run align_to_SMPL_seq.py script error

[2382752230]

Hi! Thanks for your great work!

I have an error running the align_to_SMPL_seq.py script. Does this script require the smpl sequence format? This works fine when running the example, but the following error occurs when processing AMASS/CMU/41/41_02_poses.npz.

Running batch optimization: 50%|█████ | 1/2 [04:50<04:50, 290.76s/it]
Traceback (most recent call last):
File "/home/oneday/SKEL/examples/align_to_SMPL_seq.py", line 60, in
skel_seq = skel_fitter.run_fit(smpl_seq['trans'],
File "/home/oneday/SKEL/skel/alignment/aligner.py", line 119, in run_fit
betas, poses, trans, verts = self._fit_batch(body_params, i, i_start, i_end)
File "/home/oneday/SKEL/skel/alignment/aligner.py", line 202, in _fit_batch
smpl_output = self.smpl(betas=betas_smpl, body_pose=poses_smpl[:,3:], transl=trans_smpl, global_orient=poses_smpl[:,:3])
File "/home/oneday/anaconda3/envs/skel/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1190, in _call_impl
return forward_call(*input, **kwargs)
File "/home/oneday/anaconda3/envs/skel/lib/python3.8/site-packages/smplx/body_models.py", line 366, in forward
num_repeats = int(batch_size / betas.shape[0])
ZeroDivisionError: division by zero

[MarilynKeller]

Hi, this code requires SMPL or SMPLH sequences, make sure you downloaded the proper version from amass. The example sequence I provide is also from MU, did you manage to run the fit with that example sequence?

[2382752230]

Hi,thank you for your reply.
Yes, I run you provide sample (/home/oneday/SKEL/examples/samples/amass_seq/CMU_01_01.npz) can be normal operation. However, there are errors when running other amass sequences (the version downloaded from the AMASS website is SMPL-HG), such as running:/ home/oneday/aitviewer - skel/Data/AMASS/TotalCapture/s1/acting1_poses.npz appear the following error.

Running batch optimization: 50%|█████ | 1/2 [05:06<05:06, 306.01s/it]
Traceback (most recent call last):
File "/home/oneday/SKEL/examples/align_to_SMPL_seq.py", line 60, in
skel_seq = skel_fitter.run_fit(smpl_seq['trans'],
File "/home/oneday/SKEL/skel/alignment/aligner.py", line 118, in run_fit
betas, poses, trans, verts = self._fit_batch(body_params, i, i_start, i_end)
File "/home/oneday/SKEL/skel/alignment/aligner.py", line 201, in _fit_batch
smpl_output = self.smpl(betas=betas_smpl, body_pose=poses_smpl[:,3:], transl=trans_smpl, global_orient=poses_smpl[:,:3])
File "/home/oneday/anaconda3/envs/skel/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1190, in _call_impl
return forward_call(*input, **kwargs)
File "/home/oneday/anaconda3/envs/skel/lib/python3.8/site-packages/smplx/body_models.py", line 366, in forward
num_repeats = int(batch_size / betas.shape[0])
ZeroDivisionError: division by zero

Here and the content of the error is consistent with the above, you can try to run the AMASS/CMU/41/41_02_poses.npz sequences or AMASS TotalCapture/s1/acting1_poses.npz sequence? Does it work on your site?

[2382752230]

Hi,I think the problem is the number of batches. In the code, the size of batch_size is 3000. When the smpl sequence length exceeds 3000, two batches are needed, and the size of betas will be (0,10) in the second batch, leading to an error.

I changed the size of batch_size to 5000 so that the smpl sequence can be processed in one batch, at which time the program can work normally.

Does the size of batch_size affect the registration performance? Can I set it to 5000 or 6000?

[MarilynKeller]

I see, good catch, that's probably a bug, the betas batch dim should not be zero. I'll look into it. Having more batch helps the registration usually rather than the inverse. As the first batch last frames results is used to initialized the next frames. If it works with 5k, 6k frames, there is no problem doing the fit in one batch. The limit for the batch size is your memory. That's why we implemented the batching.