The masterclass was led by Cédric Deltheil, co-founder of Finegrain, a 100% automated generative AI solution that allows retailers and e-commerce sites to generate visual filters to enhance their products, while tracking the performance and engagement of these various filters.
The goal was to immerse our students and PhD candidates in the future of foundation models, by exploring Refiners, an open-source micro-framework (MIT) developed by the startup finegrain and built on PyTorch to facilitate the adaptation of foundation models.
What was the observation behind Refiners? Adaptation techniques – effective alternatives to full model "fine-tuning" – are booming, such as LoRA (Low-rank adaptation), ControlNet, IP-Adapter, etc. This presents software engineering challenges in supporting and combining the variety of available "adapters".
To support them natively, Refiners is built around 3 key concepts:
1/ "Chain": a declarative approach for representing models
2/ "Context": a solution for storing properties accessible by all deep nodes of the "Chain"
3/ "Adapter": an API for dynamically modifying sub-parts of a model
For Finegrain, Refiners offers three major business advantages:
1. To showcase and validate their excellent technical expertise
2. Evolve a technological foundation that will drive the value proposition
3. Attract and recruit the best talent.
In fact, a community has emerged on Discord, and a "bounties" program has been set up as a call for contributions. Feel free to star us on GitHub! https://github.com/finegrain-ai/refiners
.png)
