修复周期性评估时平均损失计算错误

2026-02-13 11:29:22 +08:00 · 2026-02-13 11:29:22 +08:00 · 02f851205f
parent 92b12ef703
commit 02f851205f
1 changed files with 1 additions and 2 deletions
--- a/src/trainer/model.py
+++ b/src/trainer/model.py
@ -437,13 +437,12 @@ class MoEModel(nn.Module):
                    global_step += 1
                    original_loss = loss.item() * grad_accum_steps
                    batch_loss_sum += original_loss
-
                    # 周期性评估（与原代码相同）
                    if (
                        eval_dataloader is not None
                        and global_step % eval_frequency == 0
                    ):
-                        avg_loss = batch_loss_sum / global_step
+                        avg_loss = batch_loss_sum / eval_frequency
                        acc, _ = self.model_eval(eval_dataloader, criterion)
                        super().train()
                        if monitor is not None: