浏览代码

!8590 debugger should not read unused slots for BatchNorm

From: @john_tzanakakis
Reviewed-by: @mikef,@nsyca
Signed-off-by: @nsyca
tags/v1.1.0
mindspore-ci-bot Gitee 5 年前
父节点
当前提交
009c0e9372
共有 1 个文件被更改,包括 2 次插入2 次删除
  1. +2
    -2
      mindspore/ccsrc/runtime/device/gpu/gpu_kernel_runtime.cc

+ 2
- 2
mindspore/ccsrc/runtime/device/gpu/gpu_kernel_runtime.cc 查看文件

@@ -89,8 +89,8 @@ std::vector<int> CheckRealOutput(const std::string &node_name, const size_t &out
std::vector<int> real_outputs;
// P.FusedBatchNorm is used for training; P.BatchNorm is used for inference
// can add the filter list for more operators here....
if (node_name == "FusedBatchNorm") {
MS_LOG(INFO) << "loading node named FusedBatchNorm.";
if (node_name == "FusedBatchNorm" || node_name == "BatchNorm") {
MS_LOG(INFO) << "loading node named " << node_name;
real_outputs.insert(real_outputs.end(), {0, 3, 4});
} else {
// by default, TensorLoader will load all outputs


正在加载...
取消
保存