From a3fd42637cdd89015029c43bbfbc57f020114c81 Mon Sep 17 00:00:00 2001
From: littletomatodonkey <2120160898@bit.edu.cn>
Date: Thu, 6 May 2021 15:21:14 +0800
Subject: [PATCH] Update train_with_DALI.md

---
 docs/zh_CN/extension/train_with_DALI.md | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)

diff --git a/docs/zh_CN/extension/train_with_DALI.md b/docs/zh_CN/extension/train_with_DALI.md
index 6e7e977c..298deafa 100644
--- a/docs/zh_CN/extension/train_with_DALI.md
+++ b/docs/zh_CN/extension/train_with_DALI.md
@@ -49,8 +49,14 @@ python -m paddle.distributed.launch \
 
 ## 使用FP16训练
 
-在上述基础上，使用FP16半精度训练，可以进一步提高速度，只需在启动训练命令中添加字段`AMP.use_pure_fp16=True`：
+在上述基础上，使用FP16半精度训练，可以进一步提高速度，可以参考下面的配置与运行命令。
 
 ```shell
-python tools/static/train.py -c configs/ResNet/ResNet50.yaml -o use_dali=True -o AMP.use_pure_fp16=True
+export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+export FLAGS_fraction_of_gpu_memory_to_use=0.8
+
+python -m paddle.distributed.launch \
+    --gpus="0,1,2,3,4,5,6,7" \
+    tools/static/train.py \
+    -c configs/ResNet/ResNet50_fp16.yaml
 ```
-- 
GitLab