microsoft/Cream
Collection of neural architecture search methods and efficient Vision Transformer models for visual recognition, including distillation and compression techniques.
Velocity · 7d
+0.9
★ / day
Trend
→steady
star history
This repository aggregates Microsoft’s research on automated neural architecture design and search. It includes multiple NAS algorithms for discovering efficient transformer architectures and techniques for compressing vision models via weight multiplexing, knowledge distillation, and affinity mimicking. The work spans methods like AutoFormer, CDARTS, TinyViT, and MiniViT, providing implementations for training, searching, and compressing vision transformers.