Accelerating ML Development with Multi-Modal Datasets: Python, Parquets, and Daft
summary
Mobileye converts multi-modal rows (ten ~640×480 images plus ~200 metadata fields) into Parquet with external image URIs on S3 and reads them with Daft’s out‑of‑core engine. On a 10‑terabit scale dataset Daft runs MapReduce aggregations much faster than TFRecord and—surprisingly—serves training reads faster than TFRecord too. Result: one development environment, one pipeline, faster validation (shift‑left), and ongoing features for sparse tensor compression and probability‑based subpopulation sampling to speed I/O and training.
AI-written summary · about
chapters (7)
- 0:00Introduction to multi-modal datasets
- 3:21Row vs columnar data formats
- 7:24Challenges with single-format workflows
- 9:58Introducing Daft library
- 11:45Performance comparison with Daft
- 13:10Benefits of unified pipeline
- 18:16Summary and Q&A
related talks
26mOne Pipeline to Rule Them All: Unifying Multimodal and AI Data Processing with Daft
23mDaft - The Distributed Python Dataframe for Complex Data
34mDeep Learning at Scale
01h 23mMultimodal Deep Learning in the Real World
30mBlazing fast I/O of data in the cloud with Daft Dataframes
14mSoftware 2.0 Needs Data 2.0: A New Way of Storing and Managing Data for Efficient Deep Learning
28mBeyond Parquet and ORC: Upgrading Data Infrastructure for Multi-modal AI with Lance Col...
23mAccelerate the Machine Learning Pipeline on Very Large Datasets