Archai Documentation#

Archai is a Neural Architecture Search (NAS) framework built upon PyTorch. It provides a comprehensive solution for automating the process of finding the optimal architecture for deep learning models, making it easier for researchers and practitioners to achieve state-of-the-art results. First launched as an open-source project in 2020, Archai has made impactful progress by forming a positive feedback loop between the engineering and research aspects.

It has innovated on both search algorithms and search spaces, explored ideas on zero-cost proxies of architecture accuracy and in very recent work explored novel more efficient alternatives to the ubiquitious attention operator which is now informing next-generation search-space design. Additionally, it offers the following advantages:

  • 🔬 Easy mix-and-match between different algorithms;

  • 📈 Self-documented hyper-parameters and fair comparison;

  • ⚡ Extensible and modular to allow rapid experimentation;

  • 📂 Powerful configuration system and easy-to-use tools.

Citing Archai#

If you use Archai in a scientific publication, please consider citing it:

   title=Archai: Platform for Neural Architecture Search,
   journal=Microsoft Research,