search

otaheri / GrabNet

GrabNet: A Generative model to generate realistic 3D hands grasping unseen objects (ECCV2020)
https://grab.is.tue.mpg.de
Other
238 stars 30 forks source link

Questions for training codes #8

Closed SeanChenxy closed 3 years ago

SeanChenxy commented 3 years ago

Hi, Omid. Thanks for your sharing. I have two questions when I reading codes. 1) During training, why does NOT the self.rhm_train need grad? https://github.com/otaheri/GrabNet/blob/0b428a72d9575e196503ec56b4c4f5b4fde65914/grabnet/train/trainer.py#L78 2) What is the difference between fpose_rhand_rotmat_f and fpose_rhand_rotmat in training data? Hope for your reply. Thanks.

otaheri commented 3 years ago

Hi @SeanChenxy

Thank you for your interest in our work.

  1. This is a good question. One can also set the self.rhm_train model parameters with require_grad but I think it will have no effect on the training. During the model (GrabNet) training, the MANO parameters are generated using GrabNet which means they will require_grad enabled in any case (and they replace the initial MANO parameters that we define). So, I assume changing it to require_grad would not break anything.

  2. The detailed instruction on what is included in the dataset will be added to the website and repo soon. The ground_truth hand parameters are the ones without any extension at the end like fpose_rhand_rotmat. The _f at the end stands for "fake" and shows the parameters that we used as input when training RefineNet. In other words, the parameters with _f at the end, are the ground_truth parameters when we added random noise to them.

I hope these answer your questions.

SeanChenxy commented 3 years ago

I got it. Thanks a lot.