Releasing Neural Network Libraries v1.0

Today (6/27) is the first anniversary of Neural Network Libraries’ open source release! And we are now announcing the release of version 1.0 on this special day. We would like to express our gratitude, from the bottom of our hearts, to everyone who has used, contributed to, and given meaningful feedbacks to Neural Network Libraries.

Since the release of version 0.9, the first version to be open-sourced, a great amount of new and powerful functionalities have been added and updated.

Ubiquity

• Compatibility with both Python 2 and 3
• Compatibility with Windows
• Compatibility with MacOS
• Inference Sample with C++
• Training Sample with C++
• Build with Android NDK
• C Runtime
• Mutual Converter for ONNX
• Python binary for each CUDA runtime version

High Performance

• High-speed performance with NVIDIA Volta TensorCore
• High-speed parallel training with NVIDIA NCCL2
• High-speed parallel training by gradient exchange during backpropagation

Coherence with Neural Network Console (NNC)

• Training/Inference with models from Neural Network Console

Usability

• Computation graph performance profile
• Reconnecting multiple computation graphs

Training and Inference of Various Recent Models

Repository with a collection of recent models is created.

• nnabla-examples

GANs

• DCGAN
• Pix2Pix
• Cycle GANs
• Progressive Growing GANs

Network compaction techniques

• Distillation
• Factorized layers
• Quantization
• Structured sparcity

Visual recognition models

• ImageNet Classification with ResNet
• Object Detection with YOLOv2

Misc.

• Capsule Network
• Penn Treebank language model training with LSTM

Functions

• Quantization Layer (fixed-point and power-of-2)
• Depthwise Convolution / Deconvolution
• FFT/IFFT
• … And 141 more Functions | Parametric Functions

We will be updating our blog periodically with more detailed description of the functionalities.

We would also like to inform the users that an official cloud version of Neural Network Console, a GUI-based deep learning development tool from Sony as well, is available. High-speed training with multi-GPUs is readily available, without hassles of setting up the environment. We will further promote the coherence with Neural Network Console in the development flow, so we highly recommend that the users give it a try.

• Neural Network Console

As with C runtime or ONNX converters, we will continue to further promote Neural Network Libraries’ status as a framework that supports the development flow in an end-to-end manner, from design to implementation and deployment of deep leaning.

Thank you again for your support!