part-aware-vqvae-overlap-top3-20260529

Part-Aware VQ-VAE Overlap Top3 20260529

Public checkpoint bundle for the part-aware-vqvae VQ-VAE rerun vq_overlap_top3_20260529.

Source workspace at export time: . Training output directory: ./output/vq/2026-05-29-11-07-24_vq_overlap_top3_20260529 Git commit recorded at export time: 82d4a3f538ae8e1fc623b7c8694ba32111245062

Contents

• weights/: all saved training checkpoints: best FID, best diversity, best Top1, best Top3, best matching score, and last.
• config/training_config.json: parsed JSON config from the training log plus launch command.
• config/skeleton_partition.json: overlap body-part partition used by this run.
• config/training_eval_summary.json: parsed training-time eval summary.
• logs/: run.log, launcher stdout/stderr log, and TensorBoard event file.
• code/: minimal code snapshot needed for loading/evaluation/reproduction, plus git status/diff.
• manifest.sha256: SHA256 checksums for every uploaded file.

Reproduce Launch

cd .
conda activate tlcontrol
CUDA_VISIBLE_DEVICES=0 python -u train_vq.py --dataname t2m --seed 123 --exp-name vq_overlap_top3_20260529 --nb-code 128 --partition-file ./partition_analysis/skeleton_partition.json

Important: this is the overlap-partition VQ-VAE (partition_analysis/skeleton_partition.json), not the MMM hardcoded/default-partition checkpoint.

Main Training-Time Metrics

See METRICS.md for the parsed table. Best FID checkpoint:

• weights/net_best_fid.pth
• iter: 165000
• FID: 0.0153
• R@1/R@2/R@3: 0.5153 / 0.7021 / 0.7965
• matching: 2.9400
• diversity: 9.6057

Caveat

The metrics bundled here are training-time eval metrics from run.log; they are not a formal offline 5-seed evaluation.