Fixed TI normalization order

1 files changed, 4 insertions, 4 deletions

diff --git a/training/functional.py b/training/functional.py
index e7c4320..b830261 100644
--- a/training/functional.py
+++ b/training/functional.py
@@ -38,8 +38,8 @@ class TrainingCallbacks():
     on_accum_model: Callable[[], torch.nn.Module] = const(None)
     on_log: Callable[[], dict[str, Any]] = const({})
     on_train: Callable[[int], _GeneratorContextManager] = const(nullcontext())
-    on_before_optimize: Callable[[float, int], None] = const()
-    on_after_optimize: Callable[[float], None] = const()
+    on_before_optimize: Callable[[float, int], Any] = const()
+    on_after_optimize: Callable[[Any, float], None] = const()
     on_after_epoch: Callable[[float], None] = const()
     on_eval: Callable[[], _GeneratorContextManager] = const(nullcontext())
     on_sample: Callable[[int], None] = const()
@@ -455,13 +455,13 @@ def train_loop(
                     local_progress_bar.set_postfix(**logs)
 
                     if ((step + 1) % gradient_accumulation_steps == 0) or ((step + 1) == len(train_dataloader)):
-                        on_before_optimize(lr_scheduler.get_last_lr()[0], epoch)
+                        before_optimize_result = on_before_optimize(lr_scheduler.get_last_lr()[0], epoch)
 
                         optimizer.step()
                         lr_scheduler.step()
                         optimizer.zero_grad(set_to_none=True)
 
-                        on_after_optimize(lr_scheduler.get_last_lr()[0])
+                        on_after_optimize(before_optimize_result, lr_scheduler.get_last_lr()[0])
 
                         local_progress_bar.update(1)
                         global_progress_bar.update(1)