Released Neural Network Libraries v1.36.0!

Friday, July 14, 2023

We have released Neural Network Libraries v1.36.0 and newly released Pixel-Guided-Diffusion!
Please see “Spotlight” for important changes.

Spotlight

Pixel Guided Diffusion

We have released a fine-grained image editing method using diffusion model such as pixel-guided-diffusion here!
https://github.com/sony/pixel-guided-diffusion

This method enables pixel-level precise real image editing using the user-provided manipulation of semantic segmentation map.
Both pytorch and nnabla version implementations are available.

Dataset DDPM

We also released the nnabla implementation of DatasetDDPM on nnabla-examples.
https://github.com/sony/nnabla-examples/tree/master/diffusion-models#ddpm-segmentation

This is a label-efficient semantic segmentation model using diffusion models.
You can train segmentation models with only a few-dozen pairs of annotated data with DatasetDDPM.

[Fairness] Preferential Sampling Algorithm (part1 / part2)

We have implemented colab interactive demo for the Preferential Sampling algorithm for Discrimination-free training sets, which effectively identifies and adjusts borderline objects based on their positive class probabilities, ensuring a fair and unbiased dataset.

Name	Notebook	Task	Example
Preferential Sampling		Dataset/Model Bias Check and Mitigation by Preferential Sampling

[Fairness] Support for the “Massage Data” algorithm on various datasets using data processor class

We have implemented colab interactive demo for the Massage Data algorithm it is a Preprocessing technique designed to mitigate bias in datasets.
This tutorial covers the step-by-step process of identifying promotion and demotion candidates, strategically relabeling objects to achieve fairness, and calculating discrimination metrics for discrimination-free datasets.

Name	Notebook	Task	Example
Massage Data Preprocessing for UCI Adult dataset		Massage Data Preprocessing Algorithm on the UCI Adult dataset to check and mitigate dataset/model bias