CodeFlow-HumanML3D / metadata /rvq_metrics.md
AmberJar's picture
Upload CodeFlow HumanML3D release weights
09154c8 verified

Training Metrics

These are training-time VQ-VAE evaluation metrics parsed from logs/run.log. No formal offline 5-seed evaluation is bundled here.

checkpoint / selector iter FID R@1 / R@2 / R@3 Matching Diversity
net_best_fid.pth / best_fid 165000 0.0153 0.5153 / 0.7021 / 0.7965 2.9400 9.6057
net_best_top1.pth / best_top1 210000 0.0219 0.5372 / 0.7081 / 0.8012 2.9300 9.6410
net_best_top3.pth / best_top3 60000 0.0366 0.5160 / 0.7154 / 0.8112 2.9063 9.3176
net_best_matching.pth / best_matching 160000 0.0426 0.5312 / 0.7114 / 0.8052 2.8848 9.5794
net_last.pth / last 300000 0.0209 0.5140 / 0.7114 / 0.7999 2.9161 9.3756

Final train line: iter 300000, Recons 0.02611, PPL 83.99, Commit 0.89492.