maybe a bug

[lcc815]

Thanks for your great work. When I start training by "--resume_from some_chepoints.pth", I encounter following error. Is this a bug?

  File "/code/drivestudio/tools/train.py", line 376, in <module>
    final_step = main(args)
  File "/code/drivestudio/tools/train.py", line 260, in main
    loss_dict = trainer.compute_losses(
  File "/code/drivestudio/models/trainers/scene_graph.py", line 285, in compute_losses
    loss_dict = super().compute_losses(outputs, image_infos, cam_infos)
  File "/code/drivestudio/models/trainers/base.py", line 625, in compute_losses
    class_reg_loss = self.models[class_name].compute_reg_loss()
  File "/code/drivestudio/models/nodes/smpl.py", line 466, in compute_reg_loss
    nn_ind = self.nn_ind[instances_mask] # (num_instances, smpl_points_num, knn_neighbors)
  File "/usr/local/lib/python3.8/dist-packages/torch/nn/modules/module.py", line 1265, in __getattr__
    raise AttributeError("'{}' object has no attribute '{}'".format(
AttributeError: 'SMPLNodes' object has no attribute 'nn_ind'

Training from scratch is ok.

[ziyc]

Hi @lcc815,

Can you provide the details of experimental setting used? The dataset, scene id, start end frame index, and the camera settings ect. I'll try if I can reproduce this error from my end.

[lcc815]

Sure. my trainning command is: python train.py --config_file configs/omnire.yaml --output_root output --project omnire --resume_from omnire/exp2/checkpoint_15000.pth --run_name exp2 dataset=waymo/1cams data.scene_idx=114 data.start_timestep=0 data.end_timestep=100 Others remain the same as the repo.