https://github.com/zjykzj/modelflow
Model Eval & Export & Infer
https://github.com/zjykzj/modelflow
clip cmake cplusplus cpp linux mnn nvidia onnx onnxruntime openclip openvino python pytorch tensorrt yolo11 yolo11-seg yolov5 yolov8 yolov8-seg
Last synced: 2 months ago
JSON representation
Model Eval & Export & Infer
- Host: GitHub
- URL: https://github.com/zjykzj/modelflow
- Owner: zjykzj
- License: apache-2.0
- Created: 2021-07-03T06:53:32.000Z (almost 5 years ago)
- Default Branch: master
- Last Pushed: 2026-04-12T07:18:52.000Z (3 months ago)
- Last Synced: 2026-04-12T09:18:19.242Z (3 months ago)
- Topics: clip, cmake, cplusplus, cpp, linux, mnn, nvidia, onnx, onnxruntime, openclip, openvino, python, pytorch, tensorrt, yolo11, yolo11-seg, yolov5, yolov8, yolov8-seg
- Language: Python
- Homepage:
- Size: 8.17 MB
- Stars: 1
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
Model Eval & Export & Infer
## Table of Contents
- [Table of Contents](#table-of-contents)
- [Background](#background)
- [Maintainers](#maintainers)
- [Thanks](#thanks)
- [Contributing](#contributing)
- [License](#license)
## Background
The purpose of this repository is to better deploy computer vision algorithms, especially the implementation of object
classification, object detection, and instance segmentation algorithms.
At first, I wanted to design a unified architecture that would adapt to different pre-processing and post-processing,
network module implementations, and inference engines through a modular paradigm, similar to commonly used popular
repositories. But this approach is difficult to push forward. Every time I want to add a new algorithm, I need to
repeatedly think about how to split the algorithm according to the current architecture and how to adapt the input and
output of each module. These issues have left me exhausted and even lost my passion for warehouse optimization for a
period of time.
I have thought for a long time and confirmed that I may have fallen into some kind of development dilemma. In order to
pursue design perfection, the complexity of engineering development has been infinitely increased. After realizing the
problem of Over-Engineering, I plan to start over. In the new development, I will focus as much as possible on the goals
of this repository: model transformation and model inference, and minimize the content of architecture design. Focus the
main energy on algorithm deployment.
Note: The previous implementation was in [v0.1.0](https://github.com/zjykzj/ModelFlow/tree/v0.1.0).
## YOLOv5/YOLOv8 Eval
| Model | Inference Backend | Pre/Post-processing | Implementation File | Implemented |
|------------|-------------------|---------------------|----------------------------------|-------------|
| YOLOv5 | ONNX Runtime | PyTorch | `yolov5_runtime_w_torch.py` | ✅ |
| YOLOv5 | ONNX Runtime | NumPy | `yolov5_runtime_w_numpy.py` | ✅ |
| YOLOv5 | TensorRT | NumPy | `yolov5_tensorrt_w_numpy.py` | ✅ |
| YOLOv5 | Triton Server | NumPy | `yolov5_triton_w_numpy.py` | ✅ |
| YOLOv8 | ONNX Runtime | PyTorch | `yolov8_runtime_w_torch.py` | ✅ |
| YOLOv8 | ONNX Runtime | NumPy | `yolov8_runtime_w_numpy.py` | ✅ |
| YOLOv8 | TensorRT | NumPy | `yolov8_tensorrt_w_numpy.py` | ✅ |
| YOLOv8 | Triton Server | NumPy | `yolov8_triton_w_numpy.py` | ✅ |
| YOLOv8-seg | ONNX Runtime | PyTorch | `yolov8_seg_runtime_w_torch.py` | ✅ |
| YOLOv8-seg | ONNX Runtime | NumPy | `yolov8_seg_runtime_w_numpy.py` | ✅ |
| YOLOv8-seg | TensorRT | NumPy | `yolov8_seg_tensorrt_w_numpy.py` | ✅ |
| YOLOv8-seg | Triton Server | NumPy | `yolov8_seg_triton_w_numpy.py` | ✅ |
## CLIP/OpenCLIP Eval
CLIP 和 OpenCLIP 模型在 CIFAR-10 和 CIFAR-100 上的 Zero-Shot 和 Linear Probe 评估结果
### CIFAR-10 数据集
| 方法 / 分类头 | 模型 | 准确率 (Accuracy) | 提升（对比 Zero-Shot 单模板） |
|------------------------------|----------|----------------|----------------------|
| **Zero-Shot（单模板）** | CLIP | 88.80% | — |
| **Zero-Shot（20模板集成）** | CLIP | 89.52% | +0.72% |
| **Linear Probe + LR** | CLIP | 94.32% | +5.52% |
| **Zero-Shot（单模板）** | OpenCLIP | 88.66% | — |
| **Zero-Shot（20模板集成）** | OpenCLIP | 88.88% | +0.22% |
| **Linear Probe + LR** | OpenCLIP | **94.85%** | **+6.19%** |
| **Linear Probe + KNN (k=1)** | OpenCLIP | 91.91% | +3.25% |
| **Linear Probe + KNN (k=5)** | OpenCLIP | 93.56% | +4.90% |
### CIFAR-100 数据集
| 方法 / 分类头 | 模型 | 准确率 (Accuracy) | 提升（对比 Zero-Shot 单模板） |
|------------------------------|----------|----------------|----------------------|
| **Zero-Shot（单模板）** | CLIP | 61.70% | — |
| **Zero-Shot（20模板集成）** | CLIP | 63.96% | +2.26% |
| **Linear Probe + LR** | CLIP | 75.62% | +13.92% |
| **Zero-Shot（单模板）** | OpenCLIP | 67.02% | — |
| **Zero-Shot（20模板集成）** | OpenCLIP | 67.90% | +0.88% |
| **Linear Probe + LR** | OpenCLIP | **78.70%** | **+11.68%** |
| **Linear Probe + KNN (k=1)** | OpenCLIP | 70.83% | +3.81% |
| **Linear Probe + KNN (k=5)** | OpenCLIP | 73.32% | +6.30% |
## Maintainers
* zhujian - *Initial work* - [zjykzj](https://github.com/zjykzj)
## Contributing
Anyone's participation is welcome! Open an [issue](https://github.com/zjykzj/ModelFlow/issues) or submit PRs.
## License
[Apache License 2.0](LICENSE) © 2021 zjykzj