Change scheduler call location depending on scheduler type

sidekit/nnet/xvector.py

@@ -1016,6 +1016,10 @@ def xtrain(dataset_description,
             if local_rank < 1:
                 save_model(model, monitor, model_opts, training_opts, optimizer, scheduler, epoch)
 
+            # Call scheduler step when validation data is needed for epoch scheduler
+            if isinstance(scheduler, torch.optim.lr_scheduler.ReduceLROnPlateau):
+                scheduler.step(monitor.best_eer)
+
     # TODO manage display using Training Monitor
     if local_rank < 1:
         monitor.display_final()
@@ -1099,10 +1103,14 @@ def train_epoch(model,
             running_loss = 0.0
             accuracy = 0.0
             batch_count = 0
-        if isinstance(scheduler, torch.optim.lr_scheduler.ReduceLROnPlateau):
-            scheduler.step(training_monitor.best_eer)
-        else:
+        # Call scheduler step only for batch scheduler
+        if isinstance(scheduler, (torch.optim.lr_scheduler.OneCycleLR,
+                                  torch.optim.lr_scheduler.CyclicLR)):
             scheduler.step()
+    # Call scheduler step for epoch scheduler without validation data needed
+    if isinstance(scheduler, (torch.optim.lr_scheduler.MultiStepLR,
+                              torch.optim.lr_scheduler.StepLR)):
+        scheduler.step()
     return model