Pertubation Kernel Support for Zeroth-Order Training on SleepViT#165
Pertubation Kernel Support for Zeroth-Order Training on SleepViT#165JanCSEM wants to merge 17 commits intopulp-platform:develfrom
Conversation
|
Hi Jan, all the perturbation kernel implementations still need to be implemented once again on GAP9. |
runwangdl
changed the title
Yep - WIP |
runwangdl
left a comment
runwangdl
left a comment
There was a problem hiding this comment.
Nice work, Jan! I have one question: why do you need DeeployTest/testRunner_tiled_siracusa_mezo.py and --run_mode argument? I thought Onnx4Deeploy generates the Zeroth-Order graph, and Deeploy treats it as a normal inference graph.
| help = | ||
| "Number of cores on which the network is run. Currently, required for im2col buffer sizing on Siracusa. Default: 1." | ||
| ) | ||
| parser.add_argument('--run_mode', type = str, default = 'inference', |
There was a problem hiding this comment.
Why do you need this argument?
There was a problem hiding this comment.
I don't need it anymore, this was for my hacky multi-iteration test.
|
Ciao @JanCSEM, any update on this PR? |
Hey Victor, I opened 2 new ones: zo-support and zo-support-clean. unfortunately zo-support has many conflicts with devel because my ZO work depends on non-merged modifs made in the meantime... |
3 participants
Describe the intent of your PR here.
Added
Changed
Fixed
PR Merge Checklist
develcommit and pointing todevel.CHANGELOG.mdfile has been updated.