search

ml-struct-bio / cryodrgn

Neural networks for cryo-EM reconstruction
http://cryodrgn.cs.princeton.edu
GNU General Public License v3.0
305 stars 75 forks source link

Question about running time of cryoDRGN2 and NeRF #99

Open zhouzhenghong-gt opened 2 years ago

zhouzhenghong-gt commented 2 years ago

Thanks for your great work!! And I have two question about cryoDRGN2.

  1. Will training schedule of cryoDRGN2 lead to a long training time?
  2. In your cryoDRGN2 paper, you mention that gradient descent can't optimize pose, have you tried some NeRF related methods such as nerf--, BARF or GNeRF? Thanks for your time!
zhonge commented 2 years ago

Thanks for asking.

1) yes it will lead to a much longer training time than without pose search. Despite that, cryoDRGN2 is significantly faster and more accurate than cryoDRGN1's pose inference (described in the ICLR paper), though we're thinking of ways to make it even faster

2) yes, first order gradient based optimization has been technically available in the main cryoDRGN software since the initial release (--do-pose-sgd) but requires different training settings and doesn't lead to a large difference in many cases, I believe due to the high amount of noise in cryo-EM images. It could be evaluated further though.

On Wed, Jan 5, 2022, 7:39 AM ZZH @.***> wrote:

Thanks for your great work!! And I have two question about cryoDRGN2.

  1. Will training schedule of cryoDRGN2 lead to a long training time?
  2. In your cryoDRGN2 paper, you mention that gradient descent can't optimize pose, have you tried some NeRF related methods such as nerf--, BARF or GNeRF? Thanks for your time!

— Reply to this email directly, view it on GitHub https://github.com/zhonge/cryodrgn/issues/99, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABJVIJBRA57WDLGYXUBOODTUUQ3WTANCNFSM5LJZ53EQ . Triage notifications on the go with GitHub Mobile for iOS https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Android https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub.

You are receiving this because you are subscribed to this thread.Message ID: @.***>

ag-deepesh commented 2 years ago

Hi Ellen, Indeed Great work ! Looking forward to run CryoDRGN2. Will it not require Relion or CryoSparc to generate "homogeneous reconstruction" to get poses which is an input for CryoDRGN? Also, is the beta-version of the CryoDRGN2 code available somewhere already or you are planning to release it in sometime ? Thank you.