移动paddle_detection
This commit is contained in:
183
services/paddle_services/paddle_detection/configs/slim/README.md
Normal file
183
services/paddle_services/paddle_detection/configs/slim/README.md
Normal file
@@ -0,0 +1,183 @@
|
||||
# 模型压缩
|
||||
|
||||
在PaddleDetection中, 提供了基于[PaddleSlim](https://github.com/PaddlePaddle/PaddleSlim)进行模型压缩的完整教程和benchmark。目前支持的方法：
|
||||
|
||||
- [剪裁](prune)
|
||||
- [量化](quant)
|
||||
- [离线量化](post_quant)
|
||||
- [蒸馏](distill)
|
||||
- [联合策略](extensions)
|
||||
|
||||
推荐您使用剪裁和蒸馏联合训练，或者使用剪裁、量化训练和离线量化，进行检测模型压缩。 下面以YOLOv3为例，进行剪裁、蒸馏和量化实验。
|
||||
|
||||
## 实验环境
|
||||
|
||||
- Python 3.7+
|
||||
- PaddlePaddle >= 2.1.0
|
||||
- PaddleSlim >= 2.1.0
|
||||
- CUDA 10.1+
|
||||
- cuDNN >=7.6.5
|
||||
|
||||
**PaddleDetection、 PaddlePaddle与PaddleSlim 版本关系:**
|
||||
| PaddleDetection版本 | PaddlePaddle版本 | PaddleSlim版本 | 备注 |
|
||||
| :------------------: | :---------------: | :-------: |:---------------: |
|
||||
| release/2.3 | >= 2.1 | 2.1 | 离线量化依赖Paddle 2.2及PaddleSlim 2.2 |
|
||||
| release/2.1 | 2.2 | >= 2.1.0 | 2.1 | 量化模型导出依赖最新Paddle develop分支，可在[PaddlePaddle每日版本](https://www.paddlepaddle.org.cn/documentation/docs/zh/install/Tables.html#whl-dev)中下载安装 |
|
||||
| release/2.0 | >= 2.0.1 | 2.0 | 量化依赖Paddle 2.1及PaddleSlim 2.1 |
|
||||
|
||||
|
||||
#### 安装PaddleSlim
|
||||
- 方法一：直接安装：
|
||||
```
|
||||
pip install paddleslim -i https://pypi.tuna.tsinghua.edu.cn/simple
|
||||
```
|
||||
- 方法二：编译安装：
|
||||
```
|
||||
git clone https://github.com/PaddlePaddle/PaddleSlim.git
|
||||
cd PaddleSlim
|
||||
python setup.py install
|
||||
```
|
||||
|
||||
## 快速开始
|
||||
|
||||
### 训练
|
||||
|
||||
```shell
|
||||
python tools/train.py -c configs/{MODEL.yml} --slim_config configs/slim/{SLIM_CONFIG.yml}
|
||||
```
|
||||
|
||||
- `-c`: 指定模型配置文件。
|
||||
- `--slim_config`: 指定压缩策略配置文件。
|
||||
- 如果选择使用蒸馏，具体蒸馏方法和更多检测模型的蒸馏，请参考[蒸馏策略文档](distill/README.md)
|
||||
|
||||
|
||||
### 评估
|
||||
|
||||
```shell
|
||||
python tools/eval.py -c configs/{MODEL.yml} --slim_config configs/slim/{SLIM_CONFIG.yml} -o weights=output/{SLIM_CONFIG}/model_final
|
||||
```
|
||||
|
||||
- `-c`: 指定模型配置文件。
|
||||
- `--slim_config`: 指定压缩策略配置文件。
|
||||
- `-o weights`: 指定压缩算法训好的模型路径。
|
||||
|
||||
### 测试
|
||||
|
||||
```shell
|
||||
python tools/infer.py -c configs/{MODEL.yml} --slim_config configs/slim/{SLIM_CONFIG.yml} \
|
||||
-o weights=output/{SLIM_CONFIG}/model_final
|
||||
--infer_img={IMAGE_PATH}
|
||||
```
|
||||
|
||||
- `-c`: 指定模型配置文件。
|
||||
- `--slim_config`: 指定压缩策略配置文件。
|
||||
- `-o weights`: 指定压缩算法训好的模型路径。
|
||||
- `--infer_img`: 指定测试图像路径。
|
||||
|
||||
|
||||
## 全链条部署
|
||||
|
||||
### 动转静导出模型
|
||||
|
||||
```shell
|
||||
python tools/export_model.py -c configs/{MODEL.yml} --slim_config configs/slim/{SLIM_CONFIG.yml} -o weights=output/{SLIM_CONFIG}/model_final
|
||||
```
|
||||
|
||||
- `-c`: 指定模型配置文件。
|
||||
- `--slim_config`: 指定压缩策略配置文件。
|
||||
- `-o weights`: 指定压缩算法训好的模型路径。
|
||||
|
||||
### 部署预测
|
||||
|
||||
- Paddle-Inference预测：
|
||||
- [Python部署](../../deploy/python/README.md)
|
||||
- [C++部署](../../deploy/cpp/README.md)
|
||||
- [TensorRT预测部署教程](../../deploy/TENSOR_RT.md)
|
||||
- 服务器端部署：使用[PaddleServing](../../deploy/serving/README.md)部署。
|
||||
- 手机移动端部署：使用[Paddle-Lite](../../deploy/lite/README.md) 在手机移动端部署。
|
||||
|
||||
## Benchmark
|
||||
|
||||
### 剪裁
|
||||
|
||||
#### Pascal VOC上benchmark
|
||||
|
||||
| 模型 | 压缩策略 | GFLOPs | 模型体积(MB) | 输入尺寸 | 预测时延(SD855) | Box AP | 下载 | 模型配置文件 | 压缩算法配置文件 |
|
||||
| :---------: | :-------: | :------------: |:-------------: | :------: | :-------------: | :------: | :-----------------------------------------------------: |:-------------: | :------: |
|
||||
| YOLOv3-MobileNetV1 | baseline | 24.13 | 93 | 608 | 332.0ms | 75.1 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v1_270e_voc.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v1_270e_voc.yml) | - |
|
||||
| YOLOv3-MobileNetV1 | 剪裁-l1_norm(sensity) | 15.78(-34.49%) | 66(-29%) | 608 | - | 78.4(+3.3) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_voc_prune_l1_norm.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v1_270e_voc.yml) | [slim配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/prune/yolov3_prune_l1_norm.yml) |
|
||||
|
||||
#### COCO上benchmark
|
||||
| 模型 | 压缩策略 | GFLOPs | 模型体积(MB) | 输入尺寸 | 预测时延(SD855) | Box AP | 下载 | 模型配置文件 | 压缩算法配置文件 |
|
||||
| :---------: | :-------: | :------------: |:-------------: | :------: | :-------------: | :------: | :-----------------------------------------------------: |:-------------: | :------: |
|
||||
| PP-YOLO-MobileNetV3_large | baseline | -- | 18.5 | 608 | 25.1ms | 23.2 | [下载链接](https://paddledet.bj.bcebos.com/models/ppyolo_mbv3_large_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_mbv3_large_coco.yml) | - |
|
||||
| PP-YOLO-MobileNetV3_large | 剪裁-FPGM | -37% | 12.6 | 608 | - | 22.3 | [下载链接](https://paddledet.bj.bcebos.com/models/slim/ppyolo_mbv3_large_prune_fpgm.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_mbv3_large_coco.yml) | [slim配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/prune/ppyolo_mbv3_large_prune_fpgm.yml) |
|
||||
| YOLOv3-DarkNet53 | baseline | -- | 238.2 | 608 | - | 39.0 | [下载链接](https://paddledet.bj.bcebos.com/models/ppyolo_mbv3_large_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_darknet53_270e_coco.yml) | - |
|
||||
| YOLOv3-DarkNet53 | 剪裁-FPGM | -24% | - | 608 | - | 37.6 | [下载链接](https://paddledet.bj.bcebos.com/models/slim/yolov3_darknet_prune_fpgm.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_darknet53_270e_coco.yml) | [slim配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/prune/yolov3_darknet_prune_fpgm.yml) |
|
||||
| PP-YOLO_R50vd | baseline | -- | 183.3 | 608 | - | 44.8 | [下载链接](https://paddledet.bj.bcebos.com/models/ppyolo_r50vd_dcn_1x_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml) | - |
|
||||
| PP-YOLO_R50vd | 剪裁-FPGM | -35% | - | 608 | - | 42.1 | [下载链接](https://paddledet.bj.bcebos.com/models/slim/ppyolo_r50vd_prune_fpgm.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml) | [slim配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/prune/ppyolo_r50vd_prune_fpgm.yml) |
|
||||
|
||||
说明：
|
||||
- 目前剪裁除RCNN系列模型外，其余模型均已支持。
|
||||
- SD855预测时延为使用PaddleLite部署，使用arm8架构并使用4线程(4 Threads)推理时延。
|
||||
|
||||
### 量化
|
||||
|
||||
#### COCO上benchmark
|
||||
|
||||
| 模型 | 压缩策略 | 输入尺寸 | 模型体积(MB) | 预测时延(V100) | 预测时延(SD855) | Box AP | 下载 | Inference模型下载 | 模型配置文件 | 压缩算法配置文件 |
|
||||
| ------------------ | ------------ | -------- | :---------: | :---------: |:---------: | :---------: | :----------------------------------------------: | :----------------------------------------------: |:------------------------------------------: | :------------------------------------: |
|
||||
| PP-YOLOE-l | baseline | 640 | - | 11.2ms(trt_fp32) | 7.7ms(trt_fp16) | -- | 50.9 | [下载链接](https://paddledet.bj.bcebos.com/models/ppyoloe_crn_l_300e_coco.pdparams) | - | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyoloe/ppyoloe_crn_l_300e_coco.yml) | - |
|
||||
| PP-YOLOE-l | 普通在线量化 | 640 | - | 6.7ms(trt_int8) | -- | 48.8 | [下载链接](https://paddledet.bj.bcebos.com/models/slim/ppyoloe_l_coco_qat.pdparams) | - | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyoloe/ppyoloe_crn_l_300e_coco.yml) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/ppyoloe_l_qat.yml) |
|
||||
| PP-YOLOv2_R50vd | baseline | 640 | 208.6 | 19.1ms | -- | 49.1 | [下载链接](https://paddledet.bj.bcebos.com/models/ppyolov2_r50vd_dcn_365e_coco.pdparams) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/ppyolov2_r50vd_dcn_365e_coco.tar) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml) | - |
|
||||
| PP-YOLOv2_R50vd | PACT在线量化 | 640 | -- | 17.3ms | -- | 48.1 | [下载链接](https://paddledet.bj.bcebos.com/models/slim/ppyolov2_r50vd_dcn_qat.pdparams) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/ppyolov2_r50vd_dcn_qat.tar) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/ppyolov2_r50vd_dcn_qat.yml) |
|
||||
| PP-YOLO_R50vd | baseline | 608 | 183.3 | 17.4ms | -- | 44.8 | [下载链接](https://paddledet.bj.bcebos.com/models/ppyolo_r50vd_dcn_1x_coco.pdparams) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/ppyolo_r50vd_dcn_1x_coco.tar) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml) | - |
|
||||
| PP-YOLO_R50vd | PACT在线量化 | 608 | 67.3 | 13.8ms | -- | 44.3 | [下载链接](https://paddledet.bj.bcebos.com/models/slim/ppyolo_r50vd_qat_pact.pdparams) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/ppyolo_r50vd_qat_pact.tar) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/ppyolo_r50vd_qat_pact.yml) |
|
||||
| PP-YOLO-MobileNetV3_large | baseline | 320 | 18.5 | 2.7ms | 27.9ms | 23.2 | [下载链接](https://paddledet.bj.bcebos.com/models/ppyolo_mbv3_large_coco.pdparams) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/ppyolo_mbv3_large_coco.tar) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_mbv3_large_coco.yml) | - |
|
||||
| PP-YOLO-MobileNetV3_large | 普通在线量化 | 320 | 5.6 | -- | 25.1ms | 24.3 | [下载链接](https://paddledet.bj.bcebos.com/models/slim/ppyolo_mbv3_large_qat.pdparams) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/ppyolo_mbv3_large_qat.tar) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_mbv3_large_coco.yml) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/ppyolo_mbv3_large_qat.yml) |
|
||||
| YOLOv3-MobileNetV1 | baseline | 608 | 94.2 | 8.9ms | 332ms | 29.4 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v1_270e_coco.pdparams) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_270e_coco.tar) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml) | - |
|
||||
| YOLOv3-MobileNetV1 | 普通在线量化 | 608 | 25.4 | 6.6ms | 248ms | 30.5 | [下载链接](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_coco_qat.pdparams) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_coco_qat.tar) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml) | [slim配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/yolov3_mobilenet_v1_qat.yml) |
|
||||
| YOLOv3-MobileNetV3 | baseline | 608 | 90.3 | 9.4ms | 367.2ms | 31.4 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v3_large_270e_coco.pdparams) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v3_large_270e_coco.tar) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v3_large_270e_coco.yml) | - |
|
||||
| YOLOv3-MobileNetV3 | PACT在线量化 | 608 | 24.4 | 8.0ms | 280.0ms | 31.1 | [下载链接](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v3_coco_qat.pdparams) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v3_coco_qat.tar) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v3_large_270e_coco.yml) | [slim配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/yolov3_mobilenet_v3_qat.yml) |
|
||||
| YOLOv3-DarkNet53 | baseline | 608 | 238.2 | 16.0ms | -- | 39.0 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_darknet53_270e_coco.pdparams) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/yolov3_darknet53_270e_coco.tar) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_darknet53_270e_coco.yml) | - |
|
||||
| YOLOv3-DarkNet53 | 普通在线量化 | 608 | 78.8 | 12.4ms | -- | 38.8 | [下载链接](https://paddledet.bj.bcebos.com/models/slim/yolov3_darknet_coco_qat.pdparams) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/yolov3_darknet_coco_qat.tar) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_darknet53_270e_coco.yml) | [slim配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/yolov3_darknet_qat.yml) |
|
||||
| SSD-MobileNet_v1 | baseline | 300 | 22.5 | 4.4ms | 26.6ms | 73.8 | [下载链接](https://paddledet.bj.bcebos.com/models/ssd_mobilenet_v1_300_120e_voc.pdparams) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/ssd_mobilenet_v1_300_120e_voc.tar) |[配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ssd/ssd_mobilenet_v1_300_120e_voc.yml) | - |
|
||||
| SSD-MobileNet_v1 | 普通在线量化 | 300 | 7.1 | -- | 21.5ms | 72.9 | [下载链接](https://paddledet.bj.bcebos.com/models/slim/ssd_mobilenet_v1_300_voc_qat.pdparams) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/ssd_mobilenet_v1_300_voc_qat.tar) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ssd/ssd_mobilenet_v1_300_120e_voc.yml) | [slim配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/ssd_mobilenet_v1_qat.yml) |
|
||||
| Mask-ResNet50-FPN | baseline | (800, 1333) | 174.1 | 359.5ms | -- | 39.2/35.6 | [下载链接](https://paddledet.bj.bcebos.com/models/mask_rcnn_r50_fpn_1x_coco.pdparams) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/mask_rcnn_r50_fpn_1x_coco.tar) |[配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/mask_rcnn/mask_rcnn_r50_fpn_1x_coco.yml) | - |
|
||||
| Mask-ResNet50-FPN | 普通在线量化 | (800, 1333) | -- | -- | -- | 39.7(+0.5)/35.9(+0.3) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/mask_rcnn_r50_fpn_1x_qat.pdparams) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/mask_rcnn_r50_fpn_1x_qat.tar) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/mask_rcnn/mask_rcnn_r50_fpn_1x_coco.yml) | [slim配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/mask_rcnn_r50_fpn_1x_qat.yml) |
|
||||
|
||||
说明：
|
||||
- 上述V100预测时延非量化模型均是使用TensorRT-FP32测试，量化模型均使用TensorRT-INT8测试，并且都包含NMS耗时。
|
||||
- SD855预测时延为使用PaddleLite部署，使用arm8架构并使用4线程(4 Threads)推理时延。
|
||||
- 上述PP-YOLOE模型均在V100，开启TensorRT环境中测速，不包含NMS。（导出模型时指定：-o trt=True exclude_nms=True）
|
||||
|
||||
### 离线量化
|
||||
需要准备val集，用来对离线量化模型进行校准，运行方式：
|
||||
```shell
|
||||
python tools/post_quant.py -c configs/{MODEL.yml} --slim_config configs/slim/post_quant/{SLIM_CONFIG.yml}
|
||||
```
|
||||
例如：
|
||||
```shell
|
||||
python3.7 tools/post_quant.py -c configs/ppyolo/ppyolo_mbv3_large_coco.yml --slim_config=configs/slim/post_quant/ppyolo_mbv3_large_ptq.yml
|
||||
```
|
||||
|
||||
### 蒸馏
|
||||
|
||||
#### COCO上benchmark
|
||||
|
||||
| 模型 | 压缩策略 | 输入尺寸 | Box AP | 下载 | 模型配置文件 | 压缩算法配置文件 |
|
||||
| ------------------ | ------------ | -------- | :---------: | :----------------------------------------------------------: | :----------------------------------------------------------: | :----------------------------------------------------------: |
|
||||
| YOLOv3-MobileNetV1 | baseline | 608 | 29.4 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v1_270e_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml) | - |
|
||||
| YOLOv3-MobileNetV1 | 蒸馏 | 608 | 31.0(+1.6) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_coco_distill.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml) | [slim配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/distill/yolov3_mobilenet_v1_coco_distill.yml) |
|
||||
|
||||
- 具体蒸馏方法和更多检测模型的蒸馏，请参考[蒸馏策略文档](distill/README.md)
|
||||
|
||||
### 蒸馏剪裁联合策略
|
||||
|
||||
#### COCO上benchmark
|
||||
|
||||
| 模型 | 压缩策略 | 输入尺寸 | GFLOPs | 模型体积(MB) | 预测时延(SD855) | Box AP | 下载 | 模型配置文件 | 压缩算法配置文件 |
|
||||
| ------------------ | ------------ | -------- | :---------: |:---------: |:---------: | :---------: |:----------------------------------------------------------: | :----------------------------------------------------------: | :----------------------------------------------------------: |
|
||||
| YOLOv3-MobileNetV1 | baseline | 608 | 24.65 | 94.2 | 332.0ms | 29.4 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v1_270e_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml) | - |
|
||||
| YOLOv3-MobileNetV1 | 蒸馏+剪裁 | 608 | 7.54(-69.4%) | 30.9(-67.2%) | 166.1ms | 28.4(-1.0) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_coco_distill_prune.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml) | [slim配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/extensions/yolov3_mobilenet_v1_coco_distill_prune.yml) |
|
||||
| YOLOv3-MobileNetV1 | 剪裁+量化 | 608 | - | - | - | - | - | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v1_270e_voc.yml) | [slim配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/extensions/yolov3_mobilenetv1_prune_qat.yml) |
|
||||
@@ -0,0 +1,169 @@
|
||||
# Model Compression
|
||||
|
||||
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 |
|
||||
| :---------------------: | :------------------: | :----------------: | :----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------: |
|
||||
| release/2.1 | >= 2.1.0 | 2.1 | Quantitative model exports rely on the latest Paddle Develop branch, available in[PaddlePaddle Daily version](https://www.paddlepaddle.org.cn/documentation/docs/zh/install/Tables.html#whl-dev) |
|
||||
| release/2.0 | >= 2.0.1 | 2.0 | Quantization depends on Paddle 2.1 and PaddleSlim 2.1 |
|
||||
|
||||
|
||||
#### Install PaddleSlim
|
||||
- Method 1: Install it directly：
|
||||
```
|
||||
pip install paddleslim -i https://pypi.tuna.tsinghua.edu.cn/simple
|
||||
```
|
||||
- Method 2: Compile and install：
|
||||
```
|
||||
git clone https://github.com/PaddlePaddle/PaddleSlim.git
|
||||
cd PaddleSlim
|
||||
python setup.py install
|
||||
```
|
||||
|
||||
## Quick Start
|
||||
|
||||
### Train
|
||||
|
||||
```shell
|
||||
python tools/train.py -c configs/{MODEL.yml} --slim_config configs/slim/{SLIM_CONFIG.yml}
|
||||
```
|
||||
|
||||
- `-c`: Specify the model configuration file.
|
||||
- `--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.
|
||||
|
||||
|
||||
### Evaluation
|
||||
|
||||
```shell
|
||||
python tools/eval.py -c configs/{MODEL.yml} --slim_config configs/slim/{SLIM_CONFIG.yml} -o weights=output/{SLIM_CONFIG}/model_final
|
||||
```
|
||||
|
||||
- `-c`: Specify the model configuration file.
|
||||
- `--slim_config`: Specify the compression policy profile.
|
||||
- `-o weights`: Specifies the path of the model trained by the compression algorithm.
|
||||
|
||||
### Test
|
||||
|
||||
```shell
|
||||
python tools/infer.py -c configs/{MODEL.yml} --slim_config configs/slim/{SLIM_CONFIG.yml} \
|
||||
-o weights=output/{SLIM_CONFIG}/model_final
|
||||
--infer_img={IMAGE_PATH}
|
||||
```
|
||||
|
||||
- `-c`: Specify the model configuration file.
|
||||
- `--slim_config`: Specify the compression policy profile.
|
||||
- `-o weights`: Specifies the path of the model trained by the compression algorithm.
|
||||
- `--infer_img`: Specifies the test image path.
|
||||
|
||||
|
||||
## Full Chain Deployment
|
||||
|
||||
### the model is derived from moving to static
|
||||
|
||||
```shell
|
||||
python tools/export_model.py -c configs/{MODEL.yml} --slim_config configs/slim/{SLIM_CONFIG.yml} -o weights=output/{SLIM_CONFIG}/model_final
|
||||
```
|
||||
|
||||
- `-c`: Specify the model configuration file.
|
||||
- `--slim_config`: Specify the compression policy profile.
|
||||
- `-o weights`: Specifies the path of the model trained by the compression algorithm.
|
||||
|
||||
### prediction and deployment
|
||||
|
||||
- Paddle-Inference Prediction：
|
||||
- [Python Deployment](../../deploy/python/README.md)
|
||||
- [C++ Deployment](../../deploy/cpp/README.md)
|
||||
- [TensorRT Predictive Deployment Tutorial](../../deploy/TENSOR_RT.md)
|
||||
- Server deployment: Used[PaddleServing](../../deploy/serving/README.md)
|
||||
- Mobile deployment: Use[Paddle-Lite](../../deploy/lite/README.md) Deploy it on the mobile terminal.
|
||||
|
||||
## Benchmark
|
||||
|
||||
### Pruning
|
||||
|
||||
#### Pascal VOC Benchmark
|
||||
|
||||
| Model | Compression Strategy | GFLOPs | Model Volume(MB) | Input Size | Predict Delay(SD855) | Box AP | Download | Model Configuration File | Compression Algorithm Configuration File |
|
||||
| :----------------: | :-------------------: | :------------: | :--------------: | :--------: | :------------------: | :--------: | :------------------------------------------------------------------------------------------------: | :--------------------------------------------------------------------------------------------------------------------------------: | :---------------------------------------------------------------------------------------------------------------------------------: |
|
||||
| YOLOv3-MobileNetV1 | baseline | 24.13 | 93 | 608 | 332.0ms | 75.1 | [link](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v1_270e_voc.pdparams) | [configuration file](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v1_270e_voc.yml) | - |
|
||||
| YOLOv3-MobileNetV1 | 剪裁-l1_norm(sensity) | 15.78(-34.49%) | 66(-29%) | 608 | - | 78.4(+3.3) | [link](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_voc_prune_l1_norm.pdparams) | [configuration file](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v1_270e_voc.yml) | [slim configuration file](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/prune/yolov3_prune_l1_norm.yml) |
|
||||
|
||||
#### COCO Benchmark
|
||||
| Mode | Compression Strategy | GFLOPs | Model Volume(MB) | Input Size | Predict Delay(SD855) | Box AP | Download | Model Configuration File | Compression Algorithm Configuration File |
|
||||
| :-----------------------: | :------------------: | :----: | :--------------: | :--------: | :------------------: | :----: | :---------------------------------------------------------------------------------------: | :------------------------------------------------------------------------------------------------------------------------------: | :-----------------------------------------------------------------------------------------------------------------------------------------: |
|
||||
| PP-YOLO-MobileNetV3_large | baseline | -- | 18.5 | 608 | 25.1ms | 23.2 | [link](https://paddledet.bj.bcebos.com/models/ppyolo_mbv3_large_coco.pdparams) | [configuration file](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_mbv3_large_coco.yml) | - |
|
||||
| PP-YOLO-MobileNetV3_large | 剪裁-FPGM | -37% | 12.6 | 608 | - | 22.3 | [link](https://paddledet.bj.bcebos.com/models/slim/ppyolo_mbv3_large_prune_fpgm.pdparams) | [configuration file](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_mbv3_large_coco.yml) | [slim configuration file](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/prune/ppyolo_mbv3_large_prune_fpgm.yml) |
|
||||
| YOLOv3-DarkNet53 | baseline | -- | 238.2 | 608 | - | 39.0 | [link](https://paddledet.bj.bcebos.com/models/ppyolo_mbv3_large_coco.pdparams) | [configuration file](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_darknet53_270e_coco.yml) | - |
|
||||
| YOLOv3-DarkNet53 | 剪裁-FPGM | -24% | - | 608 | - | 37.6 | [link](https://paddledet.bj.bcebos.com/models/slim/yolov3_darknet_prune_fpgm.pdparams) | [configuration file](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_darknet53_270e_coco.yml) | [slim configuration file](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/prune/yolov3_darknet_prune_fpgm.yml) |
|
||||
| PP-YOLO_R50vd | baseline | -- | 183.3 | 608 | - | 44.8 | [link](https://paddledet.bj.bcebos.com/models/ppyolo_r50vd_dcn_1x_coco.pdparams) | [configuration file](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml) | - |
|
||||
| PP-YOLO_R50vd | 剪裁-FPGM | -35% | - | 608 | - | 42.1 | [link](https://paddledet.bj.bcebos.com/models/slim/ppyolo_r50vd_prune_fpgm.pdparams) | [configuration file](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml) | [slim configuration file](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/prune/ppyolo_r50vd_prune_fpgm.yml) |
|
||||
|
||||
Description:
|
||||
- Currently, all models except RCNN series models are supported.
|
||||
- The SD855 predicts the delay for deployment using Paddle Lite, using the ARM8 architecture and using 4 Threads (4 Threads) to reason the delay.
|
||||
|
||||
### Quantitative
|
||||
|
||||
#### COCO Benchmark
|
||||
|
||||
| Model | Compression Strategy | Input Size | Model Volume(MB) | Prediction Delay(V100) | Prediction Delay(SD855) | Box AP | Download | Download of Inference Model | Model Configuration File | Compression Algorithm Configuration File |
|
||||
| ------------------------- | -------------------------- | ----------- | :--------------: | :--------------------: | :---------------------: | :-------------------: | :-----------------------------------------------------------------------------------------: | :-----------------------------------------------------------------------------------------: | :----------------------------------------------------------------------------------------------------------------------------------------: | :--------------------------------------------------------------------------------------------------------------------------------------: |
|
||||
| PP-YOLOE-l | baseline | 640 | - | 11.2ms(trt_fp32) | 7.7ms(trt_fp16) | -- | 50.9 | [link](https://paddledet.bj.bcebos.com/models/ppyoloe_crn_l_300e_coco.pdparams) | - | [Configuration File](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyoloe/ppyoloe_crn_l_300e_coco.yml) | - |
|
||||
| PP-YOLOE-l | Common Online quantitative | 640 | - | 6.7ms(trt_int8) | -- | 48.8 | [link](https://paddledet.bj.bcebos.com/models/slim/ppyoloe_l_coco_qat.pdparams) | - | [Configuration File](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyoloe/ppyoloe_crn_l_300e_coco.yml) | [Configuration File](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/ppyoloe_l_qat.yml) |
|
||||
| PP-YOLOv2_R50vd | baseline | 640 | 208.6 | 19.1ms | -- | 49.1 | [link](https://paddledet.bj.bcebos.com/models/ppyolov2_r50vd_dcn_365e_coco.pdparams) | [link](https://paddledet.bj.bcebos.com/models/slim/ppyolov2_r50vd_dcn_365e_coco.tar) | [Configuration File](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml) | - |
|
||||
| PP-YOLOv2_R50vd | PACT Online quantitative | 640 | -- | 17.3ms | -- | 48.1 | [link](https://paddledet.bj.bcebos.com/models/slim/ppyolov2_r50vd_dcn_qat.pdparams) | [link](https://paddledet.bj.bcebos.com/models/slim/ppyolov2_r50vd_dcn_qat.tar) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/ppyolov2_r50vd_dcn_qat.yml) |
|
||||
| PP-YOLO_R50vd | baseline | 608 | 183.3 | 17.4ms | -- | 44.8 | [link](https://paddledet.bj.bcebos.com/models/ppyolo_r50vd_dcn_1x_coco.pdparams) | [link](https://paddledet.bj.bcebos.com/models/slim/ppyolo_r50vd_dcn_1x_coco.tar) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml) | - |
|
||||
| PP-YOLO_R50vd | PACT Online quantitative | 608 | 67.3 | 13.8ms | -- | 44.3 | [link](https://paddledet.bj.bcebos.com/models/slim/ppyolo_r50vd_qat_pact.pdparams) | [link](https://paddledet.bj.bcebos.com/models/slim/ppyolo_r50vd_qat_pact.tar) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/ppyolo_r50vd_qat_pact.yml) |
|
||||
| PP-YOLO-MobileNetV3_large | baseline | 320 | 18.5 | 2.7ms | 27.9ms | 23.2 | [link](https://paddledet.bj.bcebos.com/models/ppyolo_mbv3_large_coco.pdparams) | [link](https://paddledet.bj.bcebos.com/models/slim/ppyolo_mbv3_large_coco.tar) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_mbv3_large_coco.yml) | - |
|
||||
| PP-YOLO-MobileNetV3_large | Common Online quantitative | 320 | 5.6 | -- | 25.1ms | 24.3 | [link](https://paddledet.bj.bcebos.com/models/slim/ppyolo_mbv3_large_qat.pdparams) | [link](https://paddledet.bj.bcebos.com/models/slim/ppyolo_mbv3_large_qat.tar) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_mbv3_large_coco.yml) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/ppyolo_mbv3_large_qat.yml) |
|
||||
| YOLOv3-MobileNetV1 | baseline | 608 | 94.2 | 8.9ms | 332ms | 29.4 | [link](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v1_270e_coco.pdparams) | [link](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_270e_coco.tar) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml) | - |
|
||||
| YOLOv3-MobileNetV1 | Common Online quantitative | 608 | 25.4 | 6.6ms | 248ms | 30.5 | [link](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_coco_qat.pdparams) | [link](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_coco_qat.tar) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml) | [slim Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/yolov3_mobilenet_v1_qat.yml) |
|
||||
| YOLOv3-MobileNetV3 | baseline | 608 | 90.3 | 9.4ms | 367.2ms | 31.4 | [link](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v3_large_270e_coco.pdparams) | [link](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v3_large_270e_coco.tar) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v3_large_270e_coco.yml) | - |
|
||||
| YOLOv3-MobileNetV3 | PACT Online quantitative | 608 | 24.4 | 8.0ms | 280.0ms | 31.1 | [link](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v3_coco_qat.pdparams) | [link](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v3_coco_qat.tar) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v3_large_270e_coco.yml) | [slim Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/yolov3_mobilenet_v3_qat.yml) |
|
||||
| YOLOv3-DarkNet53 | baseline | 608 | 238.2 | 16.0ms | -- | 39.0 | [link](https://paddledet.bj.bcebos.com/models/yolov3_darknet53_270e_coco.pdparams) | [link](https://paddledet.bj.bcebos.com/models/slim/yolov3_darknet53_270e_coco.tar) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_darknet53_270e_coco.yml) | - |
|
||||
| YOLOv3-DarkNet53 | Common Online quantitative | 608 | 78.8 | 12.4ms | -- | 38.8 | [link](https://paddledet.bj.bcebos.com/models/slim/yolov3_darknet_coco_qat.pdparams) | [link](https://paddledet.bj.bcebos.com/models/slim/yolov3_darknet_coco_qat.tar) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_darknet53_270e_coco.yml) | [slim Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/yolov3_darknet_qat.yml) |
|
||||
| SSD-MobileNet_v1 | baseline | 300 | 22.5 | 4.4ms | 26.6ms | 73.8 | [link](https://paddledet.bj.bcebos.com/models/ssd_mobilenet_v1_300_120e_voc.pdparams) | [link](https://paddledet.bj.bcebos.com/models/slim/ssd_mobilenet_v1_300_120e_voc.tar) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ssd/ssd_mobilenet_v1_300_120e_voc.yml) | - |
|
||||
| SSD-MobileNet_v1 | Common Online quantitative | 300 | 7.1 | -- | 21.5ms | 72.9 | [link](https://paddledet.bj.bcebos.com/models/slim/ssd_mobilenet_v1_300_voc_qat.pdparams) | [link](https://paddledet.bj.bcebos.com/models/slim/ssd_mobilenet_v1_300_voc_qat.tar) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ssd/ssd_mobilenet_v1_300_120e_voc.yml) | [slim Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/ssd_mobilenet_v1_qat.yml) |
|
||||
| Mask-ResNet50-FPN | baseline | (800, 1333) | 174.1 | 359.5ms | -- | 39.2/35.6 | [link](https://paddledet.bj.bcebos.com/models/mask_rcnn_r50_fpn_1x_coco.pdparams) | [link](https://paddledet.bj.bcebos.com/models/slim/mask_rcnn_r50_fpn_1x_coco.tar) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/mask_rcnn/mask_rcnn_r50_fpn_1x_coco.yml) | - |
|
||||
| Mask-ResNet50-FPN | Common Online quantitative | (800, 1333) | -- | -- | -- | 39.7(+0.5)/35.9(+0.3) | [link](https://paddledet.bj.bcebos.com/models/slim/mask_rcnn_r50_fpn_1x_qat.pdparams) | [link](https://paddledet.bj.bcebos.com/models/slim/mask_rcnn_r50_fpn_1x_qat.tar) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/mask_rcnn/mask_rcnn_r50_fpn_1x_coco.yml) | [slim Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/mask_rcnn_r50_fpn_1x_qat.yml) |
|
||||
|
||||
Description:
|
||||
- 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 |
|
||||
| ------------------ | -------------------- | ---------- | :--------: | :-------------------------------------------------------------------------------------------: | :----------------------------------------------------------------------------------------------------------------------------------: | :-----------------------------------------------------------------------------------------------------------------------------------------------: |
|
||||
| YOLOv3-MobileNetV1 | baseline | 608 | 29.4 | [link](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v1_270e_coco.pdparams) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml) | - |
|
||||
| YOLOv3-MobileNetV1 | Distillation | 608 | 31.0(+1.6) | [link](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_coco_distill.pdparams) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml) | [slimConfiguration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/distill/yolov3_mobilenet_v1_coco_distill.yml) |
|
||||
|
||||
- For the specific distillation method and more distillation detection models, please refer to [distill](distill/README.md).
|
||||
|
||||
### Distillation Pruning Combined Strategy
|
||||
|
||||
#### COCO Benchmark
|
||||
|
||||
| Model | Compression Strategy | Input Size | GFLOPs | Model Volume(MB) | Prediction Delay(SD855) | Box AP | Download | Model Configuration File | Compression Algorithm Configuration File |
|
||||
| ------------------ | ------------------------ | ---------- | :----------: | :--------------: | :---------------------: | :--------: | :-------------------------------------------------------------------------------------------------: | :----------------------------------------------------------------------------------------------------------------------------------: | :--------------------------------------------------------------------------------------------------------------------------------------------------------: |
|
||||
| YOLOv3-MobileNetV1 | baseline | 608 | 24.65 | 94.2 | 332.0ms | 29.4 | [link](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v1_270e_coco.pdparams) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml) | - |
|
||||
| YOLOv3-MobileNetV1 | Distillation + Tailoring | 608 | 7.54(-69.4%) | 30.9(-67.2%) | 166.1ms | 28.4(-1.0) | [link](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_coco_distill_prune.pdparams) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml) | [slimConfiguration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/extensions/yolov3_mobilenet_v1_coco_distill_prune.yml) |
|
||||
@@ -0,0 +1,212 @@
|
||||
# Distillation(蒸馏)
|
||||
|
||||
## 内容
|
||||
- [YOLOv3模型蒸馏](#YOLOv3模型蒸馏)
|
||||
- [FGD模型蒸馏](#FGD模型蒸馏)
|
||||
- [CWD模型蒸馏](#CWD模型蒸馏)
|
||||
- [LD模型蒸馏](#LD模型蒸馏)
|
||||
- [PPYOLOE模型蒸馏](#PPYOLOE模型蒸馏)
|
||||
- [引用](#引用)
|
||||
|
||||
## YOLOv3模型蒸馏
|
||||
|
||||
以YOLOv3-MobileNetV1为例，使用YOLOv3-ResNet34作为蒸馏训练的teacher网络, 对YOLOv3-MobileNetV1结构的student网络进行蒸馏。
|
||||
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)。
|
||||
为了确定蒸馏的对象，我们首先需要找到student和teacher网络得到的`x,y,w,h,cls,objectness`等Tensor，用teacher得到的结果指导student训练。具体实现可参考[代码](../../../ppdet/slim/distill_loss.py)
|
||||
|
||||
| 模型 | 方案 | 输入尺寸 | epochs | Box mAP | 配置文件 | 下载链接 |
|
||||
| :---------------: | :---------: | :----: | :----: |:-----------: | :--------------: | :------------: |
|
||||
| YOLOv3-ResNet34 | teacher | 608 | 270e | 36.2 | [config](../../yolov3/yolov3_r34_270e_coco.yml) | [download](https://paddledet.bj.bcebos.com/models/yolov3_r34_270e_coco.pdparams) |
|
||||
| YOLOv3-MobileNetV1 | student | 608 | 270e | 29.4 | [config](../../yolov3/yolov3_mobilenet_v1_270e_coco.yml) | [download](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v1_270e_coco.pdparams) |
|
||||
| YOLOv3-MobileNetV1 | distill | 608 | 270e | 31.0(+1.6) | [config](../../yolov3/yolov3_mobilenet_v1_270e_coco.yml),[slim_config](./yolov3_mobilenet_v1_coco_distill.yml) | [download](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_coco_distill.pdparams) |
|
||||
|
||||
<details>
|
||||
<summary> 快速开始 </summary>
|
||||
|
||||
```shell
|
||||
# 单卡训练(不推荐)
|
||||
python tools/train.py -c configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml --slim_config configs/slim/distill/yolov3_mobilenet_v1_coco_distill.yml
|
||||
# 多卡训练
|
||||
python -m paddle.distributed.launch --log_dir=logs/ --gpus 0,1,2,3,4,5,6,7 tools/train.py -c configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml --slim_config configs/slim/distill/yolov3_mobilenet_v1_coco_distill.yml
|
||||
# 评估
|
||||
python tools/eval.py -c configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_coco_distill.pdparams
|
||||
# 预测
|
||||
python tools/infer.py -c configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_coco_distill.pdparams --infer_img=demo/000000014439_640x640.jpg
|
||||
```
|
||||
|
||||
- `-c`: 指定模型配置文件，也是student配置文件。
|
||||
- `--slim_config`: 指定压缩策略配置文件，也是teacher配置文件。
|
||||
|
||||
</details>
|
||||
|
||||
|
||||
## FGD模型蒸馏
|
||||
|
||||
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。
|
||||
在PaddleDetection中，我们实现了FGD算法，并基于RetinaNet算法进行验证，实验结果如下：
|
||||
|
||||
| 模型 | 方案 | 输入尺寸 | epochs | Box mAP | 配置文件 | 下载链接 |
|
||||
| ----------------- | ----------- | ------ | :----: | :-----------: | :--------------: | :------------: |
|
||||
| RetinaNet-ResNet101| teacher | 1333x800 | 2x | 40.6 | [config](../../retinanet/retinanet_r101_fpn_2x_coco.yml) | [download](https://paddledet.bj.bcebos.com/models/retinanet_r101_fpn_2x_coco.pdparams) |
|
||||
| RetinaNet-ResNet50 | student | 1333x800 | 2x | 39.1 | [config](../../retinanet/retinanet_r50_fpn_2x_coco.yml) | [download](https://paddledet.bj.bcebos.com/models/retinanet_r50_fpn_2x_coco.pdparams) |
|
||||
| RetinaNet-ResNet50 | FGD | 1333x800 | 2x | 40.8(+1.7) | [config](../../retinanet/retinanet_r50_fpn_2x_coco.yml),[slim_config](./retinanet_resnet101_coco_distill.yml) | [download](https://paddledet.bj.bcebos.com/models/retinanet_r101_distill_r50_2x_coco.pdparams) |
|
||||
|
||||
<details>
|
||||
<summary> 快速开始 </summary>
|
||||
|
||||
```shell
|
||||
# 单卡训练(不推荐)
|
||||
python tools/train.py -c configs/retinanet/retinanet_r50_fpn_2x_coco.yml --slim_config configs/slim/distill/retinanet_resnet101_coco_distill.yml
|
||||
# 多卡训练
|
||||
python -m paddle.distributed.launch --log_dir=logs/ --gpus 0,1,2,3,4,5,6,7 tools/train.py -c configs/retinanet/retinanet_r50_fpn_2x_coco.yml --slim_config configs/slim/distill/retinanet_resnet101_coco_distill.yml
|
||||
# 评估
|
||||
python tools/eval.py -c configs/retinanet/retinanet_r50_fpn_2x_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/retinanet_r101_distill_r50_2x_coco.pdparams
|
||||
# 预测
|
||||
python tools/infer.py -c configs/retinanet/retinanet_r50_fpn_2x_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/retinanet_r101_distill_r50_2x_coco.pdparams --infer_img=demo/000000014439_640x640.jpg
|
||||
```
|
||||
|
||||
- `-c`: 指定模型配置文件，也是student配置文件。
|
||||
- `--slim_config`: 指定压缩策略配置文件，也是teacher配置文件。
|
||||
|
||||
</details>
|
||||
|
||||
|
||||
## CWD模型蒸馏
|
||||
|
||||
CWD全称为[Channel-wise Knowledge Distillation for Dense Prediction*](https://arxiv.org/pdf/2011.13256.pdf)，通过最小化教师网络与学生网络的通道概率图之间的 Kullback-Leibler (KL) 散度，使得在蒸馏过程更加关注每个通道的最显著的区域，进而提升文本检测与图像分割任务的精度。在PaddleDetection中，我们实现了CWD算法，并基于GFL和RetinaNet模型进行验证，实验结果如下：
|
||||
|
||||
| 模型 | 方案 | 输入尺寸 | epochs | Box mAP | 配置文件 | 下载链接 |
|
||||
| ----------------- | ----------- | ------ | :----: | :-----------: | :--------------: | :------------: |
|
||||
| RetinaNet-ResNet101| teacher | 1333x800 | 2x | 40.6 | [config](../../retinanet/retinanet_r101_fpn_2x_coco.yml) | [download](https://paddledet.bj.bcebos.com/models/retinanet_r101_fpn_2x_coco.pdparams) |
|
||||
| RetinaNet-ResNet50 | student | 1333x800 | 2x | 39.1 | [config](../../retinanet/retinanet_r50_fpn_2x_coco.yml) | [download](https://paddledet.bj.bcebos.com/models/retinanet_r50_fpn_2x_coco.pdparams) |
|
||||
| RetinaNet-ResNet50 | CWD | 1333x800 | 2x | 40.5(+1.4) | [config](../../retinanet/retinanet_r50_fpn_2x_coco.yml),[slim_config](./retinanet_resnet101_coco_distill_cwd.yml) | [download](https://paddledet.bj.bcebos.com/models/retinanet_r50_fpn_2x_coco_cwd.pdparams) |
|
||||
| GFL_ResNet101-vd| teacher | 1333x800 | 2x | 46.8 | [config](../../gfl/gfl_r101vd_fpn_mstrain_2x_coco.yml) | [download](https://paddledet.bj.bcebos.com/models/gfl_r101vd_fpn_mstrain_2x_coco.pdparams) |
|
||||
| GFL_ResNet50 | student | 1333x800 | 1x | 41.0 | [config](../../gfl/gfl_r50_fpn_1x_coco.yml) | [download](https://paddledet.bj.bcebos.com/models/gfl_r50_fpn_1x_coco.pdparams) |
|
||||
| GFL_ResNet50 | CWD | 1333x800 | 2x | 44.0(+3.0) | [config](../../gfl/gfl_r50_fpn_1x_coco.yml),[slim_config](./gfl_r101vd_fpn_coco_distill_cwd.yml) | [download](https://bj.bcebos.com/v1/paddledet/models/gfl_r50_fpn_2x_coco_cwd.pdparams) |
|
||||
|
||||
<details>
|
||||
<summary> 快速开始 </summary>
|
||||
|
||||
```shell
|
||||
# 单卡训练(不推荐)
|
||||
python tools/train.py -c configs/retinanet/retinanet_r50_fpn_2x_coco.yml --slim_config configs/slim/distill/retinanet_resnet101_coco_distill_cwd.yml
|
||||
# 多卡训练
|
||||
python -m paddle.distributed.launch --log_dir=logs/ --gpus 0,1,2,3,4,5,6,7 tools/train.py -c configs/retinanet/retinanet_r50_fpn_2x_coco.yml --slim_config configs/slim/distill/retinanet_resnet101_coco_distill_cwd.yml
|
||||
# 评估
|
||||
python tools/eval.py -c configs/retinanet/retinanet_r50_fpn_2x_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/retinanet_r50_fpn_2x_coco_cwd.pdparams
|
||||
# 预测
|
||||
python tools/infer.py -c configs/retinanet/retinanet_r50_fpn_2x_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/retinanet_r50_fpn_2x_coco_cwd.pdparams --infer_img=demo/000000014439_640x640.jpg
|
||||
|
||||
# 单卡训练(不推荐)
|
||||
python tools/train.py -c configs/gfl/gfl_r50_fpn_1x_coco.yml --slim_config configs/slim/distill/gfl_r101vd_fpn_coco_distill_cwd.yml
|
||||
# 多卡训练
|
||||
python -m paddle.distributed.launch --log_dir=logs/ --gpus 0,1,2,3,4,5,6,7 tools/train.py -c configs/gfl/gfl_r50_fpn_1x_coco.yml --slim_config configs/slim/distill/gfl_r101vd_fpn_coco_distill_cwd.yml
|
||||
# 评估
|
||||
python tools/eval.py -c configs/gfl/gfl_r50_fpn_1x_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/gfl_r50_fpn_2x_coco_cwd.pdparams
|
||||
# 预测
|
||||
python tools/infer.py -c configs/gfl/gfl_r50_fpn_1x_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/gfl_r50_fpn_2x_coco_cwd.pdparams --infer_img=demo/000000014439_640x640.jpg
|
||||
```
|
||||
|
||||
- `-c`: 指定模型配置文件，也是student配置文件。
|
||||
- `--slim_config`: 指定压缩策略配置文件，也是teacher配置文件。
|
||||
|
||||
</details>
|
||||
|
||||
|
||||
## LD模型蒸馏
|
||||
|
||||
LD全称为[Localization Distillation for Dense Object Detection](https://arxiv.org/abs/2102.12252)，将回归框表示为概率分布，把分类任务的KD用在定位任务上，并且使用因地制宜、分而治之的策略，在不同的区域分别学习分类知识与定位知识。在PaddleDetection中，我们实现了LD算法，并基于GFL模型进行验证，实验结果如下：
|
||||
|
||||
| 模型 | 方案 | 输入尺寸 | epochs | Box mAP | 配置文件 | 下载链接 |
|
||||
| ----------------- | ----------- | ------ | :----: | :-----------: | :--------------: | :------------: |
|
||||
| GFL_ResNet101-vd| teacher | 1333x800 | 2x | 46.8 | [config](../../gfl/gfl_r101vd_fpn_mstrain_2x_coco.yml) | [download](https://paddledet.bj.bcebos.com/models/gfl_r101vd_fpn_mstrain_2x_coco.pdparams) |
|
||||
| GFL_ResNet18-vd | student | 1333x800 | 1x | 36.6 | [config](../../gfl/gfl_r18vd_1x_coco.yml) | [download](https://paddledet.bj.bcebos.com/models/gfl_r18vd_1x_coco.pdparams) |
|
||||
| GFL_ResNet18-vd | LD | 1333x800 | 1x | 38.2(+1.6) | [config](../../gfl/gfl_slim_ld_r18vd_1x_coco.yml),[slim_config](./gfl_ld_distill.yml) | [download](https://bj.bcebos.com/v1/paddledet/models/gfl_slim_ld_r18vd_1x_coco.pdparams) |
|
||||
|
||||
<details>
|
||||
<summary> 快速开始 </summary>
|
||||
|
||||
```shell
|
||||
# 单卡训练(不推荐)
|
||||
python tools/train.py -c configs/gfl/gfl_slim_ld_r18vd_1x_coco.yml --slim_config configs/slim/distill/gfl_ld_distill.yml
|
||||
# 多卡训练
|
||||
python -m paddle.distributed.launch --log_dir=logs/ --gpus 0,1,2,3,4,5,6,7 tools/train.py -c configs/gfl/gfl_slim_ld_r18vd_1x_coco.yml --slim_config configs/slim/distill/gfl_ld_distill.yml
|
||||
# 评估
|
||||
python tools/eval.py -c configs/gfl/gfl_slim_ld_r18vd_1x_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/gfl_slim_ld_r18vd_1x_coco.pdparams
|
||||
# 预测
|
||||
python tools/infer.py -c configs/gfl/gfl_slim_ld_r18vd_1x_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/gfl_slim_ld_r18vd_1x_coco.pdparams --infer_img=demo/000000014439_640x640.jpg
|
||||
```
|
||||
|
||||
- `-c`: 指定模型配置文件，也是student配置文件。
|
||||
- `--slim_config`: 指定压缩策略配置文件，也是teacher配置文件。
|
||||
|
||||
</details>
|
||||
|
||||
|
||||
## PPYOLOE模型蒸馏
|
||||
|
||||
PaddleDetection提供了对PPYOLOE+ 进行模型蒸馏的方案，结合了logits蒸馏和feature蒸馏。
|
||||
|
||||
| 模型 | 方案 | 输入尺寸 | epochs | Box mAP | 配置文件 | 下载链接 |
|
||||
| ----------------- | ----------- | ------ | :----: | :-----------: | :--------------: | :------------: |
|
||||
| PP-YOLOE+_x | teacher | 640 | 80e | 54.7 | [config](../../ppyoloe/ppyoloe_plus_crn_x_80e_coco.yml) | [model](https://bj.bcebos.com/v1/paddledet/models/ppyoloe_plus_crn_x_80e_coco.pdparams) |
|
||||
| PP-YOLOE+_l | student | 640 | 80e | 52.9 | [config](../../ppyoloe/ppyoloe_plus_crn_l_80e_coco.yml) | [model](https://bj.bcebos.com/v1/paddledet/models/ppyoloe_plus_crn_l_80e_coco.pdparams) |
|
||||
| PP-YOLOE+_l | distill | 640 | 80e | **54.0(+1.1)** | [config](../../ppyoloe/distill/ppyoloe_plus_crn_l_80e_coco_distill.yml),[slim_config](./ppyoloe_plus_distill_x_distill_l.yml) | [model](https://bj.bcebos.com/v1/paddledet/models/ppyoloe_plus_crn_l_80e_coco_distill.pdparams) |
|
||||
| PP-YOLOE+_l | teacher | 640 | 80e | 52.9 | [config](../../ppyoloe/ppyoloe_plus_crn_l_80e_coco.yml) | [model](https://bj.bcebos.com/v1/paddledet/models/ppyoloe_plus_crn_l_80e_coco.pdparams) |
|
||||
| PP-YOLOE+_m | student | 640 | 80e | 49.8 | [config](../../ppyoloe/ppyoloe_plus_crn_m_80e_coco.yml) | [model](https://bj.bcebos.com/v1/paddledet/models/ppyoloe_plus_crn_m_80e_coco.pdparams) |
|
||||
| PP-YOLOE+_m | distill | 640 | 80e | **51.0(+1.2)** | [config](../../ppyoloe/distill/ppyoloe_plus_crn_m_80e_coco_distill.yml),[slim_config](./ppyoloe_plus_distill_l_distill_m.yml) | [model](https://bj.bcebos.com/v1/paddledet/models/ppyoloe_plus_crn_m_80e_coco_distill.pdparams) |
|
||||
|
||||
<details>
|
||||
<summary> 快速开始 </summary>
|
||||
|
||||
```shell
|
||||
# 单卡训练(不推荐)
|
||||
python tools/train.py -c configs/ppyoloe/distill/ppyoloe_plus_crn_l_80e_coco_distill.yml --slim_config configs/slim/distill/ppyoloe_plus_distill_x_distill_l.yml
|
||||
# 多卡训练
|
||||
python -m paddle.distributed.launch --log_dir=logs/ --gpus 0,1,2,3,4,5,6,7 tools/train.py -c configs/ppyoloe/distill/ppyoloe_plus_crn_l_80e_coco_distill.yml --slim_config configs/slim/distill/ppyoloe_plus_distill_x_distill_l.yml
|
||||
# 评估
|
||||
python tools/eval.py -c configs/ppyoloe/distill/ppyoloe_plus_crn_l_80e_coco_distill.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_plus_crn_l_80e_coco_distill.pdparams
|
||||
# 预测
|
||||
python tools/infer.py -c configs/ppyoloe/distill/ppyoloe_plus_crn_l_80e_coco_distill.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_plus_crn_l_80e_coco_distill.pdparams --infer_img=demo/000000014439_640x640.jpg
|
||||
```
|
||||
|
||||
- `-c`: 指定模型配置文件，也是student配置文件。
|
||||
- `--slim_config`: 指定压缩策略配置文件，也是teacher配置文件。
|
||||
|
||||
</details>
|
||||
|
||||
|
||||
## 引用
|
||||
```
|
||||
@article{mehta2018object,
|
||||
title={Object detection at 200 Frames Per Second},
|
||||
author={Rakesh Mehta and Cemalettin Ozturk},
|
||||
year={2018},
|
||||
eprint={1805.06361},
|
||||
archivePrefix={arXiv},
|
||||
primaryClass={cs.CV}
|
||||
}
|
||||
|
||||
@inproceedings{yang2022focal,
|
||||
title={Focal and global knowledge distillation for detectors},
|
||||
author={Yang, Zhendong and Li, Zhe and Jiang, Xiaohu and Gong, Yuan and Yuan, Zehuan and Zhao, Danpei and Yuan, Chun},
|
||||
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
|
||||
pages={4643--4652},
|
||||
year={2022}
|
||||
}
|
||||
|
||||
@Inproceedings{zheng2022LD,
|
||||
title={Localization Distillation for Dense Object Detection},
|
||||
author= {Zheng, Zhaohui and Ye, Rongguang and Wang, Ping and Ren, Dongwei and Zuo, Wangmeng and Hou, Qibin and Cheng, Mingming},
|
||||
booktitle={CVPR},
|
||||
year={2022}
|
||||
}
|
||||
|
||||
@inproceedings{shu2021channel,
|
||||
title={Channel-wise knowledge distillation for dense prediction},
|
||||
author={Shu, Changyong and Liu, Yifan and Gao, Jianfei and Yan, Zheng and Shen, Chunhua},
|
||||
booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
|
||||
pages={5311--5320},
|
||||
year={2021}
|
||||
}
|
||||
```
|
||||
@@ -0,0 +1,25 @@
|
||||
_BASE_: [
|
||||
'../../gfl/gfl_r18vd_1x_coco.yml',
|
||||
]
|
||||
|
||||
# teacher pretrain model
|
||||
pretrain_weights: https://paddledet.bj.bcebos.com/models/gfl_r101vd_fpn_mstrain_2x_coco.pdparams
|
||||
|
||||
slim: Distill
|
||||
slim_method: LD
|
||||
|
||||
ResNet:
|
||||
depth: 101
|
||||
variant: d
|
||||
norm_type: bn
|
||||
freeze_at: 0
|
||||
return_idx: [1,2,3]
|
||||
num_stages: 4
|
||||
|
||||
TrainReader:
|
||||
batch_transforms:
|
||||
- PadBatch: {pad_to_stride: 32}
|
||||
- Gt2GFLTarget:
|
||||
downsample_ratios: [8, 16, 32, 64, 128]
|
||||
grid_cell_scale: 8
|
||||
compute_vlr_region: True
|
||||
@@ -0,0 +1,16 @@
|
||||
_BASE_: [
|
||||
'../../gfl/gfl_r101vd_fpn_mstrain_2x_coco.yml',
|
||||
]
|
||||
|
||||
pretrain_weights: https://paddledet.bj.bcebos.com/models/gfl_r101vd_fpn_mstrain_2x_coco.pdparams
|
||||
|
||||
slim: Distill
|
||||
slim_method: CWD
|
||||
distill_loss: CWDFeatureLoss
|
||||
distill_loss_name: ['cls_f_4', 'cls_f_3', 'cls_f_2', 'cls_f_1', 'cls_f_0']
|
||||
|
||||
CWDFeatureLoss:
|
||||
student_channels: 80
|
||||
teacher_channels: 80
|
||||
tau: 1.0
|
||||
weight: 5.0
|
||||
@@ -0,0 +1,53 @@
|
||||
# teacher and slim config
|
||||
_BASE_: [
|
||||
'../../ppyoloe/ppyoloe_plus_crn_l_80e_coco.yml',
|
||||
]
|
||||
depth_mult: 1.0
|
||||
width_mult: 1.0
|
||||
for_distill: True
|
||||
architecture: PPYOLOE
|
||||
PPYOLOE:
|
||||
backbone: CSPResNet
|
||||
neck: CustomCSPPAN
|
||||
yolo_head: PPYOLOEHead
|
||||
post_process: ~
|
||||
|
||||
pretrain_weights: https://paddledet.bj.bcebos.com/models/ppyoloe_plus_crn_l_80e_coco.pdparams
|
||||
find_unused_parameters: True
|
||||
|
||||
worker_num: 4
|
||||
TrainReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- RandomDistort: {}
|
||||
- RandomExpand: {fill_value: [123.675, 116.28, 103.53]}
|
||||
- RandomCrop: {}
|
||||
- RandomFlip: {}
|
||||
batch_transforms:
|
||||
- BatchRandomResize: {target_size: [640], random_size: True, random_interp: True, keep_ratio: False}
|
||||
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
|
||||
- Permute: {}
|
||||
- PadGT: {}
|
||||
batch_size: 8
|
||||
shuffle: True
|
||||
drop_last: True
|
||||
use_shared_memory: True
|
||||
collate_batch: True
|
||||
|
||||
|
||||
slim: Distill
|
||||
slim_method: PPYOLOEDistill
|
||||
distill_loss: DistillPPYOLOELoss
|
||||
|
||||
DistillPPYOLOELoss: # L -> M
|
||||
loss_weight: {'logits': 4.0, 'feat': 1.0}
|
||||
logits_distill: True
|
||||
logits_loss_weight: {'class': 1.0, 'iou': 2.5, 'dfl': 0.5}
|
||||
logits_ld_distill: True
|
||||
logits_ld_params: {'weight': 20000, 'T': 10}
|
||||
feat_distill: True
|
||||
feat_distiller: 'fgd' # ['cwd', 'fgd', 'pkd', 'mgd', 'mimic']
|
||||
feat_distill_place: 'neck_feats'
|
||||
teacher_width_mult: 1.0 # L
|
||||
student_width_mult: 0.75 # M
|
||||
feat_out_channels: [768, 384, 192] # The actual channel will multiply width_mult
|
||||
@@ -0,0 +1,53 @@
|
||||
# teacher and slim config
|
||||
_BASE_: [
|
||||
'../../ppyoloe/ppyoloe_plus_crn_m_80e_coco.yml',
|
||||
]
|
||||
depth_mult: 0.67
|
||||
width_mult: 0.75
|
||||
for_distill: True
|
||||
architecture: PPYOLOE
|
||||
PPYOLOE:
|
||||
backbone: CSPResNet
|
||||
neck: CustomCSPPAN
|
||||
yolo_head: PPYOLOEHead
|
||||
post_process: ~
|
||||
|
||||
pretrain_weights: https://paddledet.bj.bcebos.com/models/ppyoloe_plus_crn_m_80e_coco.pdparams
|
||||
find_unused_parameters: True
|
||||
|
||||
worker_num: 4
|
||||
TrainReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- RandomDistort: {}
|
||||
- RandomExpand: {fill_value: [123.675, 116.28, 103.53]}
|
||||
- RandomCrop: {}
|
||||
- RandomFlip: {}
|
||||
batch_transforms:
|
||||
- BatchRandomResize: {target_size: [640], random_size: True, random_interp: True, keep_ratio: False}
|
||||
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
|
||||
- Permute: {}
|
||||
- PadGT: {}
|
||||
batch_size: 8
|
||||
shuffle: True
|
||||
drop_last: True
|
||||
use_shared_memory: True
|
||||
collate_batch: True
|
||||
|
||||
|
||||
slim: Distill
|
||||
slim_method: PPYOLOEDistill
|
||||
distill_loss: DistillPPYOLOELoss
|
||||
|
||||
DistillPPYOLOELoss: # M -> S
|
||||
loss_weight: {'logits': 4.0, 'feat': 1.0}
|
||||
logits_distill: True
|
||||
logits_loss_weight: {'class': 1.0, 'iou': 2.5, 'dfl': 0.5}
|
||||
logits_ld_distill: True
|
||||
logits_ld_params: {'weight': 20000, 'T': 10}
|
||||
feat_distill: True
|
||||
feat_distiller: 'fgd' # ['cwd', 'fgd', 'pkd', 'mgd', 'mimic']
|
||||
feat_distill_place: 'neck_feats'
|
||||
teacher_width_mult: 0.75 # M
|
||||
student_width_mult: 0.5 # S
|
||||
feat_out_channels: [768, 384, 192] # The actual channel will multiply width_mult
|
||||
@@ -0,0 +1,53 @@
|
||||
# teacher and slim config
|
||||
_BASE_: [
|
||||
'../../ppyoloe/ppyoloe_plus_crn_x_80e_coco.yml',
|
||||
]
|
||||
depth_mult: 1.33
|
||||
width_mult: 1.25
|
||||
for_distill: True
|
||||
architecture: PPYOLOE
|
||||
PPYOLOE:
|
||||
backbone: CSPResNet
|
||||
neck: CustomCSPPAN
|
||||
yolo_head: PPYOLOEHead
|
||||
post_process: ~
|
||||
|
||||
pretrain_weights: https://paddledet.bj.bcebos.com/models/ppyoloe_plus_crn_x_80e_coco.pdparams
|
||||
find_unused_parameters: True
|
||||
|
||||
worker_num: 4
|
||||
TrainReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- RandomDistort: {}
|
||||
- RandomExpand: {fill_value: [123.675, 116.28, 103.53]}
|
||||
- RandomCrop: {}
|
||||
- RandomFlip: {}
|
||||
batch_transforms:
|
||||
- BatchRandomResize: {target_size: [640], random_size: True, random_interp: True, keep_ratio: False}
|
||||
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
|
||||
- Permute: {}
|
||||
- PadGT: {}
|
||||
batch_size: 8
|
||||
shuffle: True
|
||||
drop_last: True
|
||||
use_shared_memory: True
|
||||
collate_batch: True
|
||||
|
||||
|
||||
slim: Distill
|
||||
slim_method: PPYOLOEDistill
|
||||
distill_loss: DistillPPYOLOELoss
|
||||
|
||||
DistillPPYOLOELoss: # X -> L
|
||||
loss_weight: {'logits': 4.0, 'feat': 1.0}
|
||||
logits_distill: True
|
||||
logits_loss_weight: {'class': 1.0, 'iou': 2.5, 'dfl': 0.5}
|
||||
logits_ld_distill: True
|
||||
logits_ld_params: {'weight': 20000, 'T': 10}
|
||||
feat_distill: True
|
||||
feat_distiller: 'fgd' # ['cwd', 'fgd', 'pkd', 'mgd', 'mimic']
|
||||
feat_distill_place: 'neck_feats'
|
||||
teacher_width_mult: 1.25 # X
|
||||
student_width_mult: 1.0 # L
|
||||
feat_out_channels: [768, 384, 192] # The actual channel will multiply width_mult
|
||||
@@ -0,0 +1,19 @@
|
||||
_BASE_: [
|
||||
'../../retinanet/retinanet_r101_fpn_2x_coco.yml',
|
||||
]
|
||||
|
||||
pretrain_weights: https://paddledet.bj.bcebos.com/models/retinanet_r101_fpn_2x_coco.pdparams
|
||||
|
||||
slim: Distill
|
||||
slim_method: FGD
|
||||
distill_loss: FGDFeatureLoss
|
||||
distill_loss_name: ['neck_f_4', 'neck_f_3', 'neck_f_2', 'neck_f_1', 'neck_f_0']
|
||||
|
||||
FGDFeatureLoss:
|
||||
student_channels: 256
|
||||
teacher_channels: 256
|
||||
temp: 0.5
|
||||
alpha_fgd: 0.001
|
||||
beta_fgd: 0.0005
|
||||
gamma_fgd: 0.0005
|
||||
lambda_fgd: 0.000005
|
||||
@@ -0,0 +1,17 @@
|
||||
_BASE_: [
|
||||
'../../retinanet/retinanet_r101_fpn_2x_coco.yml',
|
||||
]
|
||||
|
||||
pretrain_weights: https://paddledet.bj.bcebos.com/models/retinanet_r101_fpn_2x_coco.pdparams
|
||||
|
||||
|
||||
slim: Distill
|
||||
slim_method: CWD
|
||||
distill_loss: CWDFeatureLoss
|
||||
distill_loss_name: ['cls_f_4', 'cls_f_3', 'cls_f_2', 'cls_f_1', 'cls_f_0']
|
||||
|
||||
CWDFeatureLoss:
|
||||
student_channels: 80
|
||||
teacher_channels: 80
|
||||
tau: 1.0
|
||||
weight: 5.0
|
||||
@@ -0,0 +1,12 @@
|
||||
_BASE_: [
|
||||
'../../yolov3/yolov3_r34_270e_coco.yml',
|
||||
]
|
||||
|
||||
pretrain_weights: https://paddledet.bj.bcebos.com/models/yolov3_r34_270e_coco.pdparams
|
||||
|
||||
|
||||
slim: Distill
|
||||
distill_loss: DistillYOLOv3Loss
|
||||
|
||||
DistillYOLOv3Loss:
|
||||
weight: 1000
|
||||
@@ -0,0 +1,24 @@
|
||||
_BASE_: [
|
||||
'../../yolov3/yolov3_r34_270e_coco.yml',
|
||||
]
|
||||
|
||||
pretrain_weights: https://paddledet.bj.bcebos.com/models/yolov3_r34_270e_coco.pdparams
|
||||
|
||||
slim: DistillPrune
|
||||
|
||||
distill_loss: DistillYOLOv3Loss
|
||||
|
||||
DistillYOLOv3Loss:
|
||||
weight: 1000
|
||||
|
||||
pruner: Pruner
|
||||
|
||||
Pruner:
|
||||
criterion: l1_norm
|
||||
pruned_params: ['conv2d_27.w_0', 'conv2d_28.w_0', 'conv2d_29.w_0',
|
||||
'conv2d_30.w_0', 'conv2d_31.w_0', 'conv2d_32.w_0',
|
||||
'conv2d_34.w_0', 'conv2d_35.w_0', 'conv2d_36.w_0',
|
||||
'conv2d_37.w_0', 'conv2d_38.w_0', 'conv2d_39.w_0',
|
||||
'conv2d_41.w_0', 'conv2d_42.w_0', 'conv2d_43.w_0',
|
||||
'conv2d_44.w_0', 'conv2d_45.w_0', 'conv2d_46.w_0']
|
||||
pruned_ratios: [0.5,0.5,0.5,0.5,0.5,0.5,0.7,0.7,0.7,0.7,0.7,0.7,0.8,0.8,0.8,0.8,0.8,0.8]
|
||||
@@ -0,0 +1,19 @@
|
||||
# Weights of yolov3_mobilenet_v1_voc
|
||||
pretrain_weights: https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v1_270e_voc.pdparams
|
||||
slim: PrunerQAT
|
||||
|
||||
PrunerQAT:
|
||||
criterion: fpgm
|
||||
pruned_params: ['conv2d_27.w_0', 'conv2d_28.w_0', 'conv2d_29.w_0',
|
||||
'conv2d_30.w_0', 'conv2d_31.w_0', 'conv2d_32.w_0',
|
||||
'conv2d_34.w_0', 'conv2d_35.w_0', 'conv2d_36.w_0',
|
||||
'conv2d_37.w_0', 'conv2d_38.w_0', 'conv2d_39.w_0',
|
||||
'conv2d_41.w_0', 'conv2d_42.w_0', 'conv2d_43.w_0',
|
||||
'conv2d_44.w_0', 'conv2d_45.w_0', 'conv2d_46.w_0']
|
||||
pruned_ratios: [0.1,0.2,0.2,0.2,0.2,0.1,0.2,0.3,0.3,0.3,0.2,0.1,0.3,0.4,0.4,0.4,0.4,0.3]
|
||||
print_prune_params: False
|
||||
quant_config: {
|
||||
'weight_quantize_type': 'channel_wise_abs_max', 'activation_quantize_type': 'moving_average_abs_max',
|
||||
'weight_bits': 8, 'activation_bits': 8, 'dtype': 'int8', 'window_size': 10000, 'moving_rate': 0.9,
|
||||
'quantizable_layer_type': ['Conv2D', 'Linear']}
|
||||
print_qat_model: True
|
||||
@@ -0,0 +1,85 @@
|
||||
weights: https://paddledet.bj.bcebos.com/models/pretrained/ESNet_x1_0_pretrained.pdparams
|
||||
slim: OFA
|
||||
OFA:
|
||||
ofa_config:
|
||||
task: expand_ratio
|
||||
expand_ratio: [0.5, 1]
|
||||
|
||||
skip_neck: True
|
||||
skip_head: True
|
||||
|
||||
RunConfig:
|
||||
# Skip the output layer of each block by layer name
|
||||
skip_layers: ['backbone._conv1._conv','backbone.2_1._conv_linear_1._conv',
|
||||
'backbone.2_1._conv_linear_2._conv', 'backbone.2_1._conv_dw_mv1._conv',
|
||||
'backbone.2_1._conv_pw_mv1._conv', 'backbone.2_2._conv_linear._conv',
|
||||
'backbone.2_3._conv_linear._conv', 'backbone.3_1._conv_linear_1._conv',
|
||||
'backbone.3_1._conv_linear_2._conv', 'backbone.3_1._conv_dw_mv1._conv',
|
||||
'backbone.3_1._conv_pw_mv1._conv', 'backbone.3_2._conv_linear._conv',
|
||||
'backbone.3_3._conv_linear._conv', 'backbone.3_4._conv_linear._conv',
|
||||
'backbone.3_5._conv_linear._conv', 'backbone.3_6._conv_linear._conv',
|
||||
'backbone.3_7._conv_linear._conv', 'backbone.4_1._conv_linear_1._conv',
|
||||
'backbone.4_1._conv_linear_2._conv', 'backbone.4_1._conv_dw_mv1._conv',
|
||||
'backbone.4_1._conv_pw_mv1._conv', 'backbone.4_2._conv_linear._conv',
|
||||
'backbone.4_3._conv_linear._conv']
|
||||
|
||||
# For block-wise search, make layers in each block in the same search space
|
||||
same_search_space: [
|
||||
['backbone.2_1._conv_dw_1._conv', 'backbone.2_1._conv_pw_2._conv',
|
||||
'backbone.2_1._conv_dw_2._conv', 'backbone.2_1._se.conv1', 'backbone.2_1._se.conv2'],
|
||||
['backbone.2_2._conv_pw._conv', 'backbone.2_2._conv_dw._conv',
|
||||
'backbone.2_2._se.conv1', 'backbone.2_2._se.conv2'],
|
||||
['backbone.2_3._conv_pw._conv', 'backbone.2_3._conv_dw._conv',
|
||||
'backbone.2_3._se.conv1', 'backbone.2_3._se.conv2'],
|
||||
['backbone.3_1._conv_dw_1._conv', 'backbone.3_1._conv_pw_2._conv',
|
||||
'backbone.3_1._conv_dw_2._conv', 'backbone.3_1._se.conv1', 'backbone.3_1._se.conv2'],
|
||||
['backbone.3_2._conv_pw._conv', 'backbone.3_2._conv_dw._conv',
|
||||
'backbone.3_2._se.conv1', 'backbone.3_2._se.conv2'],
|
||||
['backbone.3_3._conv_pw._conv', 'backbone.3_3._conv_dw._conv',
|
||||
'backbone.3_3._se.conv1', 'backbone.3_3._se.conv2'],
|
||||
['backbone.3_4._conv_pw._conv', 'backbone.3_4._conv_dw._conv',
|
||||
'backbone.3_4._se.conv1', 'backbone.3_4._se.conv2'],
|
||||
['backbone.3_5._conv_pw._conv', 'backbone.3_5._conv_dw._conv',
|
||||
'backbone.3_5._se.conv1', 'backbone.3_5._se.conv2'],
|
||||
['backbone.3_6._conv_pw._conv', 'backbone.3_6._conv_dw._conv',
|
||||
'backbone.3_6._se.conv1', 'backbone.3_6._se.conv2'],
|
||||
['backbone.3_7._conv_pw._conv', 'backbone.3_7._conv_dw._conv',
|
||||
'backbone.3_7._se.conv1', 'backbone.3_7._se.conv2'],
|
||||
['backbone.4_1._conv_dw_1._conv', 'backbone.4_1._conv_pw_2._conv',
|
||||
'backbone.4_1._conv_dw_2._conv', 'backbone.4_1._se.conv1', 'backbone.4_1._se.conv2'],
|
||||
['backbone.4_2._conv_pw._conv', 'backbone.4_2._conv_dw._conv',
|
||||
'backbone.4_2._se.conv1', 'backbone.4_2._se.conv2'],
|
||||
['backbone.4_3._conv_pw._conv', 'backbone.4_3._conv_dw._conv',
|
||||
'backbone.4_3._se.conv1', 'backbone.4_3._se.conv2']]
|
||||
|
||||
# demo expand ratio
|
||||
# Generally, for expand ratio, float in (0, 1] is available.
|
||||
# But please be careful if the model is complicated.
|
||||
# For picodet, there are many split and concat, the choice of channel number is important.
|
||||
ofa_layers:
|
||||
'backbone.2_1._conv_dw_1._conv':
|
||||
'expand_ratio': [0.5, 1]
|
||||
'backbone.2_2._conv_pw._conv':
|
||||
'expand_ratio': [0.5, 1]
|
||||
'backbone.2_3._conv_pw._conv':
|
||||
'expand_ratio': [0.5, 1]
|
||||
'backbone.3_1._conv_dw_1._conv':
|
||||
'expand_ratio': [0.5, 1]
|
||||
'backbone.3_2._conv_pw._conv':
|
||||
'expand_ratio': [0.5, 1]
|
||||
'backbone.3_3._conv_pw._conv':
|
||||
'expand_ratio': [0.5, 1]
|
||||
'backbone.3_4._conv_pw._conv':
|
||||
'expand_ratio': [0.5, 1]
|
||||
'backbone.3_5._conv_pw._conv':
|
||||
'expand_ratio': [0.5, 1]
|
||||
'backbone.3_6._conv_pw._conv':
|
||||
'expand_ratio': [0.5, 1]
|
||||
'backbone.3_7._conv_pw._conv':
|
||||
'expand_ratio': [0.5, 1]
|
||||
'backbone.4_1._conv_dw_1._conv':
|
||||
'expand_ratio': [0.5, 1]
|
||||
'backbone.4_2._conv_pw._conv':
|
||||
'expand_ratio': [0.5, 1]
|
||||
'backbone.4_3._conv_pw._conv':
|
||||
'expand_ratio': [0.5, 1]
|
||||
@@ -0,0 +1,10 @@
|
||||
weights: https://paddledet.bj.bcebos.com/models/mask_rcnn_r50_fpn_1x_coco.pdparams
|
||||
slim: PTQ
|
||||
|
||||
PTQ:
|
||||
ptq_config: {
|
||||
'activation_quantizer': 'HistQuantizer',
|
||||
'upsample_bins': 127,
|
||||
'hist_percent': 0.999}
|
||||
quant_batch_num: 10
|
||||
fuse: True
|
||||
@@ -0,0 +1,10 @@
|
||||
weights: https://paddledet.bj.bcebos.com/models/mot/mcfairmot_dla34_30e_1088x608_visdrone_vehicle_bytetracker.pdparams
|
||||
slim: PTQ
|
||||
|
||||
PTQ:
|
||||
ptq_config: {
|
||||
'activation_quantizer': 'HistQuantizer',
|
||||
'upsample_bins': 127,
|
||||
'hist_percent': 0.999}
|
||||
quant_batch_num: 10
|
||||
fuse: True
|
||||
@@ -0,0 +1,10 @@
|
||||
weights: https://paddledet.bj.bcebos.com/models/picodet_s_320_coco.pdparams
|
||||
slim: PTQ
|
||||
|
||||
PTQ:
|
||||
ptq_config: {
|
||||
'activation_quantizer': 'HistQuantizer',
|
||||
'upsample_bins': 127,
|
||||
'hist_percent': 0.999}
|
||||
quant_batch_num: 10
|
||||
fuse: True
|
||||
@@ -0,0 +1,10 @@
|
||||
weights: https://paddledet.bj.bcebos.com/models/ppyolo_mbv3_large_coco.pdparams
|
||||
slim: PTQ
|
||||
|
||||
PTQ:
|
||||
ptq_config: {
|
||||
'activation_quantizer': 'HistQuantizer',
|
||||
'upsample_bins': 127,
|
||||
'hist_percent': 0.999}
|
||||
quant_batch_num: 10
|
||||
fuse: True
|
||||
@@ -0,0 +1,10 @@
|
||||
weights: https://paddledet.bj.bcebos.com/models/ppyolo_r50vd_dcn_1x_coco.pdparams
|
||||
slim: PTQ
|
||||
|
||||
PTQ:
|
||||
ptq_config: {
|
||||
'activation_quantizer': 'HistQuantizer',
|
||||
'upsample_bins': 127,
|
||||
'hist_percent': 0.999}
|
||||
quant_batch_num: 10
|
||||
fuse: True
|
||||
@@ -0,0 +1,10 @@
|
||||
weights: https://paddledet.bj.bcebos.com/models/ppyoloe_crn_s_300e_coco.pdparams
|
||||
slim: PTQ
|
||||
|
||||
PTQ:
|
||||
ptq_config: {
|
||||
'activation_quantizer': 'HistQuantizer',
|
||||
'upsample_bins': 127,
|
||||
'hist_percent': 0.999}
|
||||
quant_batch_num: 10
|
||||
fuse: True
|
||||
@@ -0,0 +1,10 @@
|
||||
weights: https://paddledet.bj.bcebos.com/models/keypoint/tinypose_128x96.pdparams
|
||||
slim: PTQ
|
||||
|
||||
PTQ:
|
||||
ptq_config: {
|
||||
'activation_quantizer': 'HistQuantizer',
|
||||
'upsample_bins': 127,
|
||||
'hist_percent': 0.999}
|
||||
quant_batch_num: 10
|
||||
fuse: True
|
||||
@@ -0,0 +1,10 @@
|
||||
weights: https://paddledet.bj.bcebos.com/models/yolov3_darknet53_270e_coco.pdparams
|
||||
slim: PTQ
|
||||
|
||||
PTQ:
|
||||
ptq_config: {
|
||||
'activation_quantizer': 'HistQuantizer',
|
||||
'upsample_bins': 127,
|
||||
'hist_percent': 0.999}
|
||||
quant_batch_num: 10
|
||||
fuse: True
|
||||
@@ -0,0 +1,16 @@
|
||||
pretrain_weights: https://paddledet.bj.bcebos.com/models/faster_rcnn_r50_fpn_1x_coco.pdparams
|
||||
slim: Pruner
|
||||
|
||||
Pruner:
|
||||
criterion: fpgm
|
||||
pruned_params: ['conv2d_27.w_0', 'conv2d_28.w_0', 'conv2d_29.w_0',
|
||||
'conv2d_30.w_0', 'conv2d_31.w_0', 'conv2d_32.w_0',
|
||||
'conv2d_33.w_0', 'conv2d_34.w_0', 'conv2d_35.w_0',
|
||||
'conv2d_36.w_0', 'conv2d_37.w_0', 'conv2d_38.w_0',
|
||||
'conv2d_39.w_0', 'conv2d_40.w_0', 'conv2d_41.w_0',
|
||||
'conv2d_42.w_0', 'conv2d_43.w_0', 'conv2d_44.w_0',
|
||||
'conv2d_45.w_0', 'conv2d_46.w_0', 'conv2d_47.w_0',
|
||||
'conv2d_48.w_0', 'conv2d_49.w_0', 'conv2d_50.w_0',
|
||||
'conv2d_51.w_0', 'conv2d_52.w_0']
|
||||
pruned_ratios: [0.1,0.2,0.2,0.2,0.2,0.1,0.2,0.3,0.3,0.3,0.2,0.1,0.3,0.4,0.4,0.4,0.4,0.3,0.4,0.4,0.4,0.4,0.4,0.4,0.4,0.4]
|
||||
print_params: False
|
||||
@@ -0,0 +1,11 @@
|
||||
pretrain_weights: https://paddledet.bj.bcebos.com/models/picodet_m_320_coco.pdparams
|
||||
slim: UnstructuredPruner
|
||||
|
||||
UnstructuredPruner:
|
||||
stable_epochs: 0
|
||||
pruning_epochs: 150
|
||||
tunning_epochs: 150
|
||||
pruning_steps: 300
|
||||
ratio: 0.75
|
||||
initial_ratio: 0.15
|
||||
prune_params_type: conv1x1_only
|
||||
@@ -0,0 +1,11 @@
|
||||
pretrain_weights: https://paddledet.bj.bcebos.com/models/picodet_m_320_coco.pdparams
|
||||
slim: UnstructuredPruner
|
||||
|
||||
UnstructuredPruner:
|
||||
stable_epochs: 0
|
||||
pruning_epochs: 150
|
||||
tunning_epochs: 150
|
||||
pruning_steps: 300
|
||||
ratio: 0.85
|
||||
initial_ratio: 0.20
|
||||
prune_params_type: conv1x1_only
|
||||
@@ -0,0 +1,9 @@
|
||||
pretrain_weights: https://paddledet.bj.bcebos.com/models/ppyolo_mbv3_large_coco.pdparams
|
||||
slim: Pruner
|
||||
|
||||
Pruner:
|
||||
criterion: fpgm
|
||||
pruned_params: ['conv2d_62.w_0', 'conv2d_63.w_0', 'conv2d_64.w_0',
|
||||
'conv2d_65.w_0', 'conv2d_66.w_0', 'conv2d_67.w_0']
|
||||
pruned_ratios: [0.75, 0.75, 0.75, 0.75, 0.75, 0.75]
|
||||
print_params: True
|
||||
@@ -0,0 +1,13 @@
|
||||
pretrain_weights: https://paddledet.bj.bcebos.com/models/ppyolo_r50vd_dcn_1x_coco.pdparams
|
||||
slim: Pruner
|
||||
|
||||
Pruner:
|
||||
criterion: fpgm
|
||||
pruned_params: ['conv2d_56.w_0', 'conv2d_57.w_0', 'conv2d_58.w_0',
|
||||
'conv2d_59.w_0', 'conv2d_60.w_0', 'conv2d_61.w_0',
|
||||
'conv2d_63.w_0', 'conv2d_64.w_0', 'conv2d_65.w_0',
|
||||
'conv2d_66.w_0', 'conv2d_67.w_0', 'conv2d_68.w_0',
|
||||
'conv2d_70.w_0', 'conv2d_71.w_0', 'conv2d_72.w_0',
|
||||
'conv2d_73.w_0', 'conv2d_74.w_0', 'conv2d_75.w_0']
|
||||
pruned_ratios: [0.75,0.75,0.75,0.75,0.75,0.75,0.75,0.75,0.75,0.75,0.75,0.75,0.875,0.875,0.875,0.875,0.875,0.875]
|
||||
print_params: False
|
||||
@@ -0,0 +1,13 @@
|
||||
pretrain_weights: https://paddledet.bj.bcebos.com/models/yolov3_darknet53_270e_coco.pdparams
|
||||
slim: Pruner
|
||||
|
||||
Pruner:
|
||||
criterion: fpgm
|
||||
pruned_params: ['conv2d_52.w_0', 'conv2d_53.w_0', 'conv2d_54.w_0',
|
||||
'conv2d_55.w_0', 'conv2d_56.w_0', 'conv2d_57.w_0',
|
||||
'conv2d_59.w_0', 'conv2d_60.w_0', 'conv2d_61.w_0',
|
||||
'conv2d_62.w_0', 'conv2d_63.w_0', 'conv2d_64.w_0',
|
||||
'conv2d_66.w_0', 'conv2d_67.w_0', 'conv2d_68.w_0',
|
||||
'conv2d_69.w_0', 'conv2d_70.w_0', 'conv2d_71.w_0']
|
||||
pruned_ratios: [0.75,0.75,0.75,0.75,0.75,0.75,0.75,0.75,0.75,0.75,0.75,0.75,0.875,0.875,0.875,0.875,0.875,0.875]
|
||||
print_params: True
|
||||
@@ -0,0 +1,14 @@
|
||||
# Weights of yolov3_mobilenet_v1_voc
|
||||
pretrain_weights: https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v1_270e_voc.pdparams
|
||||
slim: Pruner
|
||||
|
||||
Pruner:
|
||||
criterion: fpgm
|
||||
pruned_params: ['conv2d_27.w_0', 'conv2d_28.w_0', 'conv2d_29.w_0',
|
||||
'conv2d_30.w_0', 'conv2d_31.w_0', 'conv2d_32.w_0',
|
||||
'conv2d_34.w_0', 'conv2d_35.w_0', 'conv2d_36.w_0',
|
||||
'conv2d_37.w_0', 'conv2d_38.w_0', 'conv2d_39.w_0',
|
||||
'conv2d_41.w_0', 'conv2d_42.w_0', 'conv2d_43.w_0',
|
||||
'conv2d_44.w_0', 'conv2d_45.w_0', 'conv2d_46.w_0']
|
||||
pruned_ratios: [0.1,0.2,0.2,0.2,0.2,0.1,0.2,0.3,0.3,0.3,0.2,0.1,0.3,0.4,0.4,0.4,0.4,0.3]
|
||||
print_params: False
|
||||
@@ -0,0 +1,14 @@
|
||||
# Weights of yolov3_mobilenet_v1_voc
|
||||
pretrain_weights: https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v1_270e_voc.pdparams
|
||||
slim: Pruner
|
||||
|
||||
Pruner:
|
||||
criterion: l1_norm
|
||||
pruned_params: ['conv2d_27.w_0', 'conv2d_28.w_0', 'conv2d_29.w_0',
|
||||
'conv2d_30.w_0', 'conv2d_31.w_0', 'conv2d_32.w_0',
|
||||
'conv2d_34.w_0', 'conv2d_35.w_0', 'conv2d_36.w_0',
|
||||
'conv2d_37.w_0', 'conv2d_38.w_0', 'conv2d_39.w_0',
|
||||
'conv2d_41.w_0', 'conv2d_42.w_0', 'conv2d_43.w_0',
|
||||
'conv2d_44.w_0', 'conv2d_45.w_0', 'conv2d_46.w_0']
|
||||
pruned_ratios: [0.1,0.2,0.2,0.2,0.2,0.1,0.2,0.3,0.3,0.3,0.2,0.1,0.3,0.4,0.4,0.4,0.4,0.3]
|
||||
print_params: False
|
||||
@@ -0,0 +1,24 @@
|
||||
pretrain_weights: https://paddledet.bj.bcebos.com/models/mask_rcnn_r50_fpn_1x_coco.pdparams
|
||||
slim: QAT
|
||||
|
||||
QAT:
|
||||
quant_config: {
|
||||
'weight_quantize_type': 'channel_wise_abs_max', 'activation_quantize_type': 'moving_average_abs_max',
|
||||
'weight_bits': 8, 'activation_bits': 8, 'dtype': 'int8', 'window_size': 10000, 'moving_rate': 0.9,
|
||||
'quantizable_layer_type': ['Conv2D', 'Linear']}
|
||||
print_model: True
|
||||
|
||||
epoch: 5
|
||||
|
||||
TrainReader:
|
||||
batch_size: 1
|
||||
|
||||
LearningRate:
|
||||
base_lr: 0.001
|
||||
schedulers:
|
||||
- !PiecewiseDecay
|
||||
gamma: 0.1
|
||||
milestones: [3, 4]
|
||||
- !LinearWarmup
|
||||
start_factor: 0.001
|
||||
steps: 100
|
||||
@@ -0,0 +1,22 @@
|
||||
pretrain_weights: https://paddledet.bj.bcebos.com/models/picodet_s_416_coco_lcnet.pdparams
|
||||
slim: QAT
|
||||
|
||||
QAT:
|
||||
quant_config: {
|
||||
'activation_preprocess_type': 'PACT',
|
||||
'weight_quantize_type': 'channel_wise_abs_max', 'activation_quantize_type': 'moving_average_abs_max',
|
||||
'weight_bits': 8, 'activation_bits': 8, 'dtype': 'int8', 'window_size': 10000, 'moving_rate': 0.9,
|
||||
'quantizable_layer_type': ['Conv2D', 'Linear']}
|
||||
print_model: False
|
||||
|
||||
TrainReader:
|
||||
batch_size: 48
|
||||
|
||||
LearningRate:
|
||||
base_lr: 0.024
|
||||
schedulers:
|
||||
- !CosineDecay
|
||||
max_epochs: 300
|
||||
- !LinearWarmup
|
||||
start_factor: 0.1
|
||||
steps: 300
|
||||
@@ -0,0 +1,26 @@
|
||||
pretrain_weights: https://paddledet.bj.bcebos.com/models/picodet_s_320_coco.pdparams
|
||||
slim: QAT
|
||||
|
||||
QAT:
|
||||
quant_config: {
|
||||
'activation_preprocess_type': 'PACT',
|
||||
'weight_quantize_type': 'channel_wise_abs_max', 'activation_quantize_type': 'moving_average_abs_max',
|
||||
'weight_bits': 8, 'activation_bits': 8, 'dtype': 'int8', 'window_size': 10000, 'moving_rate': 0.9,
|
||||
'quantizable_layer_type': ['Conv2D', 'Linear']}
|
||||
print_model: False
|
||||
|
||||
epoch: 50
|
||||
LearningRate:
|
||||
base_lr: 0.001
|
||||
schedulers:
|
||||
- !PiecewiseDecay
|
||||
gamma: 0.1
|
||||
milestones:
|
||||
- 30
|
||||
- 40
|
||||
- !LinearWarmup
|
||||
start_factor: 0.
|
||||
steps: 100
|
||||
|
||||
TrainReader:
|
||||
batch_size: 96
|
||||
@@ -0,0 +1,16 @@
|
||||
pretrain_weights: https://paddledet.bj.bcebos.com/models/ppyolo_mbv3_large_coco.pdparams
|
||||
slim: QAT
|
||||
|
||||
QAT:
|
||||
quant_config: {
|
||||
'weight_quantize_type': 'channel_wise_abs_max', 'activation_quantize_type': 'moving_average_abs_max',
|
||||
'weight_bits': 8, 'activation_bits': 8, 'dtype': 'int8', 'window_size': 10000, 'moving_rate': 0.99,
|
||||
'quantizable_layer_type': ['Conv2D', 'Linear']}
|
||||
print_model: True
|
||||
|
||||
PPYOLOFPN:
|
||||
in_channels: [160, 368]
|
||||
coord_conv: true
|
||||
conv_block_num: 0
|
||||
spp: true
|
||||
drop_block: false
|
||||
@@ -0,0 +1,39 @@
|
||||
pretrain_weights: https://paddledet.bj.bcebos.com/models/ppyolo_r50vd_dcn_1x_coco.pdparams
|
||||
slim: QAT
|
||||
|
||||
QAT:
|
||||
quant_config: {
|
||||
'activation_preprocess_type': 'PACT',
|
||||
'weight_quantize_type': 'channel_wise_abs_max', 'activation_quantize_type': 'moving_average_abs_max',
|
||||
'weight_bits': 8, 'activation_bits': 8, 'dtype': 'int8', 'window_size': 10000, 'moving_rate': 0.9,
|
||||
'quantizable_layer_type': ['Conv2D', 'Linear']}
|
||||
print_model: True
|
||||
|
||||
epoch: 50
|
||||
|
||||
LearningRate:
|
||||
base_lr: 0.0005
|
||||
schedulers:
|
||||
- !PiecewiseDecay
|
||||
gamma: 0.1
|
||||
milestones:
|
||||
- 30
|
||||
- 45
|
||||
- !LinearWarmup
|
||||
start_factor: 0.
|
||||
steps: 1000
|
||||
|
||||
OptimizerBuilder:
|
||||
optimizer:
|
||||
momentum: 0.9
|
||||
type: Momentum
|
||||
regularizer:
|
||||
factor: 0.0005
|
||||
type: L2
|
||||
|
||||
PPYOLOFPN:
|
||||
coord_conv: true
|
||||
block_size: 3
|
||||
keep_prob: 0.9
|
||||
spp: true
|
||||
drop_block: false
|
||||
@@ -0,0 +1,26 @@
|
||||
pretrain_weights: https://paddledet.bj.bcebos.com/models/ppyoloe_crn_l_300e_coco.pdparams
|
||||
slim: QAT
|
||||
|
||||
QAT:
|
||||
quant_config: {
|
||||
'weight_quantize_type': 'channel_wise_abs_max', 'activation_quantize_type': 'moving_average_abs_max',
|
||||
'weight_bits': 8, 'activation_bits': 8, 'dtype': 'int8', 'window_size': 10000, 'moving_rate': 0.9,
|
||||
'quantizable_layer_type': ['Conv2D', 'Linear']}
|
||||
print_model: True
|
||||
|
||||
epoch: 30
|
||||
snapshot_epoch: 5
|
||||
LearningRate:
|
||||
base_lr: 0.001
|
||||
schedulers:
|
||||
- !PiecewiseDecay
|
||||
gamma: 0.1
|
||||
milestones:
|
||||
- 10
|
||||
- 20
|
||||
- !LinearWarmup
|
||||
start_factor: 0.
|
||||
steps: 100
|
||||
|
||||
TrainReader:
|
||||
batch_size: 8
|
||||
@@ -0,0 +1,33 @@
|
||||
pretrain_weights: https://paddledet.bj.bcebos.com/models/ppyolov2_r50vd_dcn_365e_coco.pdparams
|
||||
slim: QAT
|
||||
|
||||
QAT:
|
||||
quant_config: {
|
||||
'activation_preprocess_type': 'PACT',
|
||||
'weight_quantize_type': 'channel_wise_abs_max', 'activation_quantize_type': 'moving_average_abs_max',
|
||||
'weight_bits': 8, 'activation_bits': 8, 'dtype': 'int8', 'window_size': 10000, 'moving_rate': 0.9,
|
||||
'quantizable_layer_type': ['Conv2D', 'Linear']}
|
||||
print_model: True
|
||||
|
||||
epoch: 50
|
||||
snapshot_epoch: 8
|
||||
LearningRate:
|
||||
base_lr: 0.0005
|
||||
schedulers:
|
||||
- !PiecewiseDecay
|
||||
gamma: 0.1
|
||||
milestones:
|
||||
- 30
|
||||
- 45
|
||||
- !LinearWarmup
|
||||
start_factor: 0.
|
||||
steps: 2000
|
||||
|
||||
TrainReader:
|
||||
batch_size: 8
|
||||
|
||||
PPYOLOPAN:
|
||||
drop_block: false
|
||||
block_size: 3
|
||||
keep_prob: 0.9
|
||||
spp: true
|
||||
@@ -0,0 +1,9 @@
|
||||
pretrain_weights: https://paddlemodels.bj.bcebos.com/object_detection/dygraph/ssd_mobilenet_v1_300_120e_voc.pdparams
|
||||
slim: QAT
|
||||
|
||||
QAT:
|
||||
quant_config: {
|
||||
'weight_quantize_type': 'channel_wise_abs_max', 'activation_quantize_type': 'moving_average_abs_max',
|
||||
'weight_bits': 8, 'activation_bits': 8, 'dtype': 'int8', 'window_size': 10000, 'moving_rate': 0.9,
|
||||
'quantizable_layer_type': ['Conv2D', 'Linear']}
|
||||
print_model: True
|
||||
@@ -0,0 +1,26 @@
|
||||
pretrain_weights: https://paddledet.bj.bcebos.com/models/keypoint/tinypose_128x96.pdparams
|
||||
slim: QAT
|
||||
|
||||
QAT:
|
||||
quant_config: {
|
||||
'activation_preprocess_type': 'PACT',
|
||||
'weight_quantize_type': 'channel_wise_abs_max', 'activation_quantize_type': 'moving_average_abs_max',
|
||||
'weight_bits': 8, 'activation_bits': 8, 'dtype': 'int8', 'window_size': 10000, 'moving_rate': 0.9,
|
||||
'quantizable_layer_type': ['Conv2D', 'Linear']}
|
||||
print_model: False
|
||||
|
||||
epoch: 50
|
||||
LearningRate:
|
||||
base_lr: 0.001
|
||||
schedulers:
|
||||
- !PiecewiseDecay
|
||||
gamma: 0.1
|
||||
milestones:
|
||||
- 30
|
||||
- 40
|
||||
- !LinearWarmup
|
||||
start_factor: 0.
|
||||
steps: 100
|
||||
|
||||
TrainReader:
|
||||
batch_size: 256
|
||||
@@ -0,0 +1,34 @@
|
||||
pretrain_weights: https://paddledet.bj.bcebos.com/models/yolov3_darknet53_270e_coco.pdparams
|
||||
slim: QAT
|
||||
|
||||
QAT:
|
||||
quant_config: {
|
||||
'weight_quantize_type': 'channel_wise_abs_max', 'activation_quantize_type': 'moving_average_abs_max',
|
||||
'weight_bits': 8, 'activation_bits': 8, 'dtype': 'int8', 'window_size': 10000, 'moving_rate': 0.9,
|
||||
'quantizable_layer_type': ['Conv2D', 'Linear']}
|
||||
print_model: True
|
||||
|
||||
epoch: 50
|
||||
|
||||
TrainReader:
|
||||
batch_size: 8
|
||||
|
||||
LearningRate:
|
||||
base_lr: 0.0001
|
||||
schedulers:
|
||||
- !PiecewiseDecay
|
||||
gamma: 0.1
|
||||
milestones:
|
||||
- 30
|
||||
- 45
|
||||
- !LinearWarmup
|
||||
start_factor: 0.
|
||||
steps: 1000
|
||||
|
||||
OptimizerBuilder:
|
||||
optimizer:
|
||||
momentum: 0.9
|
||||
type: Momentum
|
||||
regularizer:
|
||||
factor: 0.0005
|
||||
type: L2
|
||||
@@ -0,0 +1,10 @@
|
||||
# Weights of yolov3_mobilenet_v1_coco
|
||||
pretrain_weights: https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v1_270e_coco.pdparams
|
||||
slim: QAT
|
||||
|
||||
QAT:
|
||||
quant_config: {
|
||||
'weight_quantize_type': 'channel_wise_abs_max', 'activation_quantize_type': 'moving_average_abs_max',
|
||||
'weight_bits': 8, 'activation_bits': 8, 'dtype': 'int8', 'window_size': 10000, 'moving_rate': 0.9,
|
||||
'quantizable_layer_type': ['Conv2D', 'Linear']}
|
||||
print_model: True
|
||||
@@ -0,0 +1,24 @@
|
||||
# Weights of yolov3_mobilenet_v3_coco
|
||||
pretrain_weights: https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v3_large_270e_coco.pdparams
|
||||
slim: QAT
|
||||
|
||||
QAT:
|
||||
quant_config: {
|
||||
'activation_preprocess_type': 'PACT',
|
||||
'weight_quantize_type': 'channel_wise_abs_max', 'activation_quantize_type': 'moving_average_abs_max',
|
||||
'weight_bits': 8, 'activation_bits': 8, 'dtype': 'int8', 'window_size': 10000, 'moving_rate': 0.9,
|
||||
'quantizable_layer_type': ['Conv2D', 'Linear']}
|
||||
print_model: True
|
||||
|
||||
epoch: 50
|
||||
LearningRate:
|
||||
base_lr: 0.0001
|
||||
schedulers:
|
||||
- !PiecewiseDecay
|
||||
gamma: 0.1
|
||||
milestones:
|
||||
- 35
|
||||
- 45
|
||||
- !LinearWarmup
|
||||
start_factor: 0.
|
||||
steps: 1000
|
||||
Reference in New Issue
Block a user