zhanghang1989 commented on a change in pull request #16053: Fixes #15543
URL: https://github.com/apache/incubator-mxnet/pull/16053#discussion_r320900353
##########
File path: src/operator/optimizer_op-inl.h
##########
@@ -1119,22 +1112,16 @@ struct MP_NAGMomKernel {
const OpReqType req) {
float w = weight32[i];
if (param_clip_gradient >= 0.0f) {
- mom_data[i] = param_momentum*mom_data[i]
- + mshadow_op::clip::Map(param_rescale_grad
- *static_cast<float>(grad_data[i]), param_clip_gradient)
- + (param_wd*w);
- w = w - param_lr*(param_momentum*mom_data[i]
- + mshadow_op::clip::Map(param_rescale_grad
- *static_cast<float>(grad_data[i]),
- param_clip_gradient));
+ mom_data[i] = param_momentum*mom_data[i];
+ w = w-mom_data[i]+(param_momentum+1)*(mom_data[i]-param_lr
+
*(mshadow_op::clip::Map(param_rescale_grad*static_cast<float>(grad_data[i]),
+ param_clip_gradient)+(param_wd*w)));
Review comment:
Update the momentum state in the end `mom[:] -= grad`
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
