NVIDIA DALI is a GPU-accelerated library designed to eliminate data loading and pre-processing bottlenecks in deep learning workflows. It offers a collection of optimized building blocks for image, video, and audio data, enabling users to create portable, high-throughput data pipelines that can be seamlessly integrated with popular frameworks like TensorFlow, PyTorch, and PaddlePaddle.

How It Works

DALI addresses CPU-bound data processing by offloading operations to the GPU. It utilizes a custom execution engine optimized for throughput, featuring prefetching, parallel execution, and batch processing. This GPU-centric approach, combined with a flexible, functional Python API, allows for the creation of complex, multi-stage data augmentation and transformation pipelines that run efficiently, directly feeding data to the GPU for training or inference.

Quick Start & Requirements

• Install with pip install nvidia-dali-cuda120 or pip install --extra-index-url https://pypi.nvidia.com --upgrade nvidia-dali-cuda120.
• Requires NVIDIA driver supporting the target CUDA version (e.g., CUDA 12.x) and the corresponding CUDA Toolkit.
• Pre-installed in NGC containers for TensorFlow, PyTorch, and PaddlePaddle.
• Official Documentation: https://docs.nvidia.com/deeplearning/dali/user-guide/docs/index.html

Highlighted Details

• Supports numerous data formats including LMDB, TFRecord, COCO, JPEG, WAV, FLAC, and video codecs (H.264, VP9, HEVC).
• Accelerates common workloads like ResNet-50, SSD, and ASR models (Jasper, RNN-T).
• Enables direct data transfer via GPUDirect Storage and integrates with NVIDIA Triton Inference Server.
• Offers custom operator extensibility for user-specific needs.

Maintenance & Community

• Active development with a public roadmap available at https://github.com/NVIDIA/DALI/issues/5320.
• Contributions are welcomed following guidelines in CONTRIBUTING.md.

Licensing & Compatibility

• Licensed under Apache 2.0, permitting commercial use and integration with closed-source projects.

Limitations & Caveats

• Requires NVIDIA GPU hardware and specific CUDA toolkit versions for installation and operation. Building from source requires a compilation guide.