train.py: Use best validation loss for lr scheduling

bmdanielsson · bmdanielsson · commit 5566a8514227 · 2024-01-28T17:52:10.000+01:00
Previously the validation loss at the end of the epoch was used for
learning rate scheduling. Change this to use the best validation loss
for the complete epoch instead.
diff --git a/train.py b/train.py
@@ -184,7 +184,7 @@ def main(args):
         stop = time.monotonic()
         print(f' ({timedelta(seconds=stop-start)})')
 
-        scheduler.step(val_loss)
+        scheduler.step(best_val_loss)
         epoch += 1
 
 
