From 2dc20a2cd1a914c4d3f8523bb33ce77842dfe695 Mon Sep 17 00:00:00 2001
From: Chestnut <88202804+qqingzheng@users.noreply.github.com>
Date: Mon, 8 Apr 2024 23:40:07 +0800
Subject: [PATCH] Update Train_And_Eval_CausalVideoVAE.md

Former-commit-id: 589657b089930379bb75f3d49063ebb72671efc6
---
 docs/Train_And_Eval_CausalVideoVAE.md | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/docs/Train_And_Eval_CausalVideoVAE.md b/docs/Train_And_Eval_CausalVideoVAE.md
index 2cd2687b4..ccde2ba68 100644
--- a/docs/Train_And_Eval_CausalVideoVAE.md
+++ b/docs/Train_And_Eval_CausalVideoVAE.md
@@ -2,6 +2,8 @@
 
 To execute in the terminal: `bash scripts/causalvae/train.sh`
 
+> When using GAN loss for training, two backward propagations are required. However, when [custom optimizers](https://lightning.ai/docs/pytorch/stable/model/manual_optimization.html#use-multiple-optimizers-like-gans) are implemented in PyTorch Lightning, it can lead to the training step count being doubled, meaning each training loop effectively results in two steps. This issue can make it counterintuitive when setting the training step count and the starting step count for the GAN loss.
+
 ## Code Structure
 
 CausalVideoVAE is located in the directory `opensora/models/ae/videobase`. The directory structure is as follows:
@@ -123,4 +125,4 @@ Our model class inherits from the configuration and model management classes of
 ```
 model = CausalVAEModel.from_pretrained(args.ckpt)
 model = model.to(device)
-```
\ No newline at end of file
+```