In PaddleDetection, a complete tutorial and benchmarks for model compression based on [PaddleSlim](https://github.com/PaddlePaddle/PaddleSlim) are provided. Currently supported methods:
-[pruning](prune)
-[quantitative](quant)
-[distillation](distill)
-[The joint strategy](extensions)
It is recommended that you use a combination of pruning and distillation training, or use pruning and quantization for test model compression. The following takes YOLOv3 as an example to carry out cutting, distillation and quantization experiments.
## Experimental Environment
- Python 3.7+
- PaddlePaddle >= 2.1.0
- PaddleSlim >= 2.1.0
- CUDA 10.1+
- cuDNN >=7.6.5
**Version Dependency between PaddleDetection, Paddle and PaddleSlim Version**
| PaddleDetection Version | PaddlePaddle Version | PaddleSlim Version | Note |
-`--slim_config`: Specify the compression policy profile.
- If you want to use distillation, please refer to [Distillation Doc](distill/README.md) for specific distillation methods and more distillation of detection models.
- The above V100 prediction delay non-quantified model is tested by TensorRT FP32, and the quantified model is tested by TensorRT INT8, and both of them include NMS time.
- The SD855 predicts the delay for deployment using PaddleLite, using the ARM8 architecture and using 4 Threads (4 Threads) to reason the delay.
### Distillation
#### COCO Benchmark
| Model | Compression Strategy | Input Size | Box AP | Download | Model Configuration File | Compression Strategy Configuration File |
COCO数据集作为目标检测任务的训练目标难度更大，意味着teacher网络会预测出更多的背景bbox，如果直接用teacher的预测输出作为student学习的`soft label`会有严重的类别不均衡问题。解决这个问题需要引入新的方法，详细背景请参考论文:[Object detection at 200 Frames Per Second](https://arxiv.org/abs/1805.06361)。
FGD全称为[Focal and Global Knowledge Distillation for Detectors](https://arxiv.org/abs/2111.11837v1)，是目标检测任务的一种蒸馏方法，FGD蒸馏分为两个部分`Focal`和`Global`。`Focal`蒸馏分离图像的前景和背景，让学生模型分别关注教师模型的前景和背景部分特征的关键像素；`Global`蒸馏部分重建不同像素之间的关系并将其从教师转移到学生，以补偿`Focal`蒸馏中丢失的全局信息。试验结果表明，FGD蒸馏算法在基于anchor和anchor free的方法上能有效提升模型精度。
