Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

srb model — Manage Model Artifacts

Package checkpoints as self-contained SRB model artifacts with SRB identity, metrics, framework restore metadata, and optional dataset provenance, including from consumed-demo runs.

Usage

srb model <action> [options]

Actions

ActionPurpose
create CHECKPOINT --output OUTPackage one checkpoint. Requires --algo, plus --task, --embodiment, and --action-mode unless inferred from --dataset.
from-run RUN --output OUTPackage a completed SRB run’s latest checkpoint; infer algorithm and task from metadata.json. Without --dataset or consumed-demo provenance, also pass --embodiment and --action-mode.
info PATHPrint manifest JSON.
validate PATHValidate manifest, checkpoint, and bundled config paths; --strict makes warnings errors.
card PATHGenerate a Hugging Face README.md model card; schema 3 artifacts bundle the canonical card without rewriting it.
push PATHUpload or dry-run Hugging Face publication.

Both package actions accept repeated --metric key=value and --metadata key=value. Metrics must be finite numbers; text, nan, and inf are rejected. Metadata values are strings.

Portable Restore

from-run preserves available restore context:

  • Dreamer and TD-MPC2 can include config.yaml.
  • RSL-RL checks run-root config.yaml, then .hydra/config.yaml’s agent section, then metadata.json’s run.agent_cfg.
  • SKRL checks .hydra/config.yaml’s agent section, then metadata.json’s run.agent_cfg.

Reject the last source if it contains redacted or non-finite placeholders. The error names the key path. Redaction covers secret-looking key names and credential strings (URL userinfo, credential headers, and Bearer values). Warnings show key paths only.

For SB3/SBX, packaging requires inventoried vecnormalize.pkl statistics when the source RL-Zoo config declares normalization. Missing or ambiguous source evidence can remain unknown. Verify normalization state before packaging.

These commands accept artifact directories wherever they accept checkpoint paths: srb agent eval, srb agent collect, srb agent teleop, srb real_agent eval, and srb real_agent validate.

Loading an artifact can execute code (pickled checkpoints and statistics, SKRL model expressions). Load only artifacts from trusted sources.

SB3/SBX preserves RL-Zoo normalization state:

  • Packaging:
    • Normalized runs require inventoried, hashed vecnormalize.pkl.
    • Non-normalized runs record that state.
    • Missing or contradictory context fails packaging.
    • Re-package artifacts without a normalization declaration from the original run before inference.
  • Inference:
    • Restores frozen statistics without normalizing evaluation rewards.
    • Dict observations and recurrent episode state retain the policy layout.

Publish with dataset provenance from --dataset, a recorded consumed-demo run, or --dataset-repo-id plus optional --dataset-revision.

Preparation, validation, and restore stream hashes under per-file and aggregate limits, rejecting oversized or changing files. Schema 1/2 restore copies checkpoints and configs to temporary storage, verifies hashes, and removes partial copies on failure.

Examples

# Package the latest checkpoint from a run and infer identity from the dataset
srb model from-run logs/excavation/rsl_rl_ppo/<run> \
  --dataset /data/srb_excavation_policy \
  --output /models/srb_excavation_rsl_rl_ppo

# Validate and inspect the artifact
srb model validate /models/srb_excavation_rsl_rl_ppo --strict
srb model info /models/srb_excavation_rsl_rl_ppo

# Dry-run the Hugging Face upload, including the generated README.md
srb model push /models/srb_excavation_rsl_rl_ppo --dry-run

# Reuse the packaged artifact for policy rollout
srb agent collect --env excavation --model /models/srb_excavation_rsl_rl_ppo