FAQ / Troubleshooting

Hydra config errors — Ensure you are using the correct experiment config name (gem_smpl_regression or gem_smpl). Run python scripts/train.py --help to see available options.

Body model — This repo uses SMPL (body only). For full-body estimation with SOMA including hands and face, see GEM-X.

Hardware requirements — NVIDIA A100 or newer recommended. Requires CUDA 12.4+. Inference uses approximately 16 GB of GPU memory.

W&B logging — Disable with use_wandb=false appended to any training or evaluation command.

Multi-GPU training — Add pl_trainer.devices=N for DDP training across N GPUs.

Missing datasets — Only evaluation datasets are needed to run task=test. Training datasets are only required if you intend to train from scratch.