Thank you for sharing this repository. the implementation in diffcr_pseudo.py has been very helpful.
I noticed the current code provides layer-wise compression, but I couldn’t find an implementation of timestep-wise compression.
Do you plan to release the timestep-wise compression code as well?
If possible, could you also share a brief overview of how you implemented it (e.g., timestep partitioning/assignment, batch sampling strategy, any modifications to the training loop or loss aggregation, and scheduler changes)?
Even a short description or pseudocode would be greatly appreciated.
Thanks again for your excellent work and for making the code available.
Best regards,
Thank you for sharing this repository. the implementation in
diffcr_pseudo.pyhas been very helpful.I noticed the current code provides layer-wise compression, but I couldn’t find an implementation of timestep-wise compression.
Do you plan to release the timestep-wise compression code as well?
If possible, could you also share a brief overview of how you implemented it (e.g., timestep partitioning/assignment, batch sampling strategy, any modifications to the training loop or loss aggregation, and scheduler changes)?
Even a short description or pseudocode would be greatly appreciated.
Thanks again for your excellent work and for making the code available.
Best regards,