pose length error when running with run_prior=True on demo2 sequence

[zblzcj]

First of all, thank you for the great work and for open-sourcing this project!

I'm encountering an issue when running the example with run_prior: True, using the demo2 sequence. Below is the relevant output and error trace:

/app/outputs/logs/video-custom/2025-07-09/demo2-all-shot-0-0-500 ... /app/_DATA/hmp_model/mean-neutral-128-30fps.pt
mean: ['rotmat', 'pos', 'trans', 'root_vel', 'height', 'rot6d', 'angular', 'global_xform', 'velocity', 'root_orient']
std: ['rotmat', 'pos', 'trans', 'root_vel', 'height', 'rot6d', 'angular', 'global_xform', 'velocity', 'root_orient']
Loading from /app/_DATA/hmp_model/results/model
Load succeeded
2025-07-09 21:54:15.001 | INFO     | HMP.fitting:multi_stage_opt:680 - Running reconstruction with prior
loading data...
obs_data:  dict_keys(['joints2d', 'init_body_pose', 'init_body_shape', 'init_root_orient', 'init_trans', 'is_right', 'vis_mask', 'seq_interval', 'track_interval', 'track_id', 'seq_name'])
res_dict:  dict_keys(['betas', 'trans', 'world_scale', 'latent_pose', 'root_orient', 'init_body_pose', 'is_right', 'pose_body', 'cam_R', 'cam_t', 'intrins'])
betas (2, 10)
trans (2, 189, 3)
world_scale (1, 1)
latent_pose (2, 189, 15, 3)
root_orient (2, 189, 3)
init_body_pose (2, 189, 15, 3)
is_right (2, 189)
pose_body (2, 189, 15, 3)
cam_R (2, 189, 3, 3)
cam_t (2, 189, 3)
intrins (4,)
Sequence has 189 frames
torch.Size([189, 16, 3])
Error executing job with overrides: ['data=video', 'run_opt=True', 'data.seq=demo2']
Traceback (most recent call last):
  File "/app/dyn-hamr/run_opt.py", line 197, in main
    run_opt(cfg, dataset, out_dir, device)
  File "/app/dyn-hamr/run_opt.py", line 166, in run_opt
    run_prior(cfg, dataset, out_dir, device, ['smooth_fit'], \
  File "/app/dyn-hamr/HMP/fitting.py", line 1724, in run_prior
    fitting_prior(obs_data, res_dicts, hand_model, opt, data_args, prior_out, device)
  File "/app/dyn-hamr/HMP/fitting.py", line 1635, in fitting_prior
    multi_stage_opt(opt, device, obs_data, res_dict, hand_model, os.path.join(os.path.dirname(__file__), opt.HMP.config), opt.HMP.exp_name, init_method)
  File "/app/dyn-hamr/HMP/fitting.py", line 750, in multi_stage_opt
    data = get_stage2_res(opt.paths.base_dir, device, init_dict, hand_model)
  File "/app/dyn-hamr/HMP/fitting.py", line 216, in get_stage2_res
    assert len(pose) == 128
AssertionError

It seems like the code expects the sequence length to be exactly 128 frames. Should I manually trim or split my video into 128-frame segments before running with run_prior: True? Or is there another recommended way to handle longer sequences?

Thanks in advance for your help!

[ZhengdiYu]

Thank you for your interest in our work!

Regarding your issue, the current version should compatible with your settings. Could you share the .yaml config file you are using for this?

[zblzcj]

Thank you for the quick response!

I’m using the following config:

defaults:
  - data: HOT3D
  - optim
  - _self_

model:
  use_init: True
  opt_cams: False #True
  opt_scale: True
  async_tracks: True

overwrite: False
run_opt: True
run_vis: True
run_prior: True
datatype: ${data.video_dir}
vis:
  phases:
    # - motion_chunks
    # - input
    # - root_fit
    - smooth_fit
    - prior
    # - init
  render_views:
    - above
    - side
    - front
    # - src_cam
  make_grid: True
  overwrite: False

MANO:
  MODEL_PATH: ${paths.DATA_DIR}/mano
  GENDER: neutral
  NUM_HAND_JOINTS: 15
  MEAN_PARAMS: ${paths.DATA_DIR}/mano_mean_params.npz
  CREATE_BODY_POSE: FALSE

paths:
  base_dir: None
  DATA_DIR: _DATA/data/
  vposer: VPoser/pretrained/Vposer_right_mirrored # _DATA/body_models/vposer_v1_0

HMP:
  config: hmp_config.yaml
  use_hposer: False
  vid_path: ${data.root}/video/${data.seq}.mp4
  exp_name: None

fps: 30
log_root: ../outputs/logs
log_dir: ${log_root}/${data.type}-${data.split}
exp_name: ${now:%Y-%m-%d}
temporal_smooth: True
gpu: 0

hydra:
  job:
    chdir: True
  run:
    dir: ${log_dir}/${exp_name}/${data.name}

The only changes I made were setting run_prior: True and gpu: 0. Everything else remains the same as the default config from the repository.

I’m running it with the following command:

python run_opt.py data=video run_opt=True data.seq=demo2

Let me know if I should double-check anything else or provide additional logs. Thanks again!

[ZhengdiYu]

Thanks for sharing this. Could you also share the Dyn-HaMR/dyn-hamr/confs/data/video.yaml? You might have to modify start_idx and end_idx in there