Closed
[Bug-Report|缺陷反馈]: sparse_apply_adagrad_v2修改计算顺序,保证inf/nan极值场景精度与竞品一致 #6513
renshuojianAdd on  Oct 12Closed at  Oct 12
renshuojian
renshuojianCollaborators
Oct 12 Create

Describe the current behavior / 问题描述 (Mandatory / 必填)

gradgrad 与后续加法被编译器融合为 FMA 单舍入,乘法与除法结合顺序由 lr(grad/denom)需要改为(lr*grad)/denom

Environment Information / 环境信息 (Mandatory / 必填)

950

Steps to reproduce the issue / 重现步骤 (Mandatory / 必填)

测试inf/nan场景的用例

Describe the expected behavior / 预期结果 (Mandatory / 必填)

inf/nan场景的精度ok

image.png

Special notes for this issue/备注 (Optional / 选填)

likedislike
yuning_chenyuning_chenCollaborators
Oct 12 Assign rsj007 as responsible person
CANN-robotCANN-robotCollaborators
Oct 12 Close issue
CANN-robotCANN-robotCollaborators
Oct 12 Added label: resolved