search

syguan96 / DynaBOA

[T-PAMI 2022] Out-of-Domain Human Mesh Reconstruction via Dynamic Bilevel Online Adaptation
225 stars 19 forks source link

Will you release training code? #18

Closed StayYouth1993 closed 2 years ago

StayYouth1993 commented 2 years ago

How many dataset have you used? Can you share some training experience to achieve the performance as paper showed？ex：

  1. used pretrained model or train from zero
  2. how many loss used and loss weight
  3. opt method
  4. training dataset and weight
  5. training epoch

Best！

syguan96 commented 2 years ago

Hi, I only trained the basemodel with Human 3.6M and its SMPL annotation (named mosh). The model is trained from scratch. The used losses and their weight are the same as SPIN, but I removed the optimization module since Human 3.6M already contains SMPL annotation. To accelerate the training, I used apex and DDP. I remember that I trained about 90 epochs with 3x2080ti.

The training code is dirty now. You can re-train the model by yourself first. If I have time, I will reorganize the code and release it.

StayYouth1993 commented 2 years ago

Ok, thanks a lot! I'll try later!

NewCoderQ commented 2 years ago

Hi @syguan96 , any update for training code?