fastvit

Porting vision models to Keras 3 for easily accessibility. Contains MobileViT v1, MobileViT v2, fastvit

FastViT base model for use with Autodistill.

Reproducible deep-learning image-classification pipeline for industrial soft sensing. Originally built for sludge-cake quality monitoring at DC Water Blue Plains AWWTP. Six architectures (FastViT, EfficientNet, MobileNet, EfficientFormerV2, DeepTEN-ResNet, sparse-AE CNN) compared with multi-seed statistics on Modal cloud GPUs.

Distills BioCLIP 2.5 Huge (ViT-H/14, 632M params) into an 11.6M param FastViT student for on-device plant ID. 71.7% top-1 teacher agreement, 23.8 MB fp16 ONNX.

Real-time court tracking at 143+ FPS: Detects true boundary edge corners using FastViT and MobileNet, with C++ refinement and homography recovery.