Johnny Fernandes
97 lines
3.2 KiB
Markdown
97 lines
3.2 KiB
Markdown
# Training pipeline
|
||
|
||
Two stages, strictly sequential:
|
||
|
||
```
|
||
sim demos (Strömbom on tracker output, K=4 frame stack)
|
||
│
|
||
▼
|
||
bc_pretrain.py ──► runs/bc (Strömbom-imitated MLP)
|
||
│
|
||
▼ KL-regularised PPO fine-tune
|
||
│
|
||
runs/rl (deployed `rl` mode — beats BC and Strömbom)
|
||
```
|
||
|
||
## Files
|
||
|
||
```
|
||
herding_env.py — Gymnasium env (LiDAR raycast + tracker by default)
|
||
bc_pretrain.py — MSE + cosine BC of (obs, action) demos into MlpPolicy
|
||
train_ppo.py — KL-regularised PPO fine-tune of a BC checkpoint
|
||
eval.py — multi-seed analytic / learned policy comparison
|
||
runs/ — checkpoints (whitelisted entries in top-level .gitignore)
|
||
```
|
||
|
||
## Setup
|
||
|
||
```
|
||
pip install -r requirements.txt
|
||
```
|
||
|
||
CPU is the default and recommended device — SB3 PPO with an MLP policy
|
||
of this size runs faster on CPU than GPU because the bottleneck is
|
||
rollout collection, not gradient compute.
|
||
|
||
## End-to-end pipeline
|
||
|
||
```bash
|
||
# 1. Sim demos with the active-scan + Strömbom teacher under LiDAR
|
||
# perception. K=4 frame stack so the MLP has temporal context.
|
||
python -m tools.collect_demos --teacher strombom \
|
||
--out training/demos.npz --seeds-per-n 15 --subsample 3 --frame-stack 4
|
||
|
||
# 2. Behaviour-clone.
|
||
python -m training.bc_pretrain --demos training/demos.npz \
|
||
--out training/runs/bc --epochs 60 --net-arch 512,512
|
||
|
||
# 3. KL-regularised PPO fine-tune of bc.
|
||
python -m training.train_ppo \
|
||
--bc training/runs/bc --out training/runs/rl \
|
||
--total-timesteps 1000000
|
||
|
||
# 4. Multi-seed eval (env-side, fast).
|
||
python -m training.eval --policy training/runs/rl \
|
||
--max-flock 10 --max-steps 15000 --n-seeds 10
|
||
```
|
||
|
||
`bc_pretrain.py` saves the **best-val_cos** snapshot, not the final
|
||
epoch — multi-modal teachers make training noisy and the last epoch is
|
||
often worse than an earlier one.
|
||
|
||
`train_ppo.py` loads BC weights into both a trainable policy and a
|
||
frozen reference, fixes `log_std` small, and adds `β · KL(π‖π_ref)` to
|
||
the loss so the policy can only move within a trust region around BC.
|
||
See the file header for hyperparameter rationale.
|
||
|
||
## Available analytic teachers
|
||
|
||
| Name | What it does | Notes |
|
||
|---|---|---|
|
||
| `strombom` | Strömbom 2014 — collect when flock is scattered, drive CoM otherwise | Default; works for n=1–10 under tight cohesion |
|
||
| `sequential` | Pick the sheep closest to the pen and drive only it | Alternative; needs loose-cohesion regime |
|
||
|
||
Both are wrapped at demo-collection time in
|
||
`herding/control/active_scan.py:ActiveScanTeacher`, which adds an
|
||
opening in-place rotation, walk-to-centre when the LiDAR sees
|
||
nothing, and near-sheep speed modulation (same modulation
|
||
`herding/control/modulation.py` applies to every dog mode at
|
||
inference).
|
||
|
||
## Evaluating analytic teachers directly
|
||
|
||
```
|
||
python -m training.eval --policy strombom --max-flock 10 --max-steps 15000 --n-seeds 10
|
||
python -m training.eval --policy sequential --max-flock 10 --max-steps 15000 --n-seeds 10
|
||
```
|
||
|
||
## Webots inference
|
||
|
||
```
|
||
tools/run_webots.sh 10 bc # or rl, strombom, sequential
|
||
```
|
||
|
||
The dog controller loads `runs/bc` for `bc` mode and `runs/rl` for
|
||
`rl` mode. Override with `HERDING_POLICY_DIR=…` for a specific
|
||
checkpoint.
|