---

name: data-utilities description: "Use Towhee DataCollection, Entity, DataLoader, media wrappers, display, serialization, and debug visualizer/profiler outputs." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Towhee Data Utilities

Use this sub-skill when the task is about converting Towhee runtime results into inspectable rows, moving data into a pipeline, wrapping image/audio/video arrays, serializing displayed data, or reading pipeline debug visualizer/profiler outputs.

Route here for

• towhee.DataCollection(...), towhee.datacollection.DataCollection, Entity, to_list(), to_dict(), from_dict(), copy(), and show().
• towhee.DataLoader(data_source, parser=None, batch_size=None) for iterable or callable data sources.
• towhee.types.Image, AudioFrame, VideoFrame, towhee.types.image_utils.from_pil, to_pil, and color conversion helpers.
• RuntimePipeline result conversion to DataCollection or list/dict rows.
• RuntimePipeline.debug(..., profiler=True, tracer=True) outputs, Visualizer, DataVisualizer, PerformanceProfiler, node data tables, and Chrome trace dumps.

Route elsewhere

• Pipeline node construction, schemas, RuntimePipeline.batch(), debug() call options, and AutoConfig belong to the sibling pipeline-programming sub-skill.
• Operator registration, Hub operators, and CLI template generation belong to operator-hub-and-cli.
• Service startup, HTTP/GRPC, Docker, and Triton deployment belong to serving-and-triton.
• PyTorch training loops, NNOperator.train(), Trainer, and model-zoo execution belong to training-and-models.

Start with these bundled references

1. Read references/data-structures.md for DataCollection, Entity, DataLoader, media wrapper APIs, serialization contracts, and safe usage examples.
2. Read references/visualization-and-profiling.md for visualizer, tracer, and profiler output patterns.
3. Read references/troubleshooting.md when conversion, display, PIL/numpy, loader, or profiler behavior is surprising.
4. Run scripts/data_collection_smoke.py in an environment where Towhee is installed to validate the core data utilities without network, model downloads, or GPUs.

Minimal decision flow

• If the input is a pipeline result queue, wrap it once with towhee.DataCollection(result) before accessing rows repeatedly.
• If you need portable/debuggable rows, call dc.to_dict() and reconstruct with DataCollection.from_dict(saved_dict).
• If you are feeding many records into a pipeline, prefer DataLoader with a parser and, when using RuntimePipeline.batch, a positive batch_size.
• If the data are media arrays, keep the Towhee wrapper metadata (mode, sample_rate, timestamp, key_frame) aligned with the numpy shape and downstream library expectations.
• If the task asks what happened inside a pipeline run, use debug(..., profiler=True, tracer=True) and inspect v.result, v.profiler, and v.tracer rather than rebuilding nodes here.