Closed sx-liu closed 2 months ago
Seems like test_ihvp_datainf_nn
is a bit shaky.
Seems like
test_ihvp_datainf_nn
is a bit shaky.
Could lower the threshold to 0.8.
Seems like
test_ihvp_datainf_nn
is a bit shaky.Could lower the threshold to 0.8.
Should I do it here or in a separate PR?
Seems like
test_ihvp_datainf_nn
is a bit shaky.Could lower the threshold to 0.8.
Should I do it here or in a separate PR?
You could just do it here. This shaky test is going to block other PRs if we merge this PR without lowering the threshold.
Description
1. Motivation and Context
This PR is the same as the closed one #79.
To complete the proposal in the issue #76.
2. Summary of the change
The implementation is basically the same as originally designed.
Specifically,
ihvp_at_x_ekfac
and its helper functions todattri.func.ihvp
.manual
and classMLPCache
todattri.func.ihvp
.dattri.func.ihvp._check_input_size
to support the EK-FAC method.3. What tests have been added/updated for the change?