Contributor: Athanasios Masouris (ThanosM97)
Organization: OpenVINO Toolkit
Project: Train a DL model for synthetic data generation for model optimization (project page)
GitHub repository: ThanosM97/gsoc2022-openvino
The project is divided into two parts. The goal for the first part is to train a lightweight Deep Learning model to generate a dataset of synthetic images. I propose a class-conditional Generative Adversarial Network to generate images for the 10 categories of the CIFAR-10 dataset, given the class label as input. The model is trained using a knowledge distillation framework, in an attempt to compress the StyleGAN2-ADA network. For the second part, the pre-trained model of the first part is used to generate a dataset of synthetic images for CIFAR-10. Subsequently, this dataset is used for model optimization using OpenVINO's Post-training Optimization Tool. We evaluate the performance of the 8-bit post-training quantization method on a range of Computer Vision models.
A more detailed explanation of the project can be found in the project's README file, or in the wikipages.
An overview of the work I conducted during the Google Summer of Code 2022 can be found in this timeline, or by looking at my commits. Additionally, there are two blogs (Blog #1, Blog #2) explaining in detail the work conducted for the project.
- A synthetic dataset of 50,000 samples generated using the StyleGAN2-ADA model
- A pre-trained conditional GAN, DiStyleGAN, to generate images from the CIFAR-10 distribution ( weights | information )
- A webapp for conditional image generation using the pre-trained model
- Three calibration datasets for model optimization
- A range of computer vision quantized models for image classification on CIFAR-10
- Two blogs describing the development of a class-conditional GAN for synthetic image generation and the quantization of deep learning models using OpenVINO Toolkit
- GSoC22 Contributor: Athanasios Masouris (ThanosM97)
- Mentor: Mansi Sharma (mansishr)
- Mentor: Zhuo Wu (zhuo-yoyowz)