[English](../en/quantize.md) | 简体中文 # 量化加速 量化是一种流行的模型压缩方法，量化后的模型拥有更小的体积和更快的推理速度. FastDeploy基于PaddleSlim的Auto Compression Toolkit(ACT), 给用户提供了一键模型自动化压缩的工具. FastDeploy一键模型自动压缩可包含多种策略, 目前主要采用离线量化和量化蒸馏训练.同时, FastDeploy支持部署压缩后的模型, 帮助用户实现推理加速. 本文主要描述量化模型在FastDeploy上的部署. ## FastDeploy 多个引擎和硬件支持量化模型部署 当前，FastDeploy中多个推理后端可以在不同硬件上支持量化模型的部署. 支持情况如下: | 硬件/推理后端 | ONNX Runtime | Paddle Inference | TensorRT | Paddle-TensorRT | | :-----------| :-------- | :--------------- | :------- | :------- | | CPU | 支持 | 支持 | | | | GPU | | | 支持 | 支持 | ## 模型量化 ### 量化方法 基于PaddleSlim，目前FastDeploy一键模型自动压缩提供的的量化方法有量化蒸馏训练和离线量化，量化蒸馏训练通过模型训练来获得量化模型，离线量化不需要模型训练即可完成模型的量化。 FastDeploy 对两种方式产出的量化模型均能部署。 两种方法的主要对比如下表所示: | 量化方法 | 量化过程耗时 | 量化模型精度 | 模型体积 | 推理速度 | | :-----------| :--------| :-------| :------- | :------- | | 离线量化 | 无需训练，耗时短 | 比量化蒸馏训练稍低 | 两者一致 | 两者一致 | | 量化蒸馏训练 | 需要训练，耗时稍高 | 较未量化模型有少量损失 | 两者一致 |两者一致 | ### 使用FastDeploy一键模型自动化压缩工具来量化模型 FastDeploy基于PaddleSlim的Auto Compression Toolkit(ACT), 给用户提供了一键模型自动化压缩的工具，请参考如下文档进行一键模型自动化压缩。 - [FastDeploy 一键模型自动化压缩](../../tools/common_tools/auto_compression/) 当用户获得产出的压缩模型之后，即可以使用FastDeploy来部署压缩模型。 ## 量化模型 Benchmark 目前, FastDeploy支持自动化压缩,并完成部署测试的模型的Runtime Benchmark和端到端Benchmark如下所示. Benchmark表格说明: - Runtime时延为模型在各种Runtime上的推理时延,包含CPU->GPU数据拷贝,GPU推理,GPU->CPU数据拷贝时间. 不包含模型各自的前后处理时间. - 端到端时延为模型在实际推理场景中的时延, 包含模型的前后处理. - 所测时延均为推理1000次后求得的平均值, 单位是毫秒. - INT8 + FP16 为在推理INT8量化模型的同时, 给Runtime 开启FP16推理选项 - INT8 + FP16 + PM, 为在推理INT8量化模型和开启FP16的同时, 开启使用Pinned Memory的选项,可加速GPU->CPU数据拷贝的速度 - 最大加速比, 为FP32时延除以INT8推理的最快时延,得到最大加速比. - 策略为量化蒸馏训练时, 采用少量无标签数据集训练得到量化模型, 并在全量验证集上验证精度, INT8精度并不代表最高的INT8精度. - CPU为Intel(R) Xeon(R) Gold 6271C, 所有测试中固定CPU线程数为1. GPU为Tesla T4, TensorRT版本8.4.15. ### YOLO 系列 #### Runtime Benchmark | 模型 |推理后端 |部署硬件 | FP32 Runtime时延 | INT8 Runtime时延 | INT8 + FP16 Runtime时延 | INT8+FP16+PM Runtime时延 | 最大加速比 | FP32 mAP | INT8 mAP | 量化方式 | | ------------------- | -----------------|-----------| -------- |-------- |-------- | --------- |-------- |----- |----- |----- | | [YOLOv5s](../../examples/vision/detection/yolov5/quantize/) | TensorRT | GPU | 7.87 | 4.51 | 4.31 | 3.17 | 2.48 | 37.6 | 36.7 | 量化蒸馏训练 | | [YOLOv5s](../../examples/vision/detection/yolov5/quantize/) | Paddle-TensorRT | GPU | 7.99 | None | 4.46 | 3.31 | 2.41 | 37.6 | 36.8 | 量化蒸馏训练 | | [YOLOv5s](../../examples/vision/detection/yolov5/quantize/) | ONNX Runtime | CPU | 176.41 | 91.90 | None | None | 1.90 | 37.6 | 33.1 |量化蒸馏训练 | | [YOLOv5s](../../examples/vision/detection/yolov5/quantize/) | Paddle Inference| CPU | 213.73 | 130.19 | None | None | 1.64 |37.6 | 35.2 | 量化蒸馏训练 | | [YOLOv6s](../../examples/vision/detection/yolov6/quantize/) | TensorRT | GPU | 9.47 | 3.23 | 4.09 |2.81 | 3.37 | 42.5 | 40.7|量化蒸馏训练 | | [YOLOv6s](../../examples/vision/detection/yolov6/quantize/) | Paddle-TensorRT | GPU | 9.31 | None| 4.17 | 2.95 | 3.16 | 42.5 | 40.7|量化蒸馏训练 | | [YOLOv6s](../../examples/vision/detection/yolov6/quantize/) | ONNX Runtime | CPU | 334.65 | 126.38 | None | None| 2.65 |42.5| 36.8|量化蒸馏训练 | | [YOLOv6s](../../examples/vision/detection/yolov6/quantize/) | Paddle Inference | CPU | 352.87 | 123.12 |None | None| 2.87 |42.5| 40.8|量化蒸馏训练 | | [YOLOv7](../../examples/vision/detection/yolov7/quantize/) | TensorRT | GPU | 27.47 | 6.52 | 6.74| 5.19| 5.29 | 51.1| 50.4|量化蒸馏训练 | | [YOLOv7](../../examples/vision/detection/yolov7/quantize/) | Paddle-TensorRT | GPU | 27.87|None|6.91|5.86 | 4.76 | 51.1| 50.4|量化蒸馏训练 | | [YOLOv7](../../examples/vision/detection/yolov7/quantize/) | ONNX Runtime | CPU | 996.65 | 467.15 |None|None | 2.13 | 51.1 | 43.3|量化蒸馏训练 | | [YOLOv7](../../examples/vision/detection/yolov7/quantize/) | Paddle Inference | CPU | 995.85 | 477.93|None|None | 2.08 |51.1 | 46.2|量化蒸馏训练 | #### 端到端 Benchmark | 模型 |推理后端 |部署硬件 | FP32 End2End时延 | INT8 End2End时延 | INT8 + FP16 End2End时延 | INT8+FP16+PM End2End时延 | 最大加速比 | FP32 mAP | INT8 mAP | 量化方式 | | ------------------- | -----------------|-----------| -------- |-------- |-------- | --------- |-------- |----- |----- |----- | | [YOLOv5s](../../examples/vision/detection/yolov5/quantize/) | TensorRT | GPU | 24.61 | 21.20 | 20.78 | 20.94 | 1.18 | 37.6 | 36.7 | 量化蒸馏训练 | | [YOLOv5s](../../examples/vision/detection/yolov5/quantize/) | Paddle-TensorRT | GPU | 23.53 | None | 21.98 | 19.84 | 1.28 | 37.6 | 36.8 | 量化蒸馏训练 | | [YOLOv5s](../../examples/vision/detection/yolov5/quantize/) | ONNX Runtime | CPU | 197.323 | 110.99 | None | None | 1.78 | 37.6 | 33.1 |量化蒸馏训练 | | [YOLOv5s](../../examples/vision/detection/yolov5/quantize/) | Paddle Inference| CPU | 235.73 | 144.82 | None | None | 1.63 |37.6 | 35.2 | 量化蒸馏训练 | | [YOLOv6s](../../examples/vision/detection/yolov6/quantize/) | TensorRT | GPU | 15.66 | 11.30 | 10.25 |9.59 | 1.63 | 42.5 | 40.7|量化蒸馏训练 | | [YOLOv6s](../../examples/vision/detection/yolov6/quantize/) | Paddle-TensorRT | GPU | 15.03 | None| 11.36 | 9.32 | 1.61 | 42.5 | 40.7|量化蒸馏训练 | | [YOLOv6s](../../examples/vision/detection/yolov6/quantize/) | ONNX Runtime | CPU | 348.21 | 126.38 | None | None| 2.82 |42.5| 36.8|量化蒸馏训练 | | [YOLOv6s](../../examples/vision/detection/yolov6/quantize/) | Paddle Inference | CPU | 352.87 | 121.64 |None | None| 3.04 |42.5| 40.8|量化蒸馏训练 | | [YOLOv7](../../examples/vision/detection/yolov7/quantize/) | TensorRT | GPU | 36.47 | 18.81 | 20.33| 17.58| 2.07 | 51.1| 50.4|量化蒸馏训练 | | [YOLOv7](../../examples/vision/detection/yolov7/quantize/) | Paddle-TensorRT | GPU | 37.06|None|20.26|17.53 | 2.11 | 51.1| 50.4|量化蒸馏训练 | | [YOLOv7](../../examples/vision/detection/yolov7/quantize/) | ONNX Runtime | CPU | 988.85 | 478.08 |None|None | 2.07 | 51.1 | 43.3|量化蒸馏训练 | | [YOLOv7](../../examples/vision/detection/yolov7/quantize/) | Paddle Inference | CPU | 1031.73 | 500.12|None|None | 2.06 |51.1 | 46.2|量化蒸馏训练 | ### PaddleClas系列 #### Runtime Benchmark | 模型 |推理后端 |部署硬件 | FP32 Runtime时延 | INT8 Runtime时延 | INT8 + FP16 Runtime时延 | INT8+FP16+PM Runtime时延 | 最大加速比 | FP32 Top1 | INT8 Top1 | 量化方式 | | ------------------- | -----------------|-----------| -------- |-------- |-------- | --------- |-------- |----- |----- |----- | | [ResNet50_vd](../../examples/vision/classification/paddleclas/quantize/) | TensorRT | GPU | 3.55 | 0.99|0.98|1.06 | 3.62 | 79.12 | 79.06 | 离线量化 | | [ResNet50_vd](../../examples/vision/classification/paddleclas/quantize/) | Paddle-TensorRT | GPU | 3.46 |None |0.87|1.03 | 3.98 | 79.12 | 79.06 | 离线量化 | | [ResNet50_vd](../../examples/vision/classification/paddleclas/quantize/) | ONNX Runtime | CPU | 76.14 | 35.43 |None|None | 2.15 | 79.12 | 78.87| 离线量化| | [ResNet50_vd](../../examples/vision/classification/paddleclas/quantize/) | Paddle Inference | CPU | 76.21 | 24.01 |None|None | 3.17 | 79.12 | 78.55 | 离线量化| | [MobileNetV1_ssld](../../examples/vision/classification/paddleclas/quantize/) | TensorRT | GPU | 0.91 | 0.43 |0.49 | 0.54 | 2.12 |77.89 | 76.86 | 离线量化 | | [MobileNetV1_ssld](../../examples/vision/classification/paddleclas/quantize/) | Paddle-TensorRT | GPU | 0.88| None| 0.49|0.51 | 1.80 |77.89 | 76.86 | 离线量化 | | [MobileNetV1_ssld](../../examples/vision/classification/paddleclas/quantize/) | ONNX Runtime | CPU | 30.53 | 9.59|None|None | 3.18 |77.89 | 75.09 |离线量化 | | [MobileNetV1_ssld](../../examples/vision/classification/paddleclas/quantize/) | Paddle Inference | CPU | 12.29 | 4.68 | None|None|2.62 |77.89 | 71.36 |离线量化 | #### 端到端 Benchmark | 模型 |推理后端 |部署硬件 | FP32 End2End时延 | INT8 End2End时延 | INT8 + FP16 End2End时延 | INT8+FP16+PM End2End时延 | 最大加速比 | FP32 Top1 | INT8 Top1 | 量化方式 | | ------------------- | -----------------|-----------| -------- |-------- |-------- | --------- |-------- |----- |----- |----- | | [ResNet50_vd](../../examples/vision/classification/paddleclas/quantize/) | TensorRT | GPU | 4.92| 2.28|2.24|2.23 | 2.21 | 79.12 | 79.06 | 离线量化 | | [ResNet50_vd](../../examples/vision/classification/paddleclas/quantize/) | Paddle-TensorRT | GPU | 4.48|None |2.09|2.10 | 2.14 | 79.12 | 79.06 | 离线量化 | | [ResNet50_vd](../../examples/vision/classification/paddleclas/quantize/) | ONNX Runtime | CPU | 77.43 | 41.90 |None|None | 1.85 | 79.12 | 78.87| 离线量化| | [ResNet50_vd](../../examples/vision/classification/paddleclas/quantize/) | Paddle Inference | CPU | 80.60 | 27.75 |None|None | 2.90 | 79.12 | 78.55 | 离线量化| | [MobileNetV1_ssld](../../examples/vision/classification/paddleclas/quantize/) | TensorRT | GPU | 2.19 | 1.48|1.57| 1.57 | 1.48 |77.89 | 76.86 | 离线量化 | | [MobileNetV1_ssld](../../examples/vision/classification/paddleclas/quantize/) | Paddle-TensorRT | GPU | 2.04| None| 1.47|1.45 | 1.41 |77.89 | 76.86 | 离线量化 | | [MobileNetV1_ssld](../../examples/vision/classification/paddleclas/quantize/) | ONNX Runtime | CPU | 34.02 | 12.97|None|None | 2.62 |77.89 | 75.09 |离线量化 | | [MobileNetV1_ssld](../../examples/vision/classification/paddleclas/quantize/) | Paddle Inference | CPU | 16.31 | 7.42 | None|None| 2.20 |77.89 | 71.36 |离线量化 | ### PaddleDetection系列 #### Runtime Benchmark | 模型 |推理后端 |部署硬件 | FP32 Runtime时延 | INT8 Runtime时延 | INT8 + FP16 Runtime时延 | INT8+FP16+PM Runtime时延 | 最大加速比 | FP32 mAP | INT8 mAP | 量化方式 | | ------------------- | -----------------|-----------| -------- |-------- |-------- | --------- |-------- |----- |----- |----- | | [ppyoloe_crn_l_300e_coco](../../examples/vision/detection/paddledetection/quantize ) | TensorRT | GPU | 27.90 | 6.39 |6.44|5.95 | 4.67 | 51.4 | 50.7 | 量化蒸馏训练 | | [ppyoloe_crn_l_300e_coco](../../examples/vision/detection/paddledetection/quantize ) | Paddle-TensorRT | GPU | 30.89 |None | 13.78 |14.01 | 2.24 | 51.4 | 50.5| 量化蒸馏训练 | | [ppyoloe_crn_l_300e_coco](../../examples/vision/detection/paddledetection/quantize) | ONNX Runtime | CPU | 1057.82 | 449.52 |None|None | 2.35 |51.4 | 50.0 |量化蒸馏训练 | NOTE: - TensorRT比Paddle-TensorRT快的原因是在runtime移除了multiclass_nms3算子 #### 端到端 Benchmark | 模型 |推理后端 |部署硬件 | FP32 End2End时延 | INT8 End2End时延 | INT8 + FP16 End2End时延 | INT8+FP16+PM End2End时延 | 最大加速比 | FP32 mAP | INT8 mAP | 量化方式 | | ------------------- | -----------------|-----------| -------- |-------- |-------- | --------- |-------- |----- |----- |----- | | [ppyoloe_crn_l_300e_coco](../../examples/vision/detection/paddledetection/quantize ) | TensorRT | GPU | 35.75 | 15.42 |20.70|20.85 | 2.32 | 51.4 | 50.7 | 量化蒸馏训练 | | [ppyoloe_crn_l_300e_coco](../../examples/vision/detection/paddledetection/quantize ) | Paddle-TensorRT | GPU | 33.48 |None | 18.47 |18.03 | 1.81 | 51.4 | 50.5| 量化蒸馏训练 | | [ppyoloe_crn_l_300e_coco](../../examples/vision/detection/paddledetection/quantize) | ONNX Runtime | CPU | 1067.17 | 461.037 |None|None | 2.31 |51.4 | 50.0 |量化蒸馏训练 | ### PaddleSeg系列 #### Runtime Benchmark | 模型 |推理后端 |部署硬件 | FP32 Runtime时延 | INT8 Runtime时延 | INT8 + FP16 Runtime时延 | INT8+FP16+PM Runtime时延 | 最大加速比 | FP32 mIoU | INT8 mIoU | 量化方式 | | ------------------- | -----------------|-----------| -------- |-------- |-------- | --------- |-------- |----- |----- |----- | | [PP-LiteSeg-T(STDC1)-cityscapes](../../examples/vision/segmentation/paddleseg/quantize) | Paddle Inference | CPU | 1138.04| 602.62 |None|None | 1.89 |77.37 | 71.62 |量化蒸馏训练 | #### 端到端 Benchmark | 模型 |推理后端 |部署硬件 | FP32 End2End时延 | INT8 End2End时延 | INT8 + FP16 End2End时延 | INT8+FP16+PM End2End时延 | 最大加速比 | FP32 mIoU | INT8 mIoU | 量化方式 | | ------------------- | -----------------|-----------| -------- |-------- |-------- | --------- |-------- |----- |----- |----- | | [PP-LiteSeg-T(STDC1)-cityscapes](../../examples/vision/segmentation/paddleseg/quantize) | Paddle Inference | CPU | 4726.65| 4134.91|None|None | 1.14 |77.37 | 71.62 |量化蒸馏训练 |