Artifacts & Checkpoints¶

Training runs produce artifacts: model checkpoints, training videos, and TensorBoard logs. All artifacts are stored in S3 and downloadable via the SDK or dashboard.

Artifact Types¶

Type	Description	Format
`checkpoint`	Model weights saved during training	`.pt` files
`video`	Training episode recordings	`.mp4` files
`tensorboard`	TensorBoard event logs	`events.out.tfevents.*`
`policy`	Final trained policy	`.pt` / `.onnx`
`summary`	Training summary JSON	`.json`

List Artifacts¶

# All artifacts for a run
artifacts = run.list_artifacts()
for a in artifacts:
    print(f"{a['filename']} ({a['artifact_type']}, {a['size_bytes']} bytes)")

# Filter by type
checkpoints = run.list_checkpoints()
videos = run.list_videos()

Download Checkpoints¶

Best Checkpoint¶

path = run.download_best_checkpoint("./checkpoints")
print(f"Saved to {path}")
# → ./checkpoints/model_5000.pt

Specific Checkpoint¶

path = run.download_checkpoint("model_2500.pt", "./checkpoints")

By S3 Key¶

artifacts = run.list_artifacts(artifact_type="checkpoint")
for a in artifacts:
    path = run.download_artifact(a['s3_key'], "./all-checkpoints")

Download All Results¶

# Download everything (checkpoints, videos, summaries)
results_dir = run.download_results("./results")

# Include TensorBoard logs
results_dir = run.download_results("./results", include_tensorboard=True)

Get Presigned URLs¶

If you need direct S3 access (e.g., for loading in a notebook):

artifacts = run.list_checkpoints()
url = client.get_artifact_url(run.slug, artifacts[0]['s3_key'])
# url is a presigned HTTPS URL valid for download

Videos¶

Enable video recording when submitting:

run = client.submit_training_run(
    name="with-video",
    task_name="Template-Go2-Standing-Direct-v0",
    code_url="https://github.com/canard-cloud/go2-standing-env",
    enable_video=True,
    video_interval=200,  # Record every 200 steps
    num_envs=4096,
    max_iterations=5000,
)

Videos are also viewable in the dashboard run detail page.