[WIP] Byol Repro (v2) #454
Conversation
facebook-github-bot
added
the
CLA Signed
|
Note this replaces: @lluiscastrejonsubira PR here: #290 |
|
@iseessel has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
|
@iseessel has updated the pull request. You must reimport the pull request before landing. |
|
@iseessel has updated the pull request. You must reimport the pull request before landing. |
|
@iseessel has updated the pull request. You must reimport the pull request before landing. |
|
@iseessel has updated the pull request. You must reimport the pull request before landing. |
|
@iseessel has updated the pull request. You must reimport the pull request before landing. |
|
@iseessel has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
|
@iseessel has updated the pull request. You must reimport the pull request before landing. |
|
@iseessel, thanks for your great work. Can I know what accuracy does this version of BYOL achieve now? |
|
@zhifanwu This will be finished in the next month or so. I believe it is currently fully reproduced, but will want to run some more experiments to verify. |
|
Hi @iseessel are there any updates on that? Will you upload the weights to the model zoo? |
Continues work done by @lluiscastrejonsubira.
Further Improvements and bug fixes by @iseessel and @pranavsinghps1.
We are still unable to reproduce the full results of the BYOL paper. Here is our current linear transfer results. That is running
byol_8node_resnet.yamland then running/eval_resnet_8gpu_transfer_in1k_linear.yamland reporting the best test results:However, the BYOL paper reports the following, so we are ~3% lower.