objective(trial):
 
     total_loss = None
     best_epoch = best_loss = best_state_dict = None
-    patience = trial.suggest_int("patience", MIN_PATIENCE, MAX_PATIENCE)
+    patience = trial.suggest_int("patience", MIN_PATIENCE, MAX_PATIENCE) if USE_PATIENCE else 0
     for epoch in range(NUM_EPOCHS):
         # Train the model
         model.train()
@@ -651,7 +653,7 @@ def objective(trial):
             best_state_dict = model.state_dict()
         else:
             epochs_since_best = epoch - best_epoch
-            if epochs_since_best >= patience:
+            if USE_PATIENCE and epochs_since_best >= patience:
                 # Stop early, no improvement in awhile
                 break