Explore a groundbreaking approach to efficient machine learning deployment across diverse hardware platforms in this tinyML Talks webcast. Learn about the Once-for-All (OFA) network, which decouples training and search to support various architectural settings. Discover the novel progressive shrinking algorithm, a generalized pruning method that reduces model size across multiple dimensions. Understand how OFA outperforms state-of-the-art NAS methods on edge devices, achieving significant improvements in ImageNet top1 accuracy and latency compared to MobileNetV3 and EfficientNet. Gain insights into OFA's success in the 4th Low Power Computer Vision Challenge and its potential to revolutionize efficient inference on diverse hardware platforms while reducing GPU hours and CO2 emissions.
Train One Network and Specialize It for Efficient Deployment
Overview
Syllabus
tinyML. Talks Enabling ultra-low Power Machine Learning at the Edge "Once-for-All: Train One Network and Specialize it for Efficient Deployment"
Our 1st generation solution
New Challenge: Efficient Inference on Diverse Hardware Platfo
Our new solution, OFA: Decouple Training and Search
Challenge: Efficient Inference on Diverse Hardware Platforms
Once-for-All Network: Decouple Model Training and Architecture Design
Solution: Progressive Shrinking
Connection to Network Pruning
Performances of Sub-networks on ImageNe
How about search?
Accuracy & Latency Improvement
More accurate than training from scratch
OFA: 80% Top-1 Accuracy on ImageNet
Specialized Architecture for Different Hardware Platform
AutoML Outperforms Human Designing better MLM 1st place in CVPR 19 Visual Wake Words Challenge
What if we also optimize the compiler and run
How to save CO2 emission
OFA for FPGA Accelerators
Next tiny ML Talk
Taught by
tinyML
