Released v1.0.9! Support CUDA10, add various layers, improve ONNX support etc.

We have released nnabla packages v1.0.9 a few days ago. The main features updated from v1.0.0 (previous blog post) are described as following:

Python

• Python packages compatible with CUDA10 are now available at PyPI
• Python packages compatible with distributed GPU training are available at PyPI
• Dynamic Loss Scaling utility for Mixed Precision Training
• Truncated BPTT in dynamic graph
• Numpy-like array indexing
• Add an optional argument for specifying initial parameter values in batch normalization

C++

• Command line interface for C++-standalone training

Addition and improvement of layer functions

• Image resize function Interpolate (2D bilinear only so far)
• Add inplace option for LeakyReLU
• Add numpy.arange-like function Arange
• Add Sort function
• Adding options to Max/Min function to get outputs compatible with argmin/argmax（Min, Max）
• Add deterministic option to CUDNN Convolution
• Add reflection option to Pad

Model reduction features

• Add Prune function for weight pruning
• Add regularization option for CP decomposition

Format converters

• Improved ONNX support
  • Operator import coverage
  • Operator export coverage

Utilities

• Drawing computation graph SimpleGraphViewer
• Plotting training logs produced by Monitors: Command line|Python API

Baseline model implementation: nnabla-examples

• Metric-based meta learning examples：Prototypical Networks, Matching Networks
• Improving YOLOv2 training example：Significant training time reduction, resolved numerical instability during training, code cleanup, logging loss and training time, modified evaluation code to be Python 2/3 compatible
• Add ShuffleNet example
• Add ShiftNet example
• Add Dockerfile