H-BAC Shrinks a 327 MB Plant Disease Model Down to 6 MB
A unified pruning, quantization and distillation pipeline shrinks a Vision Transformer 54.5x while holding 95.13% accuracy on out-of-distribution chilli disease images.
- ViT-B/16 compressed 54.5x from 327.42 MB to 6.01 MB while holding 95.13% accuracy on OOD chilli disease data.
- Pipeline combines Hessian-Balanced Adaptive Block Pruning, attention-based knowledge distillation, and dynamic INT8 quantization.
- Evaluated on a cross-village, cross-device split, not a leaky in-distribution test set.
- Compressed models achieved 74 to 98 percent size reduction across four configurations.
- A directly-trained 6.01 MB INT8 student hit 94.87%, nearly matching the full pipeline.
- Authors openly note H-BAC and distillation are not yet shown worth their compute cost here.
A 6 MB chilli disease model tests how much ViT compression buys
A team led by Mahadev Sunil Kumar compressed a ViT-B/16 plant disease classifier from 327.42 MB to 6.01 MB while preserving its reported 95.13% accuracy on images from held-out villages and devices. A smaller model trained from scratch and quantized to the same size reached 94.87%, giving developers a strong baseline to test before adopting the full pruning and distillation pipeline.
Why fields expose model bloat
Indian chilli growers often need to distinguish healthy plants from early and severe disease without reliable access to crop specialists or cloud services. Running inference on an Android phone avoids network latency, data charges, and connectivity failures, but it places strict limits on storage, memory, and compute.
ViT-B/16 divides an image into 16-by-16-pixel patches and processes them through a stack of transformer blocks. The architecture can classify plant disease accurately, but its 327.42 MB FP32 checkpoint adds substantial download and storage costs. Its parameter count can also create runtime memory and latency problems, although this study measures model size rather than on-device performance.
H-BAC chooses what to cut
The H-BAC paper combines structural pruning, attention-based knowledge distillation, and dynamic INT8 quantization. H-BAC stands for Hessian-Balanced Adaptive Block Pruning.
- Structural pruning
- Removes model capacity according to block-level sensitivity scores, producing a physically smaller network.
- Hessian-based sensitivity
- Approximates the curvature of the loss around each transformer block. In plain terms, it estimates how sharply accuracy may deteriorate when a block changes. Lower-sensitivity blocks can absorb more pruning.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
