selfimprovingv2a / README.md
mwilcoxson's picture
All assets available: VP datasets, eval windows, expert videos, goal-transfer experts; full SHA256SUMS
e502842 verified
|
Raw History Blame Contribute Delete
3.6 kB
# Release assets
Checkpoints and data for the Mirror experiments. They are hosted on the Hugging Face dataset
[`mwilcoxson/selfimprovingv2a`](https://huggingface.co/datasets/mwilcoxson/selfimprovingv2a),
laid out exactly as this directory; download into `assets/` (see the main README). Only the
index files are tracked in git: this README, `assets.json` (the asset paths of every
manifest row), `goal_transfer_experts.json` and `SHA256SUMS` (verify a download with
`cd assets && sha256sum -c --ignore-missing SHA256SUMS`).
| Directory | Contents | Read by | Size | Status |
|---|---|---|---|---|
| `video_planners/mirror/<env>/` | Mirror VP (`meta.kind=mirror`): `codec.msgpack`, `vp.msgpack` (EMA flow), `meta.json` | `--config.planner=learned --config.learned_planner_path=...` (`train_mirror.py`, `train_supervised_idm.py`, `eval_random_idm.py`, `eval_plan_samples.py`) | 0.7 GB | available |
| `video_planners/grounding/<env>/` | Grounding's subgoal video planner (`meta.kind=grounding`) | `train_grounding.py --config.planner=learned` | 0.7 GB | available |
| `supervised_idms/<env>/` | `idm.msgpack`: a flow IDM (Mirror's architecture) trained on expert actions; `idm.json`: architecture | `train_supervised_idm.py --config.idm_path=...` | 1.4 GB | available |
| `experts/dmc/<env>/` | BRO policy at 1M steps: `policy.msgpack`, `policy.json` | `--config.expert_checkpoint=<dir>` | | available |
| `experts/ogbench/<env>/` | FQL agent: `flags.json`, `params_2000000.pkl` | `--config.expert_checkpoint=<dir>` | | available |
| `experts/atari/<game>/` | DQN (atari-agents, 50M steps): `manifest.json`, `weights.npz` | `--config.expert_manifest=<manifest.json>` | 0.3 GB (all experts) | available |
| `experts/ogbench_goals/<env>/task<k>/` | FQL experts of the other OGBench goals (`goal_transfer_experts.json`) | `eval_goal_transfer.py --expert_checkpoint=<dir>` | 1.1 GB | available |
| `eval_datasets/<env>/` | Labeled expert windows for the offline IDM audit (`labeled_eval_dataset`) | `train_mirror.py`, `train_random_idm.py`, `train_bco.py` | 7.6 GB | available |
| `expert_videos/<env>/{bco,laifo}.npz` | Expert videos (`expert_dataset`) | `train_bco.py`, `train_laifo.py` | 6.6 GB | available |
| `vp_datasets/<env>[_episodes]/` | Video-planner training datasets (`mirror/vp_data.py`) | `train_video_planner.py`, `train_supervised_idm.py --config.dataset_path=...` | ~3.3 TB | available |
There is one asset of each kind per task (24 tasks: 4 OGBench, 13 DMC, 7 Atari).
## How the planners and IDMs were trained
- Each planner was trained with `train_video_planner.py` at the default budgets, on its
task's dataset under `vp_datasets/`: a spatial codec for 250k steps and a latent flow for
10M updates, keeping the best-validation codec and the EMA flow with the best validation
loss. `meta.json` records the dataset (with the sha256 of its `dataset.json`) and the
steps it kept (`dataset` names the released copy under `vp_datasets/`).
- The Mirror VP trains on `vp_datasets/<env>/`. Grounding's subgoal planner
trains on `vp_datasets/<env>/` for DMC/Atari and on `vp_datasets/<env>_episodes/` for
OGBench.
- `vp_datasets/<env>/` holds expert demonstrations plus on-policy rollouts. Frames are 84x84
on DMC and Atari and OGBench's native 64x64. `<ogbench task>_episodes/` holds 1000 expert
episodes, each run to success plus 10 steps (`successes.npy` gives the first success).
- The supervised IDMs were trained on the same datasets with expert actions. They are
independent of the planners and are used only to evaluate a planner
(`train_supervised_idm.py`).