Missing file phc_shape_pnn_train_iccv.yaml and regarding training from scratch

[canoneod]

Hi! Thanks for the great work:)

I'm trying to train pnn from scratch with amass dataset and trying to find phc_shape_pnn_train_iccv.yaml which I can't. Seems to hold some information regarding training as below description in Train section.

Training PHC is not super automated yet, so it requires some (a lot of) manual steps, and invovles changing the config file a couple of times during training based on the training phase. The phc_shape_pnn_train_iccv.yaml config file provides a starting point for training primitives.

First, we will train one primitive, and keep an eye on its performance (--has_eval) flag. In the config, the "training_prim" is the primitive that is being trained. This need to be updated accordingly.

Also, I understood that training should be done iteratively in manual manner chaning argument (epoch and idx) of

python scripts/pmcp/forward_pmcp.py --exp [exp_name] --epoch [epoch] --idx {idx}

python phc/run_hydra.py learning=im_pnn_big env=env_im_pnn robot=smpl_humanoid env.motion_file=[motion_file] epoch=[include epoch+1 from previous step] env.fitting=True env.training_prim=1

To which criteria should I look upon to decide which idx to get rid of? (Based on evaluation during training?)

For the last question, how do you separate motion clip that is easy and hard? I assume there should be orders in AMASS clips to train full pnn.

Thanks in advance!