Bug in save_debug with pytorch3D 0.6.2

[Arthur151]

Geat work! The performance is really impressive!

I have struggled to install the old version of pytorch3D (0.4.0) but failed. Therefore, I made some change to the source code of pytorch3D 0.6.2 to make the training run sucessfully.

File "train.py", line 167, in <module>
    loss=optNet(outs,sample_pix_num,ratio,frame_ids,debug_root)		
  File "/home/yusun/.local/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1110, in _call_impl
    return forward_call(*input, **kwargs)
  File "/home/yusun/data_drive/existing_works/SelfReconCode/model/network.py", line 506, in forward
    self.save_debug(oldTmpVs,self.Tmpfs,defMeshes,self.deformer.defs[0].offset.view(N,-1,3).detach(),masks,gtMs,mgtMs,gtCs,batch_inds,row_inds,col_inds,initTmpPs,[d_cond.detach(),[poses.detach(),trans.detach()]],rendcond,ratio)
  File "/home/yusun/data_drive/existing_works/SelfReconCode/model/network.py", line 385, in save_debug
    defMeshes=Meshes(verts=[vs for vs in torch.split(TmpVs[None,:,:].expand(N,-1,3)+offset,1)],faces=[Tmpfs for _ in range(N)],textures=TexturesVertex([torch.ones_like(TmpVs) for _ in range(N)]))
  File "/home/yusun/anaconda3/envs/icon/lib/python3.8/site-packages/pytorch3d-0.6.2-py3.8-linux-x86_64.egg/pytorch3d/structures/meshes.py", line 436, in __init__
    raise ValueError(msg)
ValueError: Textures do not match the dimensions of Meshes.

But there is a bug when running into save_debug.

I suppose that fixing this bug, then we can use pytorch3D 0.6.2 as the default installation version, which is way more convenient to install.

Looking forward to your reply!

[jby1993]

The API has some changes for higher versions and I do not check it carefully, you can try to update it by yourself. Feel free to propose a pull request.