jepa icon indicating copy to clipboard operation
jepa copied to clipboard

Context encoder and predictor weights

Open mirlanium opened this issue 9 months ago • 2 comments

@MidoAssran @soniajoseph Thanks for the great work and maintaining the repo. I would like to fine tune the model, but if I am not mistaken, the context encoder and predictor weights are not published. Are you planning to publish those? If not, can I use target encoder weights as context encoder weights and just train a predictor from scratch?

mirlanium avatar Mar 06 '25 13:03 mirlanium

Hi!

The context encoder weights and the predictor weights are contained as dictionary entries within the published model, alongside the target encoder.

Lmk if confusing. I was thinking of putting all three on Huggingface too.

On Thu, Mar 6, 2025 at 8:41 AM mirl6n @.***> wrote:

@MidoAssran https://github.com/MidoAssran @soniajoseph https://github.com/soniajoseph Thanks for the great work and maintaining the repo. I would like to fine tune the model, but if I am not mistaken, the context encoder and predictor weights are not published. Are you planning to publish those? If not, can I use target encoder weights as context encoder weights and just train a predictor from scratch?

— Reply to this email directly, view it on GitHub https://github.com/facebookresearch/jepa/issues/91, or unsubscribe https://github.com/notifications/unsubscribe-auth/AHBIYMFWYX3ZQSJX7WY2YOT2TBGCHAVCNFSM6AAAAABYOXH4CWVHI2DSMVQWIX3LMV43ASLTON2WKOZSHEYDANBUHEYTONI . You are receiving this because you were mentioned.Message ID: @.***> [image: mirl6n]mirl6n created an issue (facebookresearch/jepa#91) https://github.com/facebookresearch/jepa/issues/91

@MidoAssran https://github.com/MidoAssran @soniajoseph https://github.com/soniajoseph Thanks for the great work and maintaining the repo. I would like to fine tune the model, but if I am not mistaken, the context encoder and predictor weights are not published. Are you planning to publish those? If not, can I use target encoder weights as context encoder weights and just train a predictor from scratch?

— Reply to this email directly, view it on GitHub https://github.com/facebookresearch/jepa/issues/91, or unsubscribe https://github.com/notifications/unsubscribe-auth/AHBIYMFWYX3ZQSJX7WY2YOT2TBGCHAVCNFSM6AAAAABYOXH4CWVHI2DSMVQWIX3LMV43ASLTON2WKOZSHEYDANBUHEYTONI . You are receiving this because you were mentioned.Message ID: @.***>

soniajoseph avatar Mar 06 '25 14:03 soniajoseph

@soniajoseph Thanks for the quick reply! Any plans to release the conditional diffusion decoder?

mirlanium avatar Mar 06 '25 16:03 mirlanium