sim_data / fridge_m_v2 /SETUP.md
Ravenh97's picture
fridge_m_v2: initial SETUP.md
f9d1e6d verified
|
Raw
History Blame Contribute Delete
4.73 kB
# fridge_m_v2 sim datagen — setup record
Generated 2026-05-14 by an automated agent on viscam6 (4× RTX A5000), with
helper nodes viscam7 (4× A5000) and viscam3 (4× RTX 3090). Sim-only run;
no concurrent gen on this dataset. Push target:
`Ravenh97/sim_data:fridge_m_v2/`.
## Motivation
Regenerate fridge_m sim data after the user updated:
- The scene XML (`assets/scenes/ithor_minimal_fridge_m/FloorPlan1_physics.xml`)
— modified 2026-05-14 ~20:56 PDT.
- The robot init params and placement ranges in
`scripts/datagen/data_gen_fridge_m_2.sh` — modified 2026-05-14 ~21:02 PDT.
## Pipeline
Same chunked sim pipeline that worked for fridge_m / fridge_7 / 3drawers:
- Each chunk runs 4 sim shards on 4 GPUs writing to `sim_chunks/chunk_NNN/`.
- After all 4 shards exit (h5 locks released), `check_mismatch.py` asserts
`T == exo_mp4_frames == wrist_mp4_frames` for every traj.
- NO gen step (sim-only request).
Frozen scripts are at `/scr/ravenh/fridge_m_v2/scripts/` on each node so
mid-run edits to the source `data_gen_fridge_m_2.sh` cannot silently
change the run. The frozen `data_gen_v2.sh` only differs from the source
by absolutizing the `PIPELINE` path (snapshot lives under `/scr` so the
SCRIPT_DIR-relative path in the original would not resolve).
## v2 parameters (from frozen `data_gen_v2.sh`)
```
--scene_dataset ithor_minimal_fridge_m
--object_instance_index 1
--joint_index 1
--base_pose_sampling_radius_range 0.8 1.2
--base_pose_sampling_theta_range_rad -0.5 1.0
--robot_placement_rotation_range_rad 0.5
--robot_placement_exclusion_threshold 0.0
--robot_base_z 0.86
--robot_base_z_jitter -0.1 0.1
--robot_init_qpos_noise_range 0.4 0.4 0.4 0.4 0.4 0.4 0.4
--camera_extrinsics scripts/datagen/picked_camera_extrinsics.json
--camera_extrinsics_pos_noise 0.05
--camera_extrinsics_rot_noise_deg 5.0
--output_suffix _cam_rand
--record_depth, --record_segmentation, --record_robot_masked_rgb
--filter_for_successful_trajectories
```
Pipeline: `molmospaces/scripts/datagen/run_pipeline_record_state.py`
(records full mocap/qpos state per step; same wrapper as fridge_m_0.7_pd).
## Diff vs fridge_m (v1)
| param | v1 (data_gen_fridge_m.sh) | v2 (data_gen_fridge_m_2.sh) |
| ------------------------------------- | ------------------------- | --------------------------- |
| base_pose_sampling_radius_range | 0.75 1.1 | 0.8 1.2 |
| base_pose_sampling_theta_range_rad | -0.5 0.9 | -0.5 1.0 |
| robot_placement_rotation_range_rad | 0.2 | 0.5 |
| robot_base_z_jitter | 0.0 0.1 | -0.1 0.1 |
| robot_init_qpos_noise_range | (unset) | 0.4 × 7 |
| pipeline | run_pipeline_no_clip.py | run_pipeline_record_state.py|
Scene XML was also modified in place — fingerprints recorded below.
## Fingerprints
- `assets/scenes/ithor_minimal_fridge_m/FloorPlan1_physics.xml` md5 `83a5b0c9d5a28208343e70c2b1151af6`
- `assets/scenes/ithor_minimal_fridge_m/FloorPlan1_physics.json` md5 `1670a043ef6ddfd3c872d9fa221a7bbe`
- `scripts/datagen/data_gen_fridge_m_2.sh` md5 `86be8ec056ebbcb8668c2f80a4ed8eeb`
- frozen `data_gen_v2.sh` (absolute PIPELINE path) md5 `cf816d56054141e9ea8e7b51226623cd`
- molmospaces git HEAD `8003c7009d28746621f969ea44e96475f851d2cd`
## Node assignments
| node | GPUs | chunk range | SEED_BASE | STOP_AT_TRAJS | /scr free at launch |
| -------- | ----------------- | ----------- | --------- | ------------- | ------------------- |
| viscam6 | 4× A5000 (0-3) | 001-015 | 20000 | 370 | 5.1 TB |
| viscam7 | 4× A5000 (0-3) | 101-115 | 30000 | 370 | 3.3 TB |
| viscam3 | 4× RTX 3090 (0-3) | 201-210 | 40000 | 260 | 248 GB |
Expected per-chunk yield (from fridge_m_chunked memory): ~45-50 trajs.
Target ~1000 total.
## HF layout
```
Ravenh97/sim_data:fridge_m_v2/
SETUP.md <- this file
sim_chunks/
chunk_001/ ... <- viscam6
chunk_002/ ...
...
chunk_101/ ... <- viscam7
...
chunk_201/ ... <- viscam3
...
```
push_watchdog.py uploads (1 commit per push) every 50 finalized trajs;
skips h5 files until the final flush (then includes them).