JAX Data Loading: Using the Grain Dataset API for Simple and Declarative Data Processing

From the creator

Accelerators are getting faster, but is your data loading keeping up? In this video, we explore the Grain Dataset API, a powerful Python library designed to optimize data processing for machine learning. Learn how to build efficient, deterministic data pipelines that ensure your accelerators aren't left waiting. Dive into the chaining syntax for transformations—including mapping, shuffling, filtering, and batching. You'll also discover how to preserve random access for easy debugging and how to implement robust, asynchronous checkpointing with Orbax to save your data loading state alongside your model. Chapters: 0:00 - The Data Loading Bottleneck 0:27 - Recap: Grain & DataLoader 0:58 - The Grain Dataset API Overview 1:44 - Supported Data Sources (ArrayRecord, TFDS, Parquet) 2:02 - Transformation Pipeline: Shuffle, Map, Filter, Batch 2:33 - Code Example: Filtering News Headlines 3:12 - Checkpointing with get_state and set_state 3:56 - Asynchronous Checkpointing with Orbax 5:01 - Next Steps & Keras Hub Resources: Grain GitHub Repository → https://goo.gle/4rpUDdN Grain Documentation→https://goo.gle/4qvLEY5 Orbax documentation→ https://goo.gle/4jMGmVC Hear about Grain from the Engineer Lead →https://goo.gle/45XoQID Ready to load up some models? Check out this video about using Hugging Face Hub with KerasHub → https://goo.gle/4sRdy2l Subscribe to Google for Developers → https://goo.gle/developers Speaker: Yufeng Guo, Products Mentioned: Keras, Gemma, JAX

Published 2026-01-22 · Watch on YouTube ↗
Availability and playback restrictions are controlled by the original channel.