---

name: rich-media-summaries description: "Image, audio, video, histogram, PR-curve, text, and mesh summary workflows for tensorboardX." metadata: disco-role: operating disable-model-invocation: true license: MIT

rich-media-summaries

Use this sub-skill when the tensorboardX task is about non-scalar summary payloads.

Covers

• SummaryWriter.add_image() and add_images()
• SummaryWriter.add_image_with_boxes()
• SummaryWriter.add_figure()
• SummaryWriter.add_audio()
• SummaryWriter.add_video()
• SummaryWriter.add_text()
• SummaryWriter.add_histogram() and add_histogram_raw()
• SummaryWriter.add_pr_curve() and add_pr_curve_raw()
• SummaryWriter.add_mesh()
• direct tensorboardX.summary.* helpers for the same payload families
• tensorboardX.utils shape and video helpers used by these summaries

Route out

• scalar logging, writer lifecycle, hparams, purge_step, or event-file checks: ../logging-core/SKILL.md
• model graphs, ONNX/OpenVINO, and projector embeddings: ../graph-and-embedding-plugins/SKILL.md
• global writer, multiprocessing, S3/GCS, and Comet integrations: ../remote-and-parallel-integrations/SKILL.md

Read these references first

• references/api-reference.md: method signatures and summary builders.
• references/data-formats.md: tensor shapes, channel conventions, label handling, and optional dependency requirements.
• references/workflows.md: end-to-end payload recipes and the safest sequence for each summary family.
• references/troubleshooting.md: shape, dtype, and optional dependency failures.

Bundled smoke script

• scripts/tbx_media_summary_smoke.py: creates a tiny temporary run with images, boxes, histogram, PR curve, text, mesh, and optional figure/audio/video checks when the dependencies are installed.

Working rules

• Prefer a CPU-only tiny fixture unless the user explicitly wants a different backend; this sub-skill does not require GPU support.
• Use tensorboardX.summary helpers only when the caller needs direct protobuf construction or when SummaryWriter is too high-level.
• Treat moviepy, imageio, matplotlib, soundfile, and pillow as optional dependencies that are activated only for the payloads that need them.
• Keep payloads small and deterministic so the helper script remains safe to run in a fresh environment.
• If the user only needs scalar logging, route them back to logging-core instead of expanding this route.

Common user requests

• "How do I log a batch of images?"
• "Why is my audio summary empty or clipped?"
• "How do I log a matplotlib figure?"
• "What shape should my video tensor have?"
• "How do I encode a PR curve or mesh summary?"

What to avoid

• Do not depend on the original repository checkout.
• Do not point future agents to the source repo examples as runtime instructions.
• Do not treat a missing optional dependency as a failure for the whole skill when the user did not ask for that payload.