We have released Neural Network Libraries v1.31.0!
nnabla-rl v0.12.0, nnabla-nas v0.13.0 and nnabla-browser v0.2.0 are also released!

Spotlight

nnabla-rl v0.12.0 is released!

nnabla-rl v0.12.0 is released! In v0.12.0, we have added latest deep RL algorithms such as QRSAC and REDQ! We have also added optimal control algorithms iLQR and MPPI in this new release. From v0.12.0, visualization of training graphs and training curves with nnabla-browser is also supported.

Try nnabla-rl with:

$ pip install nnabla-rl

Check also the release note of nnabla-rl for details.

[XAI]Attention Branch Network (ABN) example

We have implemented colab interactive demo for Attention Branch Network, which extends a response-based visual explanation and accuracy improvement. It is realized by applying the focus area obtained from explanation to Attention mechanism.

Name	Notebook	Task
Attention Branch Network		Visualization

Add tutorials for fairness in machine learning

We have added two colab demos to our fairness tutorial for bias mitigation algorithm. Prejudice Remover Regularizer mitigates bias on training, and Rejection Option-based classification mitigates bias on post-processing.

Name	Notebook	Task
Prejudice Remover Regularizer for Images		Model Bias Check and Mitigation
Rejection Option-based Classification for image		Bias mitigation

Improve Diffusion Probabilistic Model (DPM) example

We have added faster sampling methods and conditional generation methods in Diffusion Model exmample. Please see the PRs below for more details.
– ODE-based sampler for faster inference
– class conditional and Cascaded DPM

nnabla

Core

• Add narrow function for NdArray (CPU / GPU)

OP layer

• Error function (erf) implementation (CPU / GPU)
• Add TopK support to ONNX importer

Bugfix

• implement drop_last option for slice
• fix numpy indice error due to numpy version update
• add code to allow unlink variable
• Support compile nnabla-ext-cuda with cuTENSOR in non-docker environment
• Correct macro error in convolution.c

Build

• add tensorboard version dependency of file format converter
• build android target with dynamic link
• Sync api level version from NNabla.
• Sync api_level version from NNabla for TopK

Format Converter

• Add TopK support to ONNX importer
• Add GlobalMaxPool, RandomNormal, RandomUniform support to ONNX importer
• Fix issues in ONNX importer
• restore the CL-initializer in legacy_nnp_graph.py to graph_def.py

Examples

• add example of infer semantic segmentation by cpp
• Add ODE based samplers

Known Issues

• FFT/iFFT pytest cases fails on Win32+CUDA11.4 Cuda backend

nnabla-RL

Bugfix

• Fix algos. Properly apply grad clip and weight decay
• Correct variable to use during rnn training
• Check np_random instance and use correct randint alternative
• Fix pendulum-env render
• Fix ScreenRenderEnv to support gym 0.25.0

Core

• Run PPO on single process when actor num is 1
• Add qrsac algorithm
• Add REDQ algorithm
• Update to support discrete tuple
• Add icra2018 qtopt
• Add goal_env module
• Add PPO tuple state support
• Add iLQR and LQR
• Add mppi
• Add ddp
• Add gmm and Update gaussian

Utility

• Support nnabla-browser

Documentation

• Fix module path of sac
• Improve README with graph visulization feature with nnabla-browser

Build

• Extend github build timelimit to 5 minutes

nnabla-browser

Bugfix

• Add rogue parameter support
• Add rogue parameter support (support nnabla-rl)

nnabla-NAS

Core

• Redefine the batch size definition
• Rework config and hparams

Bugfix

• Fix wrong path to working directory when saving and loading files
• Remove the duplicate processes of saving learned weights in OFA searcher
• Fix KeyError: ‘comm’.
• [OFA] Add loss_weights arg in configs
• Fix OFA valid/test
• Fix imagenet datapath
• Fix wrong working directory for OFAResnet50 and OFAXception search spaces

Build

• Add dependencies for RTD documentation build.
• Correct the package installation files.