openpi-v4-memory-artifacts

openpi v4 dual-bank memory policy: training artifacts

Companion to the ZJU-Walker/memory_project_v4 code repository. Files keep their project-relative paths: download with --local-dir <project root> and they land where openpi/src/openpi/shared/project_paths.py expects them.

Dataset (LeRobot, 70 episodes, 41.6 GB): kewalk123/bin_memory_0830_0831_v36_subtask revision bd97941eca402f8be854ee8a0a3bbad14df37292 -> data/lerobot/yam/bin_memory_0830_0831_v36_subtask/{data,meta}. Base model: gs://openpi-assets/checkpoints/pi05_base/params (downloaded by the trainer).

Files and SHA256

• data/0830_0831_episode_manifest_v36_frozen.json: 9085fe50d7b02ea65930f3647ce0413e0583a66d430484e06c60812c52af8442
• data/v4_fact_labels_0830_0831.json: 4b6027bf2cf43db992479709619e42ab1d1ddea792e0453eae6cf8091514d378
• data/0830_0831_episode_manifest_v36_frozen_block_confound.json: 06d2b5dffa3617d943a657837478ca5aa7568e35a1f84e147d0a14f9b39793bb
• data/0830_0831_episode_manifest_v36_frozen_e_visibility.json: 633e1a49b254b769d2aab60f9f4a0b11321648458ca298834132c120e4f84fca
• data/0830_0831_episode_manifest_v36_frozen_d_valid.json: ecc8ee474a1d5c107bee084d28a9352e12464acd365245b4a97e122dfba8c9d1
• v4/assets/pi05_yam_0830_0831_v36/yam/bin_memory_0830_0831_v36_subtask/norm_stats.json: 5535ea95ad7ed1edc399ba47e278285a1fdec02a589451cec0dc9003d458519c
• v4/assets/pi05_yam_0830_0831_v36/yam/bin_memory_0830_0831_v36_subtask/norm_stats_provenance.json: 36e0ab51d53272038e1b204c752b30fa6ab000096bff9ff6dccd605166188c58
• data/*/demo*/subtask_labels.json: 70 raw per-episode label files (each pinned by label_sha256 in the manifest)

The manifest and fact-label digests are pinned in openpi/src/openpi/training/config.py (memory_episode_manifest_sha256, memory_v4_fact_labels_sha256); the three manifest sidecars and the 70 label files are pinned inside the manifest itself. Training refuses to start if any of them differ.

Quick start

git clone [email protected]:ZJU-Walker/memory_project_v4.git && cd memory_project_v4
bash openpi/cluster_v4/coauthor/run_all.sh   # env -> data -> 8-GPU training (asks for a W&B key; WANDB=0 to skip)