This repository contains the model artifacts used for my master thesis Benchmarking Computer Vision Tasks on Edge AI Hardware.
Benchmarking framework: https://github.com/hig-dev/edgebench
- https://github.com/hig-dev/mmpose/blob/main/quick_export.py
- https://github.com/hig-dev/mmpose/blob/main/quick_export_edgetpu.py
- https://github.com/hig-dev/mmpose/blob/main/quick_export_hailo.py
- https://github.com/hig-dev/mmpose/blob/main/quick_export_hhb.py
- https://github.com/hig-dev/mmpose/blob/main/quick_export_tidl.py
- https://github.com/hig-dev/mmpose/blob/main/quick_export_vela.py
| Directory | Description |
|---|---|
executorch(xnnpack)/ |
ExecuTorch model files optimized with XNNPACK backend for CPU inference |
hef/ |
Hailo Executable Format files for Raspberry Pi AI HAT+ |
hhb/ |
Heterogeneous Honey Badger (HHB) compiled models for BeagleV-Ahead |
onnx/ |
ONNX models with opset 20 |
pytorch/ |
PyTorch model files (.pth format) |
tflite(float32)/ |
TensorFlow Lite models with float32 precision |
tflite(int8-edgetpu)/ |
TensorFlow Lite models quantized to int8 for Coral Edge TPU |
tflite(int8-ethosu55)/ |
TensorFlow Lite models quantized to int8 for Arm Ethos-U55 NPU |
tflite(int8)/ |
TensorFlow Lite models with int8 quantization |
tidl(am67a)/ |
Texas Instruments Deep Learning (TIDL) models optimized for AM67A processors |