(WIP: do not review or read) Docs: failure mining operational guide#434
(WIP: do not review or read) Docs: failure mining operational guide#434eugenevinitsky wants to merge 2 commits into
Conversation
There was a problem hiding this comment.
Your free trial has ended. If you'd like to continue receiving code reviews, you can add a payment method here.
There was a problem hiding this comment.
Pull request overview
Adds an operational guide for the mine_failures workflow (capturing compact replays + generating an HTML triage UI) and links it from the existing README “Failure mining” section.
Changes:
- Add
docs/mining.mdwith a failure-mining walkthrough, CLI examples, cluster submission pattern, and viewer notes. - Add a README pointer linking to the deeper mining guide.
Reviewed changes
Copilot reviewed 2 out of 2 changed files in this pull request and generated 5 comments.
| File | Description |
|---|---|
| README.md | Adds a pointer to the new failure-mining operational guide. |
| docs/mining.md | New end-to-end guide for mine_failures, including threshold semantics, cluster usage, and viewer documentation. |
Comments suppressed due to low confidence (1)
docs/mining.md:103
pufferl.trainis referenced as if it were a module; the training entrypoint here ispuffer train .../python -m pufferlib.pufferl train ...(and the function ispufferlib.pufferl.train). Consider updating the wording to the correct command/module path.
`mine_failures` does not read the sibling `config.yaml` next to
`load_model_path` (only `pufferl.train` does). If the checkpoint was trained
with non-default `policy.*` or `rnn.*` dimensions (e.g. `input_size=128`,
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
| browser-viewable HTML index of episodes. Pairs with `pufferl.mine_failures` | ||
| and `pufferlib/mining_viz.py`. |
| Open the index in a browser: | ||
|
|
||
| ```bash | ||
| open ./failure_mining/baseline_011000/renders/index.html | ||
| ``` |
| per-step agent state, traffic state, and observation arrays for a single | ||
| episode. Bundles are produced C-side when `capture_compact_replay=True` is |
| reads the bundle and replays it in-browser on a top-down canvas, with optional | ||
| overlays for the agent's observed FOV, partner circle, goal route, and waypoint | ||
| markers. |
| - Frame scrubber + play/pause + speed control. | ||
| - Toggle observation overlay (FOV rectangle, partner circle, observed-entity | ||
| highlights, goal route, waypoint markers). | ||
| - Toggle road segment / road edge / lane line rendering. | ||
| - Map background (CARLA / nuPlan / Waymo road graph from the bundle's | ||
| embedded `simulation_mode`). |
4e8139a to
9deb766
Compare
There was a problem hiding this comment.
Your free trial has ended. If you'd like to continue receiving code reviews, you can add a payment method here.
9deb766 to
8441077
Compare
There was a problem hiding this comment.
Your free trial has ended. If you'd like to continue receiving code reviews, you can add a payment method here.
There was a problem hiding this comment.
Your free trial has ended. If you'd like to continue receiving code reviews, you can add a payment method here.
e3abf94 to
2e8cc74
Compare
There was a problem hiding this comment.
Your free trial has ended. If you'd like to continue receiving code reviews, you can add a payment method here.
2e8cc74 to
86d30e2
Compare
There was a problem hiding this comment.
Your free trial has ended. If you'd like to continue receiving code reviews, you can add a payment method here.
There was a problem hiding this comment.
Your free trial has ended. If you'd like to continue receiving code reviews, you can add a payment method here.
96c75e3 to
8d043a5
Compare
There was a problem hiding this comment.
Your free trial has ended. If you'd like to continue receiving code reviews, you can add a payment method here.
There was a problem hiding this comment.
Your free trial has ended. If you'd like to continue receiving code reviews, you can add a payment method here.
fc91378 to
1fd6b0b
Compare
There was a problem hiding this comment.
Your free trial has ended. If you'd like to continue receiving code reviews, you can add a payment method here.
There was a problem hiding this comment.
Your free trial has ended. If you'd like to continue receiving code reviews, you can add a payment method here.
638e41a to
21f3fe8
Compare
There was a problem hiding this comment.
Your free trial has ended. If you'd like to continue receiving code reviews, you can add a payment method here.
21f3fe8 to
8da7751
Compare
There was a problem hiding this comment.
Your free trial has ended. If you'd like to continue receiving code reviews, you can add a payment method here.
8da7751 to
9edd792
Compare
New docs/mining.md covering the mine_failures workflow:
- score_threshold semantics (default -inf saves nothing)
- the required --vec.backend Serial flag (pufferl's default
Multiprocessing backend forks workers post-torch-import and
deadlocks on CUDA)
- loading checkpoints with non-default policy.* / rnn.* dims
(mine_failures doesn't auto-merge the sibling config.yaml that
train() does)
- on-cluster submit_cluster.py pattern with --main override
- viewer features
README.md gains a short pointer paragraph at the end of the existing
Failure mining section.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Mirror the cluster_training.md change — setup_container.sh install already lands submitit in the venv, and sourcing the venv on login makes it importable. No --user bootstrap needed. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
9edd792 to
0ca867b
Compare
New `docs/mining.md` covering the failure mining workflow plus a short README pointer.
Chained on top of #433 (cluster training docs) — that PR un-ignores `docs/` in `.gitignore` and adds the `submit_cluster.py` singularity-wrap patch that the on-cluster mining example here depends on. After #433 merges, this PR rebases onto `emerge/temp_training` automatically.
What's in
`docs/mining.md` — `mine_failures` workflow, `score_threshold` semantics, the required `--vec.backend Serial` flag (the default Multiprocessing backend forks workers post-torch-import and deadlocks on CUDA), loading checkpoints with non-default `policy.` / `rnn.` dims (mine_failures doesn't auto-merge the sibling `config.yaml` that `train()` does), the on-cluster `submit_cluster.py` pattern with `--main` override, and viewer features.
`README.md` — short pointer paragraph at the end of the existing "Failure mining" section.
🤖 Generated with Claude Code