Data Wrangling for ML

Who: ML Engineers, Data Scientists
Output: Training/evaluation datasets Pixeltable is your system of record—all data, cached results, and references stay in sync.

Data Lifecycle

1. Acquire Data
2. Enrich & Annotate
3. Curate
4. Share & Export

Ingest

Load from any source: import_csv(), import_parquet(), HuggingFace, S3/GCS/Azure, RDBMS via Python DB API

Import from S3

Load images/videos from cloud storage

Import HuggingFace

Load datasets from HuggingFace Hub

Explore

Statistics & sampling: select(), .sample(), .head()

Data Sampling

Sample and filter large datasets efficiently

Enrich

Transform & extract: add_computed_column(), FrameIterator, DocumentSplitter

Extract Video Frames

Process video into frame-level data

Transcribe Audio

Audio to text with Whisper

Pre-Annotate

Model-in-the-loop: Auto-generate labels with AI models

Object Detection: yolox.yolox(), huggingface.detr_for_object_detection()
Vision LLMs: openai.vision(), anthropic.messages(), gemini.messages()
Classification: huggingface.image_classification()

Object Detection

Run YOLOX detection on images

Vision Batch Analysis

Analyze images with GPT-4o

Annotate

Human-in-the-loop: Refine labels with human annotatorsLabel Studio sync, FiftyOne export, add_embedding_index() for curation search

Label Studio Integration

Sync annotations bidirectionally

FiftyOne Export

Visualize and curate datasets

Model-in-the-loop vs Human-in-the-loop: Use pre-annotation to generate initial labels with AI models, then refine with human annotators. Pixeltable keeps both in sync—model outputs and human corrections live in the same table.

Search & Filter

Find similar examples with embedding search, filter by quality metricsadd_embedding_index(), .similarity(), .where(), .order_by()

Similar Image Search

Find visually similar samples

Semantic Text Search

Search by meaning, not keywords

Experiment

Test transformations before committing: Run SELECT to preview results on samples before adding computed columns

# Test on 5 rows first (no storage cost)
t.select(t.image, new_label=my_classifier(t.image)).head(5)

# Happy? Commit to full dataset
t.add_computed_column(new_label=my_classifier(t.image))

Iterative Development

Test UDFs and expressions before committing

Snapshot

Version control: create_snapshot(), create_view(), history(), lineage tracking

Version Control Guide

Track changes and revert to previous states

Why curate? ML models are only as good as their training data. Use Pixeltable’s search and filtering to find edge cases, remove duplicates, balance classes, and iterate on your data quality before export.

End-to-End Examples

Object Detection Pipeline

Complete workflow: ingest video → extract frames → detect objects → export

Audio Transcription Pipeline

Transcribe and analyze audio at scale

Structured Vision Output

Extract structured data from images with GPT-4o

Generate Captions

Auto-generate image descriptions

Use Cases

Deployment

Pixeltable Cloud

Data Lifecycle

Import from S3

Import HuggingFace

Data Sampling

Extract Video Frames

Transcribe Audio

Object Detection

Vision Batch Analysis

Label Studio Integration

FiftyOne Export

Similar Image Search

Semantic Text Search

Iterative Development

Version Control Guide

Data Sharing

Export to PyTorch

Data Interoperability

End-to-End Examples

Object Detection Pipeline

Audio Transcription Pipeline

Structured Vision Output

Generate Captions

Use Cases

Deployment

Pixeltable Cloud

​Data Lifecycle

Import from S3

Import HuggingFace

Data Sampling

Extract Video Frames

Transcribe Audio

Object Detection

Vision Batch Analysis

Label Studio Integration

FiftyOne Export

Similar Image Search

Semantic Text Search

Iterative Development

Version Control Guide

Data Sharing

Export to PyTorch

Data Interoperability

​End-to-End Examples

Object Detection Pipeline

Audio Transcription Pipeline

Structured Vision Output

Generate Captions

Data Lifecycle

End-to-End Examples