search

dc3042 / CROM_offline_training

21 stars 4 forks source link

About the exact number of gpus in CROM experiments #4

Open BraveDrXuTF opened 11 months ago

BraveDrXuTF commented 11 months ago

Thank you for your excellent work of CROM and I learned a lot. I'm curious about the exact number of GPUs you use in your experiments. Especially, in Figure 28 CROM costs around 30 mins on elastic case, where I think it impossible for a single gpu to do 30000 epochs (in section D.1) in 30 mins.

dc3042 commented 9 months ago

Hi, thank you for your interest in the repository. To answer your question, the dataset, and the model are both small enough that the entire dataset can fit into a single batch on a single gpu. Since generalizability is not the goal of this model, there is also no need for a separate validation or testing dataset. This allows for the fast training as shown in the paper. Hope this helps!

BraveDrXuTF commented 1 month ago

Is it possible to open your datasets?

peterchencyc commented 1 month ago

Thanks for your interest! Which data are you interested in? We have put some sample data in the git repo: https://www.dropbox.com/sh/c71axdcpxng2nu3/AABp0YZ6ho__Ih3P2cmjMC2Ga?dl=0

Thanks!

On Tue, Sep 3, 2024 at 9:03 AM BraveDrXuTF @.***> wrote:

Is it possible to open your datasets?

— Reply to this email directly, view it on GitHub https://github.com/dc3042/CROM_offline_training/issues/4#issuecomment-2326471421, or unsubscribe https://github.com/notifications/unsubscribe-auth/AC6JUJL3SGW3XF5TS2ID2LTZUWXQVAVCNFSM6AAAAABNSDAQYKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDGMRWGQ3TCNBSGE . You are receiving this because you are subscribed to this thread.Message ID: @.***>