提交 2f5c0c71 编写于 作者: G gaotingquan 提交者: Tingquan Gao

fix: replace ResNet50.yaml with ResNet50_amp_O2_ultra.yaml when AMPO2

It is needed that use dali, NHWC and 4 channels when training ResNet50 with AMPO2
上级 ea66adbe
......@@ -179,6 +179,11 @@ for batch_size in ${batch_size_list[*]}; do
func_sed_params "$FILENAME" "${line_epoch}" "$epoch"
gpu_id=$(set_gpu_id $device_num)
# It is needed that using dali, NHWC and 4 channels when training ResNet50 with AMPO2
if [[ $model_name == "ResNet50" && $precision == "fp16" ]]; then
sed -i "s/ResNet50.yaml/ResNet50_amp_O2_ultra.yaml/g" $FILENAME
fi
# if bs is big, then copy train_list.txt to generate more train log
# At least 25 log number would be good to calculate ips for benchmark system.
# So the copy number for train_list is as follows:
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册