GT4SD/gt4sd-core

What it solves

GT4SD (Generative Toolkit for Scientific Discovery) simplifies the process of hypothesis generation in scientific discovery. It addresses the fragmentation of generative AI tools by providing a unified, high-level API and a standardized interface to access and use various state-of-the-art generative models for materials science and NLP, reducing the effort required to implement and run these models.

How it works

GT4SD acts as a wrapper and orchestrator for multiple existing generative AI packages (such as GuacaMol, Moses, TorchDrug, and HuggingFace Transformers). It provides three primary interaction methods:

1. Python API: Allows users to instantiate algorithms via a registry and run sampling/inference directly in code.
2. CLI Tools: Provides dedicated commands (gt4sd-inference, gt4sd-trainer, gt4sd-saving, and gt4sd-upload) to run inference pipelines, train models, save trained algorithms, and upload them to a public hub.
3. Property Prediction: Includes a properties submodule to compute characteristics (like molecular weight or similarity) of the generated samples.

Who it’s for

Scientists and researchers in materials science, chemistry, and NLP who want to use generative AI to design new molecules, polymers, or peptides without having to write complex boilerplate code for every different model architecture.

Highlights

• Unified Interface: Harmonizes APIs across diverse libraries like Diffusers, GFlowNets, and Regression Transformers.
• Comprehensive Pipeline: Supports the full lifecycle from training and fine-tuning to inference and deployment.
• ** uma Hub Integration**: Facilitates uploading trained models to a public hub for sharing.
• Broad Model Support: Integrates a wide array of generative techniques, including VAEs, LSTMs, Diffusion models, and Graph Convolutional Policy Networks.