Released Neural Network Libraries v1.23.0!

Thursday, November 25, 2021


Posted by shin

We have released Neural Network Libraries v1.23.0!

We have added new examples including implicit differentiable render (IDR) for 3D shape reconstruction, elastic weight consolidation (EWC) and cross-domain correspondence (CDC) for few-shot image generation, and many others! Plus, our format converter also supports ONNX opset 13 now!


Implicit differentiable renderer

We have implemented Implicit Differentiable Renderer (IDR). The goal of IDR is to extract an object surface, given multi-view posed images and corresponding masks with no 3D supervision. The method is comprised of the following:

  • Signed distance function as implicit function
  • Geometric initialization to initialize the network such that we have sphere-like initial geometry
  • Crafted ray tracing; combination of the ray and unit-sphere intersection, sphere tracing, and ray marching
  • Losses; RGB, mask, and Eikonal loss.

Add ONNX opset13 support

We have updated nnabla-converter to support both import and export for ONNX opset13, and have also updated support status for nearly 50 functions. Please check support status on our website.

Add EWC and CDC to StyleGAN2

Few-Shot Image Generation with EWC

Few-Shot Image Generation with CDC

Few-shot image generation has increasingly gained attention as a research topic, as it enables image generation from a limited amount of training samples, as opposed to thousands in conventional generation models. We have implemented two methods for few-shot image generation using StyleGAN2 as the model, with EWC (Elastic Weight Consolidation) and CDC (Cross-Domain Correspondence).

Pointnet for 3d object classification implementation

We have implemented PointNet for 3D object classification, which directly takes point clouds as input, without having to convert it to 3D voxel grids or images. Our implementation has been tested with ModelNet40 dataset.

Add GCN implementation

We have implemented Graph Convolutional Network (GCN), which performs a semi-supervised learning on graph-structured data. Following the paper, we use GCN to semi-supervised node classification of papers using Cora dataset, where we assume that only some of the papers are labeled and predict the subjects of the remaining papers.



Core Functionalities