Skip to content

pi0.5 keypoint action path + held-out-operator split for the fold-freeform fit tests: keypoints_wristframe_pi transform (138-D wrist-first keypoints + cartesian ee_pose prompt proprio), HumanBimanualKeypoints converter, PI resizes openpi action projection - #603

Draft
aidang3019 wants to merge 11 commits into
aidan/fold-freeform-framesfrom
aidan/fold-freeform-keypoints

Conversation

@aidang3019

Copy link
Copy Markdown
Contributor

pi0.5 keypoint action path + held-out-operator split for the fold-freeform fit tests: keypoints_wristframe_pi transform (138-D wrist-first keypoints + cartesian ee_pose prompt proprio), HumanBimanualKeypoints converter, PI resizes openpi action projections for action_dim!=32, opsplit data configs (cam6d/wrist6d/kp), keypoint evaluator/viz, norm + train launchers, CPU smoke

smoke: CPU keypoint smoke needs 1-process DDP (find_unused_parameters) — pl_model on_fit_start barriers; add working smoke_keypoints_cpu_ddp launcher

dataset v2: fold_clothes-derivative task-name rule (fold+clothes, no packaging; 492 eps / 11.2 h / 8 tasks) + operator split v2 (27 held-out ops, 8.7%)

revert dataset to folding_clothes only (v1: 462 eps / 10.45 h, operator split 25 held-out ops) per Aidan — drop the v2 task-name-rule expansion

opsplit launchers: validate every 20 epochs instead of 50 (viz still every 100, ckpt every 50)

Co-Authored-By: Claude Fable 5.1 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_012AaLp3ZC4xVapZZvE5Pd4Z

opsplit launchers: 72h wall, exclude the two thermally throttled H200 nodes

Co-Authored-By: Claude Fable 5.1 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_012AaLp3ZC4xVapZZvE5Pd4Z

opsplit: hold out a full 10% of frames by operator (was 8%)

Co-Authored-By: Claude Fable 5.1 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_012AaLp3ZC4xVapZZvE5Pd4Z

eval: palm-origin cam-frame metric for keypoint runs (PIEvalKeypointPalm) + eval-only launcher for the epoch-49 checkpoint

Co-Authored-By: Claude Fable 5.1 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_012AaLp3ZC4xVapZZvE5Pd4Z

eval launcher: typed h200 gres (untyped gres over a partition list landed on a V100)

Co-Authored-By: Claude Fable 5.1 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_012AaLp3ZC4xVapZZvE5Pd4Z

eval: offline launcher for cartesian opsplit checkpoints (same path as the keypoint palm eval)

Co-Authored-By: Claude Fable 5.1 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_012AaLp3ZC4xVapZZvE5Pd4Z

eval launcher: per-job inductor/triton cache dirs (same-node autotune race)

Co-Authored-By: Claude Fable 5.1 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_012AaLp3ZC4xVapZZvE5Pd4Z

Aidan Gao and others added 11 commits September 1, 2026 13:56
…eform fit tests: keypoints_wristframe_pi transform (138-D wrist-first keypoints + cartesian ee_pose prompt proprio), HumanBimanualKeypoints converter, PI resizes openpi action projections for action_dim!=32, opsplit data configs (cam6d/wrist6d/kp), keypoint evaluator/viz, norm + train launchers, CPU smoke
…) — pl_model on_fit_start barriers; add working smoke_keypoints_cpu_ddp launcher
…packaging; 492 eps / 11.2 h / 8 tasks) + operator split v2 (27 held-out ops, 8.7%)
…or split 25 held-out ops) per Aidan — drop the v2 task-name-rule expansion
…every 100, ckpt every 50)

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AaLp3ZC4xVapZZvE5Pd4Z
… nodes

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AaLp3ZC4xVapZZvE5Pd4Z
…alm) + eval-only launcher for the epoch-49 checkpoint

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AaLp3ZC4xVapZZvE5Pd4Z
…nded on a V100)

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AaLp3ZC4xVapZZvE5Pd4Z
…s the keypoint palm eval)

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AaLp3ZC4xVapZZvE5Pd4Z
… race)

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AaLp3ZC4xVapZZvE5Pd4Z

Copy link
Copy Markdown
Contributor Author

Stacked on aidan/eval_gating. pi0.5 fit tests on the mecka freeform folding_clothes set (462 eps / 10.45 h) with a held-out-operator split (31 operators / 10.2 % of frames): cam-frame 6D vs wrist-frame 6D vs a new 138-D MANO-keypoint action path.

  • fold-freeform-frames: cam6d/wrist6d data configs, cam-frame train_viz evaluator, CPU norm-stats + 2xH200 launchers.
  • fold-freeform-keypoints: keypoints_wristframe_pi transform (+ cartesian ee_pose prompt proprio), HumanBimanualKeypoints converter, PI resizes the openpi action projections for action_dim != 32, keypoint evaluator/viz, opsplit configs, palm-origin offline evaluator (PIEvalKeypointPalm) and eval-only launchers.
    Runs: 12707852 (wrist) / 12707853 (cam) / 12707854 (keypoints) on PACE; W&B rl2-group/zarr_test *_fold_clothes_freeform_fit_test_flat5e-5_opsplit.
    Path-matched offline result at epoch 99: palm-origin cam-frame MSE 2.80e-3 (wrist) / 2.89e-3 (cam) / 2.87e-3 (keypoints) m2 - parity within 3 %.

🤖 Generated with Claude Code

https://claude.ai/code/session_012AaLp3ZC4xVapZZvE5Pd4Z

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant