OpenGait/docs/scoliosis_training_change_log.md

Scoliosis Training Change Log

This file is the single run-by-run changelog for Scoliosis1K training and evaluation in this repo.

Use it for:

• what changed between runs
• which dataset/config/checkpoint was used
• what the resulting metrics were
• whether a run is still in progress

Conventions

• Add one entry before launching a new training run.
• Update the same entry after training/eval completes.
• Record only the delta from the previous relevant run, not a full config dump.
• For skeleton-map control runs, use plain-text ScoNet-MT-ske naming in the notes even though the code class is ScoNet.

Runs

Date	Run	Model	Dataset	Main change vs previous relevant run	Status	Eval result
2026-03-07	DRF	DRF	Scoliosis1K-drf-pkl-118	First OpenGait DRF integration on paper 1:1:8 split using shared OpenGait skeleton/PAV path	complete	58.08 Acc / 78.80 Prec / 60.22 Rec / 56.99 F1
2026-03-08	DRF_paper	DRF	Scoliosis1K-drf-pkl-118-paper	More paper-literal preprocessing: summed/sparser heatmap path, dataset-level PAV normalization, body-prior path refinement	complete	51.67 Acc / 72.37 Prec / 56.22 Rec / 50.92 F1
2026-03-08	ScoNet_skeleton_118	ScoNet-MT-ske control	Scoliosis1K-drf-pkl-118-paper	Plain skeleton-map baseline on the paper-literal export to isolate DRF vs skeleton-path failure	complete	38.85 Acc / 61.23 Prec / 46.75 Rec / 35.96 F1
2026-03-08	ScoNet_skeleton_118_sigma8	ScoNet-MT-ske control	Scoliosis1K_sigma_8.0/pkl	Reused upstream/default sigma-8 heatmap export instead of the DRF paper-literal export	complete	36.45 Acc / 69.17 Prec / 43.82 Rec / 32.78 F1
2026-03-08	ScoNet_skeleton_118_sigma15_bs12x8	ScoNet-MT-ske control	Scoliosis1K-drf-pkl-118-sigma15	Lowered skeleton-map sigma from 8.0 to 1.5 to tighten the pose rasterization	complete	46.33 Acc / 68.09 Prec / 51.92 Rec / 44.69 F1
2026-03-09	ScoNet_skeleton_118_sigma15_joint8_sharedalign_2gpu_bs12x8	ScoNet-MT-ske control	Scoliosis1K-drf-pkl-118-sigma15-joint8-sharedalign	Fixed limb/joint channel misalignment, used mixed sigma limb=1.5 / joint=8.0, kept SGD	complete	50.47 Acc / 69.31 Prec / 54.58 Rec / 48.63 F1
2026-03-09	ScoNet_skeleton_118_sigma15_joint8_limb4_adamw_2gpu_bs12x8	ScoNet-MT-ske control	Scoliosis1K-drf-pkl-118-sigma15-joint8-sharedalign-limb4	Rebalanced channel intensity with limb_gain=4.0; switched optimizer from SGD to AdamW	complete	48.60 Acc / 65.97 Prec / 53.19 Rec / 46.41 F1
2026-03-09	ScoNet_skeleton_118_sigma15_joint8_sharedalign_nocut_adamw_1gpu_bs8x8	ScoNet-MT-ske control	Scoliosis1K-drf-pkl-118-sigma15-joint8-sharedalign	Switched runtime transform from BaseSilCuttingTransform to BaseSilTransform (no-cut), kept AdamW, reduced 8x8 due to 5070 Ti OOM at 12x8	interrupted	superseded by proxy route before eval
2026-03-09	ScoNet_skeleton_118_sigma15_joint8_sharedalign_nocut_adamw_proxy_1gpu	ScoNet-MT-ske proxy	Scoliosis1K-drf-pkl-118-sigma15-joint8-sharedalign	Fast proxy route: no-cut, AdamW, 8x8, total_iter=2000, eval_iter=500, test_seq_subset_size=128	interrupted	superseded by geometry-fixed proxy before completion
2026-03-10	ScoNet_skeleton_118_sigma15_joint8_geomfix_proxy_1gpu	ScoNet-MT-ske proxy	Scoliosis1K-drf-pkl-118-sigma15-joint8-geomfix	Geometry ablation: aspect-ratio-preserving crop+pad instead of square-warp resize; AdamW, no-cut, 8x8, total_iter=2000, eval_iter=500, fixed test subset seed 118	complete	proxy subset unstable: 500 24.22/8.07/33.33/13.00, 1000 60.16/68.05/58.13/55.25, 1500 26.56/58.33/35.64/17.68, 2000 27.34/63.96/37.02/20.14 (Acc/Prec/Rec/F1)

Current best skeleton baseline

Current best ScoNet-MT-ske-style result:

• ScoNet_skeleton_118_sigma15_joint8_sharedalign_2gpu_bs12x8
• 50.47 Acc / 69.31 Prec / 54.58 Rec / 48.63 F1

Notes

• ckpt/ScoNet-20000-better.pt is intentionally not listed here because it is a silhouette checkpoint, not a skeleton-map run.
• DRF runs are included because they are part of the same reproduction/debugging loop, but this log should stay focused on train/eval changes, not broader code refactors.
• The long ScoNet_skeleton_118_sigma15_joint8_sharedalign_nocut_adamw_1gpu_bs8x8 run was intentionally interrupted and superseded by the shorter proxy run once fast-iteration support was added.
• The geometry-fixed proxy run fit the train split quickly but did not produce a stable proxy validation curve, so it should not be promoted to a full 20k run.