search

mlcommons / training_policies

Issues related to MLPerf™ training policies, including rules and suggested changes
https://mlcommons.org/en/groups/training
Apache License 2.0
92 stars 65 forks source link

[Minigo] Pre-trained checkpoint conversion script across frameworks #420

Open nv-rborkar opened 3 years ago

nv-rborkar commented 3 years ago

As per #317 , we would like all references that use pre-trained weights or start from a checkpoint to provide conversion scripts for use across different frameworks (Pytorch, TF, & other frameworks of choice for submission) to ensure submitters use the same pre-approved checkpoint.

AI Intel (reference owner) for Minigo Example script & recipe in #391

johntran-nv commented 3 years ago

@christ1ne could you assign to the reference owner?

christ1ne commented 3 years ago

@johntran-nv are you referring to the following example for converting the weights? https://github.com/mlcommons/training/pull/436 That's a from PyTorch model. Is there an TF model weights export example? Thanks!

@delock who's the reference owner.

johntran-nv commented 3 years ago

Yes that's what we're referring to. Sorry but I don't know of a similar TF export example. We could ask the Google folks for help, if you want.

christ1ne commented 3 years ago

I think we should ask the Google folks for an TF example. Meanwhile, I am not sure if all the TF models can be converted to Pytorch as they are today.