Skip to content

P
PaddleDetection

PaddlePaddle / PaddleDetection
大约 1 年 前同步成功

通知 695
Star 11112
Fork 2696
P
PaddleDetection
切换分支/标签
查找文件 普通视图历史永久链接Permalink
README.md 15.7 KB
Newer Older
G
[cherry-pick] Fix v0.2 docs (#258)  
Guanghua Yu 已提交
1 
# 压缩benchmark
Q
Update docs (#251)  
qingqing01 已提交
2 3 4 5 6 7 8 9 10 11 

在PaddleDetection, 提供了基于PaddleSlim进行模型压缩的完整教程和实验结果。详细教程请参考：

- [量化](quantization)
- [裁剪](prune)
- [蒸馏](distillation)
- [搜索](nas)

下面给出压缩的benchmark实验结果。
K
add slim model zoo (#193)  
Kaipeng Deng 已提交
12 13 14 15 16 17 18 19 
## 测试环境

- Python 2.7.1
- PaddlePaddle >=1.6
- CUDA 9.0
- cuDNN >=7.4
- NCCL 2.1.2
K
Add distill prune demo (#210)  
Kaipeng Deng 已提交
20 
## 剪裁模型库
K
add slim model zoo (#193)  
Kaipeng Deng 已提交
21 22 23 

### 训练策略
G
[cherry-pick] Fix v0.2 docs (#258)  
Guanghua Yu 已提交
24 
- 剪裁模型训练时使用[PaddleDetection模型库](https://paddledetection.readthedocs.io/zh/latest/MODEL_ZOO_cn.html)发布的模型权重作为预训练权重。
K
Add distill prune demo (#210)  
Kaipeng Deng 已提交
25 26 27 
- 剪裁训练使用模型默认配置，即除`pretrained_weights`外配置不变。
- 剪裁模型全部为基于敏感度的卷积通道剪裁。
- YOLOv3模型主要剪裁`yolo_head`部分，即剪裁参数如下。
K
add slim model zoo (#193)  
Kaipeng Deng 已提交
28 29 30 31 

```
--pruned_params="yolo_block.0.0.0.conv.weights,yolo_block.0.0.1.conv.weights,yolo_block.0.1.0.conv.weights,yolo_block.0.1.1.conv.weights,yolo_block.0.2.conv.weights,yolo_block.0.tip.conv.weights,yolo_block.1.0.0.conv.weights,yolo_block.1.0.1.conv.weights,yolo_block.1.1.0.conv.weights,yolo_block.1.1.1.conv.weights,yolo_block.1.2.conv.weights,yolo_block.1.tip.conv.weights,yolo_block.2.0.0.conv.weights,yolo_block.2.0.1.conv.weights,yolo_block.2.1.0.conv.weights,yolo_block.2.1.1.conv.weights,yolo_block.2.2.conv.weights,yolo_block.2.tip.conv.weights"
```
K
Add distill prune demo (#210)  
Kaipeng Deng 已提交
32 
- YOLOv3模型剪裁中剪裁策略`r578`表示`yolo_head`中三个输出分支一次使用`0.5, 0.7, 0.8`的剪裁率剪裁，即剪裁率如下。
K
add slim model zoo (#193)  
Kaipeng Deng 已提交
33 34 35 36 37 

```
--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"
```
K
Add distill prune demo (#210)  
Kaipeng Deng 已提交
38 
- YOLOv3模型剪裁中剪裁策略`sensity`表示`yolo_head`中各参数剪裁率如下，该剪裁率为使用`yolov3_mobilnet_v1`模型在COCO数据集上敏感度实验分析得出。
K
add slim model zoo (#193)  
Kaipeng Deng 已提交
39 40 41 42 43 44 45 

```
--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"
```

### YOLOv3 on COCO
K
update prune model zoo (#232)  
Kaipeng Deng 已提交
46 47 48 49 50 51 52 53 54 55 
| 骨架网络         |  剪裁策略 | FLOPs剪裁率 | 模型体积剪裁率 | 输入尺寸 | Box AP  |                           下载                          |
| :----------------| :-------: | :---------: | :------------: | :------: | :-----: | :-----------------------------------------------------: |
| ResNet50-vd-dcn  |  sensity  |   18.41%    |     15.46%     |   608    |  39.8   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_r50vd_dcn_prune1x.tar) |
| ResNet50-vd-dcn  |   r578    |   43.69%    |     36.61%     |   608    |  38.3   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_r50vd_dcn_prune578.tar) |
| MobileNetV1      |  sensity  |   28.76%    |     28.54%     |   608    |  30.2   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_prune1x.tar) |
| MobileNetV1      |  sensity  |   28.76%    |     28.54%     |   416    |  29.7   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_prune1x.tar) |
| MobileNetV1      |  sensity  |   28.76%    |     28.54%     |   320    |  27.2   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_prune1x.tar) |
| MobileNetV1      |   r578    |   67.56%    |     66.90%     |   608    |  27.8   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_prune578.tar) |
| MobileNetV1      |   r578    |   67.56%    |     66.90%     |   416    |  26.8   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_prune578.tar) |
| MobileNetV1      |   r578    |   67.56%    |     66.90%     |   320    |  24.0   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_prune578.tar) |
K
add slim model zoo (#193)  
Kaipeng Deng 已提交
56 57 58 

### YOLOv3 on Pascal VOC
K
update prune model zoo (#232)  
Kaipeng Deng 已提交
59 60 61 62 63 64 65 66 
| 骨架网络         |  剪裁策略 | FLOPs剪裁率 | 模型体积剪裁率 | 输入尺寸 | Box AP  |                           下载                          |
| :----------------| :-------: | :---------: | :------------: | :------: | :-----: | :-----------------------------------------------------: |
| MobileNetV1      |  sensity  |   34.55%    |     28.75%     |   608    |  78.4   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune1x.tar) |
| MobileNetV1      |  sensity  |   34.55%    |     28.75%     |   416    |  78.7   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune1x.tar) |
| MobileNetV1      |  sensity  |   34.55%    |     28.75%     |   320    |  76.1   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune1x.tar) |
| MobileNetV1      |   r578    |   69.57%    |     67.00%     |   608    |  77.6   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune578.tar) |
| MobileNetV1      |   r578    |   69.57%    |     67.00%     |   416    |  77.7   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune578.tar) |
| MobileNetV1      |   r578    |   69.57%    |     67.00%     |   320    |  75.5   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune578.tar) |
K
add slim model zoo (#193)  
Kaipeng Deng 已提交
67
K
Add distill prune demo (#210)  
Kaipeng Deng 已提交
68 69 
### 蒸馏通道剪裁模型
G
[cherry-pick] Fix v0.2 docs (#258)  
Guanghua Yu 已提交
70 
可通过高精度模型蒸馏通道剪裁后模型的方式，训练方法及相关示例见[蒸馏通道剪裁模型](https://github.com/PaddlePaddle/PaddleDetection/blob/master/slim/extensions/distill_pruned_model/distill_pruned_model_demo.ipynb)。
K
Add distill prune demo (#210)  
Kaipeng Deng 已提交
71 72 73 

COCO数据集上蒸馏通道剪裁模型库如下。
K
update prune model zoo (#232)  
Kaipeng Deng 已提交
74 75 76 77 78 79 
| 骨架网络         |  剪裁策略 | FLOPs剪裁率 | 模型体积剪裁率 | 输入尺寸 |      teacher模型       | Box AP  |                           下载                          |
| :----------------| :-------: | :---------: | :------------: | :------: | :--------------------- | :-----: | :-----------------------------------------------------: |
| ResNet50-vd-dcn  |   r578    |   43.69%    |     36.61%     |   608    | YOLOv3-ResNet50-vd-dcn |  39.7   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_r50vd_dcn_prune578_distill.tar) |
| MobileNetV1      |   r578    |   67.56%    |     66.90%     |   608    | YOLOv3-ResNet34        |  29.0   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_prune578_distillby_r34.tar) |
| MobileNetV1      |   r578    |   67.56%    |     66.90%     |   416    | YOLOv3-ResNet34        |  28.0   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_prune578_distillby_r34.tar) |
| MobileNetV1      |   r578    |   67.56%    |     66.90%     |   320    | YOLOv3-ResNet34        |  25.1   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_prune578_distillby_r34.tar) |
K
Add distill prune demo (#210)  
Kaipeng Deng 已提交
80 81 82 

Pascal VOC数据集上蒸馏通道剪裁模型库如下。
K
update prune model zoo (#232)  
Kaipeng Deng 已提交
83 84 85 86 87 
| 骨架网络         |  剪裁策略 | FLOPs剪裁率 | 模型体积剪裁率 | 输入尺寸 |      teacher模型       | Box AP  |                           下载                          |
| :----------------| :-------: | :---------: | :------------: | :------: | :--------------------- | :-----: | :-----------------------------------------------------: |
| MobileNetV1      |   r578    |   69.57%    |     67.00%     |   608    | YOLOv3-ResNet34        |  78.8   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune578_distillby_r34.tar) |
| MobileNetV1      |   r578    |   69.57%    |     67.00%     |   416    | YOLOv3-ResNet34        |  78.7   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune578_distillby_r34.tar) |
| MobileNetV1      |   r578    |   69.57%    |     67.00%     |   320    | YOLOv3-ResNet34        |  76.3   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune578_distillby_r34.tar) |
K
Add distill prune demo (#210)  
Kaipeng Deng 已提交
88
K
[cherry-pick] add prune benchmark (#262)  
Kaipeng Deng 已提交
89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 
### YOLOv3通道剪裁模型推理时延

- 时延单位均为`ms/images`
- Tesla P4时延为单卡并开启TensorRT推理时延
- 高通835/高通855/麒麟970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

| 骨架网络         | 数据集 | 剪裁策略 | FLOPs剪裁率 | 模型体积剪裁率 | 输入尺寸 | Tesla P4 | 麒麟970 | 高通835 | 高通855 |
| :--------------- | :----: | :------: | :---------: | :------------: | :------: | :------: | :-----: | :-----: | :-----: |
| MobileNetV1      |  VOC   | baseline |      -      |       -        |   608    |  16.556  | 748.404 | 734.970 | 289.878 |
| MobileNetV1      |  VOC   | baseline |      -      |       -        |   416    |   9.031  | 371.214 | 349.065 | 140.877 |
| MobileNetV1      |  VOC   | baseline |      -      |       -        |   320    |   6.235  | 221.705 | 200.498 |  80.515 |
| MobileNetV1      |  VOC   |   r578   |   69.57%    |     67.00%     |   608    |  10.064  | 314.531 | 323.537 | 123.414 |
| MobileNetV1      |  VOC   |   r578   |   69.57%    |     67.00%     |   416    |   5.478  | 151.562 | 146.014 |  56.420 |
| MobileNetV1      |  VOC   |   r578   |   69.57%    |     67.00%     |   320    |   3.880  |  91.132 |  87.440 |  31.470 |
| ResNet50-vd-dcn  |  COCO  | baseline |      -      |       -        |   608    |  36.127  |    -    |    -    |    -    |
| ResNet50-vd-dcn  |  COCO  | baseline |      -      |       -        |   416    |  20.437  |    -    |    -    |    -    |
| ResNet50-vd-dcn  |  COCO  | baseline |      -      |       -        |   320    |  14.037  |    -    |    -    |    -    |
| ResNet50-vd-dcn  |  COCO  | sensity  |   18.41%    |     15.46%     |   608    |  33.245  |    -    |    -    |    -    |
| ResNet50-vd-dcn  |  COCO  | sensity  |   18.41%    |     15.46%     |   416    |  19.246  |    -    |    -    |    -    |
| ResNet50-vd-dcn  |  COCO  | sensity  |   18.41%    |     15.46%     |   320    |  13.656  |    -    |    -    |    -    |
| ResNet50-vd-dcn  |  COCO  |   r578   |   43.69%    |     36.61%     |   608    |  29.138  |    -    |    -    |    -    |
| ResNet50-vd-dcn  |  COCO  |   r578   |   43.69%    |     36.61%     |   416    |  16.439  |    -    |    -    |    -    |
| ResNet50-vd-dcn  |  COCO  |   r578   |   43.69%    |     36.61%     |   320    |  11.339  |    -    |    -    |    -    |
K
add slim model zoo (#193)  
Kaipeng Deng 已提交
113 114 115 116 117 

## 蒸馏模型库

### 训练策略
G
[cherry-pick] Fix v0.2 docs (#258)  
Guanghua Yu 已提交
118 
- 蒸馏模型训练时teacher模型使用[PaddleDetection模型库](https://paddledetection.readthedocs.io/zh/latest/MODEL_ZOO_cn.html)发布的模型权重作为预训练权重。
K
add slim model zoo (#193)  
Kaipeng Deng 已提交
119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 
- 蒸馏模型训练时student模型使用backbone的预训练权重

### YOLOv3 on COCO

| 骨架网络         |    蒸馏策略   | 输入尺寸 | Box AP  |                           下载                          |
| :----------------| :-----------: | :------: |:------: | :-----------------------------------------------------: |
| MobileNetV1      | split_distiil |   608    |  31.4   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_distilled.tar) |
| MobileNetV1      | split_distiil |   416    |  30.0   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_distilled.tar) |
| MobileNetV1      | split_distiil |   320    |  27.1   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_distilled.tar) |

### YOLOv3 on Pascal VOC

| 骨架网络         |    蒸馏策略   | 输入尺寸 | Box AP  |                           下载                          |
| :----------------| :-----------: | :------: |:------: | :-----------------------------------------------------: |
| MobileNetV1      |  l2_distiil   |   608    |  79.0   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_voc_distilled.tar) |
| MobileNetV1      |  l2_distiil   |   416    |  78.2   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_voc_distilled.tar) |
| MobileNetV1      |  l2_distiil   |   320    |  75.5   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_voc_distilled.tar) |

## 量化模型库

### 训练策略

- 量化策略`post`为使用离线量化得到的模型，`aware`为在线量化训练得到的模型。

### YOLOv3 on COCO

| 骨架网络         | 预训练权重 | 量化策略 | 输入尺寸 | Box AP  |                           下载                          |
| :----------------| :--------: | :------: | :------: |:------: | :-----------------------------------------------------: |
| MobileNetV1      |  ImageNet  |   post   |   608    |  27.9   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_quant_post.tar) |
| MobileNetV1      |  ImageNet  |   post   |   416    |  28.0   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_quant_post.tar) |
| MobileNetV1      |  ImageNet  |   post   |   320    |  26.0   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_quant_post.tar) |
| MobileNetV1      |  ImageNet  |  aware   |   608    |  28.1   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_quant_aware.tar) |
| MobileNetV1      |  ImageNet  |  aware   |   416    |  28.2   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_quant_aware.tar) |
| MobileNetV1      |  ImageNet  |  aware   |   320    |  25.8   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_quant_aware.tar) |
| ResNet34         |  ImageNet  |   post   |   608    |  35.7   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r34_coco_quant_post.tar) |
| ResNet34         |  ImageNet  |  aware   |   608    |  35.2   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r34_coco_quant_aware.tar) |
| ResNet34         |  ImageNet  |  aware   |   416    |  33.3   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r34_coco_quant_aware.tar) |
| ResNet34         |  ImageNet  |  aware   |   320    |  30.3   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r34_coco_quant_aware.tar) |
| R50vd-dcn        | object365  |  aware   |   608    |  40.6   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_obj365_pretrained_coco_quant_aware.tar) |
| R50vd-dcn        | object365  |  aware   |   416    |  37.5   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_obj365_pretrained_coco_quant_aware.tar) |
| R50vd-dcn        | object365  |  aware   |   320    |  34.1   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_obj365_pretrained_coco_quant_aware.tar) |

### BlazeFace on WIDER FACE

| 模型             | 量化策略 | 输入尺寸 | Easy Set | Medium Set | Hard Set |                           下载                          |
| :--------------- | :------: | :------: | :------: | :--------: | :------: | :-----------------------------------------------------: |
| BlazeFace        |   post   |   640    |   87.8   |    85.1    |   74.9   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/blazeface_origin_quant_post.tar) |
| BlazeFace        |  aware   |   640    |   90.5   |    87.9    |   77.6   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/blazeface_origin_quant_aware.tar) |
| BlazeFace-Lite   |   post   |   640    |   89.4   |    86.7    |   75.7   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/blazeface_lite_quant_post.tar) |
| BlazeFace-Lite   |  aware   |   640    |   89.7   |    87.3    |   77.0   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/blazeface_lite_quant_aware.tar) |
| BlazeFace-NAS    |   post   |   640    |   81.6   |    78.3    |   63.6   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/blazeface_nas_quant_post.tar) |
| BlazeFace-NAS    |  aware   |   640    |   83.1   |    79.7    |   64.2   | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/blazeface_nas_quant_aware.tar) |