search

pengHTYX / Era3D

GNU Affero General Public License v3.0
543 stars 24 forks source link

About predict_relative_views #50

Open lzt02 opened 2 weeks ago

lzt02 commented 2 weeks ago

I noticed that in the config, there are predict_delative-views: [0, 2, 4, 8, 12, 14], which appear to have 0 as the front view and 8 as the back view. However, when loading the training set data, there is code: if self. validation: rond_ele0_idx=12, which appears to have 12 as the front view. I would like to ask for the correct order of views and their corresponding ids.

Thanks in advance for your help.

pengHTYX commented 1 week ago

hi, sorry for the late reply. The id of front view is 12 and [0, 2, 4, 8, 12, 14] is the relative offset to the front view.

lzt02 commented 1 week ago

Ok, I get it, thanks for your reply! By the way, I noticed that the training code has two prediction types for diffusion: "v-prediction" and "epsilon". Which one did you use when training Era3D

hi, sorry for the late reply. The id of front view is 12 and [0, 2, 4, 8, 12, 14] is the relative offset to the front view.

pengHTYX commented 1 week ago

We use 'v_prediction' for diffusion training.