Hzzzzx0/official-textop-robot-humanml-artifacts

Official TextOp Deliverable

This folder is a compact handoff package for the current official TextOp baseline on our dataset.

Included files:

1. mvae_ckpt_10000.pth Recommended MVAE checkpoint for downstream use.
2. dar_ckpt_50000.pth Latest saved DAR checkpoint from the current training run.
3. mvae_run.log MVAE resumed training log.
4. dar_run.log DAR training log.
5. dar_cfg.yaml Exact DAR config saved by Hydra for this run.
6. vae_src.log Source path of the MVAE checkpoint cached into the DAR run.
7. mvae_reconstruction_sample.npz One exported MVAE reconstruction sample.
8. mvae_reconstruction_summary.json Reconstruction error summary for that sample batch.

Current status:

1. MVAE was trained to a late, flatter regime and stopped to free GPU for DAR.
2. DAR was trained successfully to step 50000.
3. DAR loss is normal and clearly lower than the early stage.

Useful numbers:

1. MVAE loss/train_total: 0.459224 -> 0.006279
2. DAR loss/train_total: 0.498519 -> 0.100912
3. DAR loss/train_rec: 0.109399 -> 0.021164
4. DAR loss/train_latent_rec: 0.389046 -> 0.079731
5. MVAE reconstruction sample mean MSE: 0.004995
6. MVAE reconstruction sample mean MAE: 0.039733

What "reconstruction" means here:

1. Take a real motion segment from the validation set.
2. Encode the future motion into MVAE latent.
3. Decode it back into motion using the same model.
4. Compare decoded motion against the original future motion.

So reconstruction checks whether the MVAE has learned a good motion representation. It is not the same thing as DAR text-driven generation.