Our Mission for embedded AI
In our mission to unleash AI on the edge, Datakalab has applied its proprietary quantization method on classical architecture and converted everythong to run in ONNX format for seamless compatibility with many hardware such as :
- Intel
- Nvidia
- NXP
- Texas Intrument
You can find an exhaustive list of devices support by ORT and the Execution Providers available
If you are looking for further compression methods (continual learning, pruning, batch norm folding, etc.) to fit your models on Intel devices, we’d love to hear from you!
Time on
Model name | <p align="center"> Relative Accuracy Drop [(fp32-int8)/fp32] | <p align="center"> Relative Model Size Improvement [(fp32-int8)/fp32] | <p align="center"> Memory Size Ratio [fp32/int8] |
---|---|---|---|
Mobilenet V2 | <p align="center"> 0.12% | <p align="center"> + 72.2% | <p align="center"> x 3.6 |
Resnet50 | <p align="center"> 0.13% | <p align="center"> + 74.5% | <p align="center"> x 3.9 |
Efficientnetlite b0 | <p align="center"> 0.1% | <p align="center"> + 72.9% | <p align="center"> x 3.7 |
YoloV3 | <p align="center"> 0.93% <sup>1</sup> | <p align="center"> + 77.3% | <p align="center"> x 3.8 |
<sup>1</sup> Accuracy Ratio on mAP results
Thanks
All those benchmarks where made using ONNX® and OnnxRuntime® Thanks a lot @Onnx®
Contact us
Feel free to reach out by email at hello at datakalab.com