AGENT SKILL

training-and-evaluation

Compose and debug Diffusion Policy Hydra training, evaluation, Ray multirun, checkpoint evaluation, and benchmark logging commands.

License 许可
MIT
最近更新
2026-10-08
Stars
★ 11
Forks
3

01 前言 Front Matter

name
training-and-evaluation
license
MIT
metadata
{ "disco-role": "operating" }
description
Compose and debug Diffusion Policy Hydra training, evaluation, Ray multirun, checkpoint evaluation, and benchmark logging commands.
disable-model-invocation
true

02 资源文件 3

03 SKILL.md 正文

Training and Evaluation

Use this sub-skill when you need to build or debug Diffusion Policy training and evaluation workflows.

Route here for

  • Single-seed Hydra training from train.py.
  • Checkpoint evaluation from eval.py.
  • Ray multirun orchestration from ray_train_multirun.py and ray_exec.py.
  • Offline aggregation of train_*/logs.json.txt from multirun_metrics.py-style logs.
  • Simulation benchmark routing for low-dim, image, hybrid, Robomimic, and video workspaces.
  • W&B, checkpoint, and output-tree diagnostics for training/evaluation runs.

Do not handle here

Operating workflow

  1. Choose the workspace config family and matching task config.
  2. Compose the resolved config with scripts/compose_experiment_config.py.
  3. Launch a single-seed train.py run or a Ray multirun.
  4. Inspect logs.json.txt, checkpoints, eval_log.json, and metrics/logs.json.txt.
  5. Use scripts/summarize_multirun_metrics.py to summarize finished train_* logs offline.

Core API names

  • BaseWorkspace.run() drives the training/evaluation lifecycle.
  • BaseWorkspace.save_checkpoint() and BaseWorkspace.load_checkpoint() manage checkpoint state.
  • policy.set_normalizer(...) and policy.predict_action(...) are the key policy hooks.
  • env_runner.run(policy) produces rollout metrics and videos.
  • JsonLogger, read_json_log, and TopKCheckpointManager shape the run logs.

Start here

Read the bundled references for command recipes, config structure, and troubleshooting:

04 同仓库更多 Skills