Furiosa Models
furiosa-models is an open model zoo project for FuriosaAI NPU.
It provides a set of public pre-trained, pre-quantized models for learning and demo purposes or
for developing your applications.
furiosa-models also includes pre-packaged post/processing utilities, and compiler configurations optimized
for FuriosaAI NPU. However, all models are standard ONNX or tflite models,
and they can run even on CPU and GPU as well.
Releases
Online Documentation
If you are new, you can start from Getting Started. You can also find the latest online documents, including programming guides, API references, and examples from the following:
- Furiosa Models - Latest Documentation
- Model Object
- Model List
- Command Line Tool
- Furiosa SDK - Tutorial and Code Examples
Model List
The table summarizes all models available in furiosa-models. If you visit each model link,
you can find details about loading a model, their input and output tensors, pre/post-processing, and usage examples.
| Model | Task | Size | Accuracy |
|---|---|---|---|
| ResNet50 | Image Classification | 25M | 75.618% (ImageNet1K-val) |
| EfficientNetB0 | Image Classification | 6.4M | 72.44% (ImageNet1K-val) |
| EfficientNetV2-S | Image Classification | 26M | 83.532% (ImageNet1K-val) |
| SSDMobileNet | Object Detection | 7.2M | mAP 0.232 (COCO 2017-val) |
| SSDResNet34 | Object Detection | 20M | mAP 0.220 (COCO 2017-val) |
| YOLOv5M | Object Detection | 21M | mAP 0.272 (Bdd100k-val) [1] |
| YOLOv5L | Object Detection | 46M | mAP 0.284 (Bdd100k-val) [1] |
| YOLOv7w6Pose | Pose Estimation | 80M | N/A |
[1]: The accuracy of the yolov5 f32 model trained with bdd100k-val dataset, is mAP 0.295 (for yolov5m) and mAP 0.316 (for yolov5l).