NVIDIA's Open-Source DALI Reaches Version 1.0

Announced nearly three years ago by NVIDIA as one of their open-source projects was the DALI library for GPU-accelerated data augmentation and image loading. The DALI library today reached the v1.0.0 milestone.

NVIDIA DALI is summed up as a data loading library with a focus on data loading and pre-processing for deep learning software. DALI provides various building blocks particularly around image, video, and audio processing. Of course, the GPU-accelerated library is optimized for NVIDIA's software/hardware architecture. DALI allows more of the data loading and pre-processing traditionally managed by the CPU to instead be handled by the GPU in a more efficient manner.

With DALI 1.0 there is improved API documentation, several new operators, new Python modules, enabling additional codecs and demuxers from FFmpeg, many bug fixes, and dozens of other improvements.

NVIDIA's DALI does support CPU-based execution while the GPU paths are written for CUDA. DALI 1.0 is compatible across deep learning frameworks like TensorFlow, PyTorch, MXNet, and PaddlePaddle.

More details on NVIDIA's DALI 1.0 open-source library via GitHub.