We have released Neural Network Libraries v.1.0.17! Added Xception model, STFT, and synced batch normalization!

Spotlight

STFT and inverse STFT

Short-form Fourier Transformation is a popular technique frequently used in speech/audio processing.

Synchronized Batch Normalization

Synchronized batch normalization is useful for distributed training with multiple GPUs.

Xception

Current Lineup of pre-trained models:
* ResNet: 18, 34, 50, 101, 152 layers available
* MobileNet: Both v1 and v2 available
* SENet: 154 layers available
* SqueezeNet: v1.1
* VGG: 11, 13, 16 layers available
* DenseNet: 161 layers available
* Network in Network
* Inception: v3
* Xception

Other Updates

Function Layers

• Various Activation Funcs (hard_sigmoid, hard_tanh, log_sigmoid, relu6, softplus, softsign, tanh_shrink, sinc)
• Implement assign function

• gather_nd and selection by index variable (advanced indexing)

Utilities

• Add tests for min and max function when returning indices.
• Improve CLI train, profile and forward
• Add variable name option to dump command.

Examples

• Synchronized Batch Normalization Example
• ImageNet training with DALI Data Iterator

Build

• Support ubuntu18.04 for multi-gpu build

• Fix build and create archive

• Add missing ‘cstring’ include

Bug Fix

• Fix LSTM Memory Leakage Problem
• Fix access violation in C++ MNIST training example
• Fix bugs in min and max functions when used to return indices.