Skip to content

P
PaddleDetection

PaddlePaddle / PaddleDetection
大约 2 年 前同步成功

通知 708
Star 11112
Fork 2696
P
PaddleDetection
“1ca5cf77c8259ebc53cbc50f52e1fab4f62b1c56”上不存在“git@gitcode.net:s920243400/PaddleDetection.git”
未验证 提交 91609c99 编写于 作者: G Guanghua Yu 提交者： GitHub
浏览文件
操作

[cherry-pcik]add Distill model and refactor slim (#2541) 

* add Distill model

* add distill+prune
上级 395d07d3
显示空白变更内容
内联 并排
Showing with 439 addition and 127 deletion
configs/slim/README.md
浏览文件 @ 91609c99
...@@ -4,33 +4,11 @@ ...@@ -4,33 +4,11 @@
- [剪裁](prune) - [剪裁](prune)
- [量化](quant) - [量化](quant)
- [蒸馏](distill)
- [联合策略](extensions)
推荐您使用剪裁和蒸馏联合训练，或者使用剪裁和量化，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。 推荐您使用剪裁和蒸馏联合训练，或者使用剪裁和量化，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。
## 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 | 289.9ms | 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/master/dygraph/configs/yolov3/yolov3_mobilenet_v1_270e_voc.yml) | - |
| YOLOv3-MobileNetV1 | 剪裁-l1_norm(sensity) | 15.78(-34.49%) | 66(-29%) | 608 | - | 77.6(+2.5) | [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/master/dygraph/configs/yolov3/yolov3_mobilenet_v1_270e_voc.yml) | [slim配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/master/dygraph/configs/slim/prune/yolov3_prune_l1_norm.yml) |
### 量化
#### 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 | 28.8 | [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/master/dygraph/configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml) | - |
| YOLOv3-MobileNetV1 | 普通在线量化 | 608 | 27.5 (-1.3) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_coco_qat.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/master/dygraph/configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml) | [slim配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/master/dygraph/configs/slim/quant/yolov3_mobilenet_v1_qat.yml) |
| YOLOv3-MobileNetV3 | baseline | 608 | 31.4 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v3_large_270e_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/master/dygraph/configs/yolov3/yolov3_mobilenet_v3_large_270e_coco.yml) | - |
| YOLOv3-MobileNetV3 | PACT在线量化 | 608 | 29.0 (-2.4) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v3_coco_qat.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/master/dygraph/configs/yolov3/yolov3_mobilenet_v3_large_270e_coco.yml) | [slim配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/master/dygraph/configs/slim/quant/yolov3_mobilenet_v3_qat.yml) |
- 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
## 实验环境 ## 实验环境
- Python 3.7+ - Python 3.7+
...@@ -39,6 +17,8 @@ ...@@ -39,6 +17,8 @@
- CUDA 9.0+ - CUDA 9.0+
- cuDNN >=7.5 - cuDNN >=7.5
**注意：** 量化训练需要依赖Paddle develop分支，可在[PaddlePaddle每日版本](https://www.paddlepaddle.org.cn/documentation/docs/zh/install/Tables.html#whl-dev)中下载安装合适的PaddlePaddle版本。
## 快速开始 ## 快速开始
### 训练 ### 训练
...@@ -84,3 +64,56 @@ python tools/export_model.py -c configs/{MODEL.yml} --slim_config configs/slim/{ ...@@ -84,3 +64,56 @@ python tools/export_model.py -c configs/{MODEL.yml} --slim_config configs/slim/{
- `-c`: 指定模型配置文件。 - `-c`: 指定模型配置文件。
- `--slim_config`: 指定压缩策略配置文件。 - `--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。 - `-o weights`: 指定压缩算法训好的模型路径。
## 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 | 289.9ms | 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 | - | 77.6(+2.5) | [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/slim/prune/yolov3_prune_l1_norm.yml) |
- 目前剪裁支持YOLO系列、SSD、TTFNet、BlazeFace，其余模型正在开发支持中。
- 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 | Box AP | 下载 | 模型配置文件 | 压缩算法配置文件 |
| ------------------ | ------------ | -------- | :---------: | :----------------------------------------------------------: | :----------------------------------------------------------: | :----------------------------------------------------------: |
| YOLOv3-MobileNetV1 | baseline | 608 | 28.8 | [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 | 普通在线量化 | 608 | 30.3 (+1.5) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_coco_qat.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/quant/yolov3_mobilenet_v1_qat.yml) |
| YOLOv3-MobileNetV3 | baseline | 608 | 31.4 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v3_large_270e_coco.pdparams) | [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 | 29.5 (-1.9) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v3_coco_qat.pdparams) | [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 | 39.0 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_darknet53_270e_coco.pdparams) | [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 | 38.7 (-0.3) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/yolov3_darknet_coco_qat.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/quant/yolov3_darknet_qat.yml) |
| SSD-MobileNet_v1 | baseline | 300 | 73.8 | [下载链接](https://paddledet.bj.bcebos.com/models/ssd_mobilenet_v1_300_120e_voc.pdparams) | [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 | 73.1(-0.7) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/ssd_mobilenet_v1_300_voc_qat.pdparams) | [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) | 39.2/35.6 | [下载链接](https://paddledet.bj.bcebos.com/models/mask_rcnn_r50_fpn_1x_coco.pdparams) | [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) | [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) |
### 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 | 28.8 | [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 | 29.0 | [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) | Box AP | 下载 | 模型配置文件 | 压缩算法配置文件 |
| ------------------ | ------------ | -------- | :---------: |:---------: | :---------: |:----------------------------------------------------------: | :----------------------------------------------------------: | :----------------------------------------------------------: |
| YOLOv3-MobileNetV1 | baseline | 608 | 24.65 | 94.6 | 28.8 | [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%) | 32.0(-66.0%) | 28.7(-0.1) | [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/extensions/yolov3_mobilenet_v1_coco_distill_prune.yml) |
configs/slim/distill/README.md 0 → 100644
浏览文件 @ 91609c99
# Distillation(蒸馏)
## 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,objness`等Tensor，用teacher得到的结果指导student训练。具体实现可参考[代码](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/ppdet/slim/distill.py)
## Citations
```
@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}
}
```
configs/slim/distill/yolov3_mobilenet_v1_coco_distill.yml 0 → 100644
浏览文件 @ 91609c99
_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
configs/slim/extensions/yolov3_mobilenet_v1_coco_distill_prune.yml 0 → 100644
浏览文件 @ 91609c99
_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]
configs/slim/prune/yolov3_prune_fpgm.yml
浏览文件 @ 91609c99
...@@ -5,11 +5,11 @@ slim: Pruner ...@@ -5,11 +5,11 @@ slim: Pruner
Pruner: Pruner:
criterion: fpgm criterion: fpgm
pruned_params: ['yolo_block.0.0.0.conv.weights', 'yolo_block.0.0.1.conv.weights', 'yolo_block.0.1.0.conv.weights', pruned_params: ['conv2d_27.w_0', 'conv2d_28.w_0', 'conv2d_29.w_0',
'yolo_block.0.1.1.conv.weights', 'yolo_block.0.2.conv.weights', 'yolo_block.0.tip.conv.weights', 'conv2d_30.w_0', 'conv2d_31.w_0', 'conv2d_32.w_0',
'yolo_block.1.0.0.conv.weights', 'yolo_block.1.0.1.conv.weights', 'yolo_block.1.1.0.conv.weights', 'conv2d_34.w_0', 'conv2d_35.w_0', 'conv2d_36.w_0',
'yolo_block.1.1.1.conv.weights', 'yolo_block.1.2.conv.weights', 'yolo_block.1.tip.conv.weights', 'conv2d_37.w_0', 'conv2d_38.w_0', 'conv2d_39.w_0',
'yolo_block.2.0.0.conv.weights', 'yolo_block.2.0.1.conv.weights', 'yolo_block.2.1.0.conv.weights', 'conv2d_41.w_0', 'conv2d_42.w_0', 'conv2d_43.w_0',
'yolo_block.2.1.1.conv.weights', 'yolo_block.2.2.conv.weights', 'yolo_block.2.tip.conv.weights'] '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] 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 print_params: False
configs/slim/prune/yolov3_prune_l1_norm.yml
浏览文件 @ 91609c99
...@@ -5,11 +5,11 @@ slim: Pruner ...@@ -5,11 +5,11 @@ slim: Pruner
Pruner: Pruner:
criterion: l1_norm criterion: l1_norm
pruned_params: ['yolo_block.0.0.0.conv.weights', 'yolo_block.0.0.1.conv.weights', 'yolo_block.0.1.0.conv.weights', pruned_params: ['conv2d_27.w_0', 'conv2d_28.w_0', 'conv2d_29.w_0',
'yolo_block.0.1.1.conv.weights', 'yolo_block.0.2.conv.weights', 'yolo_block.0.tip.conv.weights', 'conv2d_30.w_0', 'conv2d_31.w_0', 'conv2d_32.w_0',
'yolo_block.1.0.0.conv.weights', 'yolo_block.1.0.1.conv.weights', 'yolo_block.1.1.0.conv.weights', 'conv2d_34.w_0', 'conv2d_35.w_0', 'conv2d_36.w_0',
'yolo_block.1.1.1.conv.weights', 'yolo_block.1.2.conv.weights', 'yolo_block.1.tip.conv.weights', 'conv2d_37.w_0', 'conv2d_38.w_0', 'conv2d_39.w_0',
'yolo_block.2.0.0.conv.weights', 'yolo_block.2.0.1.conv.weights', 'yolo_block.2.1.0.conv.weights', 'conv2d_41.w_0', 'conv2d_42.w_0', 'conv2d_43.w_0',
'yolo_block.2.1.1.conv.weights', 'yolo_block.2.2.conv.weights', 'yolo_block.2.tip.conv.weights'] '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] 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 print_params: False
configs/yolov3/README.md
浏览文件 @ 91609c99
...@@ -9,27 +9,30 @@ ...@@ -9,27 +9,30 @@
| DarkNet53(paper) | 608 | 8 | 270e | ---- | 33.0 | - | - | | DarkNet53(paper) | 608 | 8 | 270e | ---- | 33.0 | - | - |
| DarkNet53(paper) | 416 | 8 | 270e | ---- | 31.0 | - | - | | DarkNet53(paper) | 416 | 8 | 270e | ---- | 31.0 | - | - |
| DarkNet53(paper) | 320 | 8 | 270e | ---- | 28.2 | - | - | | DarkNet53(paper) | 320 | 8 | 270e | ---- | 28.2 | - | - |
| DarkNet53 | 608 | 8 | 270e | ---- | 39.0 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_darknet53_270e_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/master/dygraph/configs/yolov3/yolov3_darknet53_270e_coco.yml) | | DarkNet53 | 608 | 8 | 270e | ---- | 39.0 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_darknet53_270e_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_darknet53_270e_coco.yml) |
| DarkNet53 | 416 | 8 | 270e | ---- | 37.5 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_darknet53_270e_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/master/dygraph/configs/yolov3/yolov3_darknet53_270e_coco.yml) | | DarkNet53 | 416 | 8 | 270e | ---- | 37.5 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_darknet53_270e_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_darknet53_270e_coco.yml) |
| DarkNet53 | 320 | 8 | 270e | ---- | 34.6 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_darknet53_270e_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/master/dygraph/configs/yolov3/yolov3_darknet53_270e_coco.yml) | | DarkNet53 | 320 | 8 | 270e | ---- | 34.6 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_darknet53_270e_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_darknet53_270e_coco.yml) |
| ResNet50_vd | 608 | 8 | 270e | ---- | 39.1 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_r50vd_dcn_270e_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/master/dygraph/configs/yolov3/yolov3_r50vd_dcn_270e_coco.yml) | | ResNet50_vd | 608 | 8 | 270e | ---- | 39.1 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_r50vd_dcn_270e_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_r50vd_dcn_270e_coco.yml) |
| MobileNet-V1 | 608 | 8 | 270e | ---- | 28.8 | [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/master/dygraph/configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml) | | ResNet34 | 608 | 8 | 270e | ---- | 36.2 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_r34_270e_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_r34_270e_coco.yml) |
| MobileNet-V1 | 416 | 8 | 270e | ---- | 28.7 | [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/master/dygraph/configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml) | | ResNet34 | 416 | 8 | 270e | ---- | 34.3 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_r34_270e_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_r34_270e_coco.yml) |
| MobileNet-V1 | 320 | 8 | 270e | ---- | 26.5 | [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/master/dygraph/configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml) | | ResNet34 | 320 | 8 | 270e | ---- | 31.2 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_r34_270e_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_r34_270e_coco.yml) |
| MobileNet-V3 | 608 | 8 | 270e | ---- | 31.4 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v3_large_270e_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/master/dygraph/configs/yolov3/yolov3_mobilenet_v3_large_270e_coco.yml) | | MobileNet-V1 | 608 | 8 | 270e | ---- | 28.8 | [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) |
| MobileNet-V3 | 416 | 8 | 270e | ---- | 29.7 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v3_large_270e_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/master/dygraph/configs/yolov3/yolov3_mobilenet_v3_large_270e_coco.yml) | | MobileNet-V1 | 416 | 8 | 270e | ---- | 28.7 | [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) |
| MobileNet-V3 | 320 | 8 | 270e | ---- | 26.9 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v3_large_270e_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/master/dygraph/configs/yolov3/yolov3_mobilenet_v3_large_270e_coco.yml) | | MobileNet-V1 | 320 | 8 | 270e | ---- | 26.5 | [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) |
| MobileNet-V3 | 608 | 8 | 270e | ---- | 31.4 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v3_large_270e_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v3_large_270e_coco.yml) |
| MobileNet-V3 | 416 | 8 | 270e | ---- | 29.7 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v3_large_270e_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v3_large_270e_coco.yml) |
| MobileNet-V3 | 320 | 8 | 270e | ---- | 26.9 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v3_large_270e_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v3_large_270e_coco.yml) |
### YOLOv3 on Pasacl VOC ### YOLOv3 on Pasacl VOC
| 骨架网络 | 输入尺寸 | 每张GPU图片个数 | 学习率策略 |推理时间(fps)| Box AP | 下载 | 配置文件 | | 骨架网络 | 输入尺寸 | 每张GPU图片个数 | 学习率策略 |推理时间(fps)| Box AP | 下载 | 配置文件 |
| :----------- | :--: | :-----: | :-----: |:------------: |:----: | :-------: | :----: | | :----------- | :--: | :-----: | :-----: |:------------: |:----: | :-------: | :----: |
| MobileNet-V1 | 608 | 8 | 270e | - | 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/master/dygraph/configs/yolov3/yolov3_mobilenet_v1_270e_voc.yml) | | MobileNet-V1 | 608 | 8 | 270e | - | 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) |
| MobileNet-V1 | 416 | 8 | 270e | - | 76.1 | [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/master/dygraph/configs/yolov3/yolov3_mobilenet_v1_270e_voc.yml) | | MobileNet-V1 | 416 | 8 | 270e | - | 76.1 | [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) |
| MobileNet-V1 | 320 | 8 | 270e | - | 73.6 | [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/master/dygraph/configs/yolov3/yolov3_mobilenet_v1_270e_voc.yml) | | MobileNet-V1 | 320 | 8 | 270e | - | 73.6 | [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) |
| MobileNet-V3 | 608 | 8 | 270e | - | 79.6 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v3_large_270e_voc.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/master/dygraph/configs/yolov3/yolov3_mobilenet_v3_large_270e_voc.yml) | | MobileNet-V3 | 608 | 8 | 270e | - | 79.6 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v3_large_270e_voc.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v3_large_270e_voc.yml) |
| MobileNet-V3 | 416 | 8 | 270e | - | 78.6 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v3_large_270e_voc.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/master/dygraph/configs/yolov3/yolov3_mobilenet_v3_large_270e_voc.yml) | | MobileNet-V3 | 416 | 8 | 270e | - | 78.6 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v3_large_270e_voc.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v3_large_270e_voc.yml) |
| MobileNet-V3 | 320 | 8 | 270e | - | 76.4 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v3_large_270e_voc.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/master/dygraph/configs/yolov3/yolov3_mobilenet_v3_large_270e_voc.yml) | | MobileNet-V3 | 320 | 8 | 270e | - | 76.4 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v3_large_270e_voc.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v3_large_270e_voc.yml) |
**注意：** YOLOv3均使用8GPU训练，训练270个epoch **注意：** YOLOv3均使用8GPU训练，训练270个epoch
......
configs/yolov3/_base_/yolov3_r34.yml 0 → 100644
浏览文件 @ 91609c99
architecture: YOLOv3
pretrain_weights: https://paddledet.bj.bcebos.com/models/pretrained/ResNet34_pretrained.pdparams
norm_type: sync_bn
YOLOv3:
backbone: ResNet
neck: YOLOv3FPN
yolo_head: YOLOv3Head
post_process: BBoxPostProcess
ResNet:
depth: 34
return_idx: [1, 2, 3]
freeze_at: -1
freeze_norm: false
norm_decay: 0.
YOLOv3Head:
anchors: [[10, 13], [16, 30], [33, 23],
[30, 61], [62, 45], [59, 119],
[116, 90], [156, 198], [373, 326]]
anchor_masks: [[6, 7, 8], [3, 4, 5], [0, 1, 2]]
loss: YOLOv3Loss
YOLOv3Loss:
ignore_thresh: 0.7
downsample: [32, 16, 8]
label_smooth: false
BBoxPostProcess:
decode:
name: YOLOBox
conf_thresh: 0.005
downsample_ratio: 32
clip_bbox: true
nms:
name: MultiClassNMS
keep_top_k: 100
score_threshold: 0.01
nms_threshold: 0.45
nms_top_k: 1000
configs/yolov3/yolov3_r34_270e_coco.yml 0 → 100644
浏览文件 @ 91609c99
_BASE_: [
'../datasets/coco_detection.yml',
'../runtime.yml',
'_base_/optimizer_270e.yml',
'_base_/yolov3_r34.yml',
'_base_/yolov3_reader.yml',
]
snapshot_epoch: 5
weights: output/yolov3_r34_270e_coco/model_final
ppdet/engine/callbacks.py
浏览文件 @ 91609c99
...@@ -146,13 +146,17 @@ class Checkpointer(Callback): ...@@ -146,13 +146,17 @@ class Checkpointer(Callback):
self.use_ema = ('use_ema' in cfg and cfg['use_ema']) self.use_ema = ('use_ema' in cfg and cfg['use_ema'])
self.save_dir = os.path.join(self.model.cfg.save_dir, self.save_dir = os.path.join(self.model.cfg.save_dir,
self.model.cfg.filename) self.model.cfg.filename)
if hasattr(self.model.model, 'student_model'):
self.weight = self.model.model.student_model
else:
self.weight = self.model.model
if self.use_ema: if self.use_ema:
self.ema = ModelEMA( self.ema = ModelEMA(
cfg['ema_decay'], self.model.model, use_thres_step=True) cfg['ema_decay'], self.weight, use_thres_step=True)
def on_step_end(self, status): def on_step_end(self, status):
if self.use_ema: if self.use_ema:
self.ema.update(self.model.model) self.ema.update(self.weight)
def on_epoch_end(self, status): def on_epoch_end(self, status):
# Checkpointer only performed during training # Checkpointer only performed during training
...@@ -169,7 +173,7 @@ class Checkpointer(Callback): ...@@ -169,7 +173,7 @@ class Checkpointer(Callback):
if self.use_ema: if self.use_ema:
weight = self.ema.apply() weight = self.ema.apply()
else: else:
weight = self.model.model weight = self.weight
elif mode == 'eval': elif mode == 'eval':
if 'save_best_model' in status and status['save_best_model']: if 'save_best_model' in status and status['save_best_model']:
for metric in self.model._metrics: for metric in self.model._metrics:
...@@ -181,7 +185,7 @@ class Checkpointer(Callback): ...@@ -181,7 +185,7 @@ class Checkpointer(Callback):
if self.use_ema: if self.use_ema:
weight = self.ema.apply() weight = self.ema.apply()
else: else:
weight = self.model.model weight = self.weight
logger.info("Best test {} ap is {:0.3f}.".format( logger.info("Best test {} ap is {:0.3f}.".format(
key, self.best_ap)) key, self.best_ap))
if weight: if weight:
......
ppdet/engine/trainer.py
浏览文件 @ 91609c99
...@@ -52,17 +52,14 @@ class Trainer(object): ...@@ -52,17 +52,14 @@ class Trainer(object):
"mode should be 'train', 'eval' or 'test'" "mode should be 'train', 'eval' or 'test'"
self.mode = mode.lower() self.mode = mode.lower()
self.optimizer = None self.optimizer = None
self.slim = None self.is_loaded_weights = False
# build model # build model
if 'model' not in self.cfg:
self.model = create(cfg.architecture) self.model = create(cfg.architecture)
else:
# model slim build self.model = self.cfg.model
if 'slim' in cfg and cfg.slim: self.is_loaded_weights = True
if self.mode == 'train':
self.load_weights(cfg.pretrain_weights)
self.slim = create(cfg.slim)
self.slim(self.model)
# build data loader # build data loader
self.dataset = cfg['{}Dataset'.format(self.mode.capitalize())] self.dataset = cfg['{}Dataset'.format(self.mode.capitalize())]
...@@ -192,11 +189,18 @@ class Trainer(object): ...@@ -192,11 +189,18 @@ class Trainer(object):
self._metrics.extend(metrics) self._metrics.extend(metrics)
def load_weights(self, weights): def load_weights(self, weights):
if self.is_loaded_weights:
return
self.start_epoch = 0 self.start_epoch = 0
load_pretrain_weight(self.model, weights) load_pretrain_weight(self.model, weights)
logger.debug("Load weights {} to start training".format(weights)) logger.debug("Load weights {} to start training".format(weights))
def resume_weights(self, weights): def resume_weights(self, weights):
# support Distill resume weights
if hasattr(self.model, 'student_model'):
self.start_epoch = load_weight(self.model.student_model, weights,
self.optimizer)
else:
self.start_epoch = load_weight(self.model, weights, self.optimizer) self.start_epoch = load_weight(self.model, weights, self.optimizer)
logger.debug("Resume weights of epoch {}".format(self.start_epoch)) logger.debug("Resume weights of epoch {}".format(self.start_epoch))
...@@ -419,7 +423,7 @@ class Trainer(object): ...@@ -419,7 +423,7 @@ class Trainer(object):
}] }]
# dy2st and save model # dy2st and save model
if self.slim is None or self.cfg['slim'] != 'QAT': if 'slim' not in self.cfg or self.cfg['slim'] != 'QAT':
static_model = paddle.jit.to_static( static_model = paddle.jit.to_static(
self.model, input_spec=input_spec) self.model, input_spec=input_spec)
# NOTE: dy2st do not pruned program, but jit.save will prune program # NOTE: dy2st do not pruned program, but jit.save will prune program
...@@ -433,7 +437,7 @@ class Trainer(object): ...@@ -433,7 +437,7 @@ class Trainer(object):
input_spec=pruned_input_spec) input_spec=pruned_input_spec)
logger.info("Export model and saved in {}".format(save_dir)) logger.info("Export model and saved in {}".format(save_dir))
else: else:
self.slim.save_quantized_model( self.cfg.slim.save_quantized_model(
self.model, self.model,
os.path.join(save_dir, 'model'), os.path.join(save_dir, 'model'),
input_spec=input_spec) input_spec=input_spec)
......
ppdet/modeling/backbones/darknet.py
浏览文件 @ 91609c99
...@@ -46,7 +46,6 @@ class ConvBNLayer(nn.Layer): ...@@ -46,7 +46,6 @@ class ConvBNLayer(nn.Layer):
stride=stride, stride=stride,
padding=padding, padding=padding,
groups=groups, groups=groups,
weight_attr=ParamAttr(name=name + '.conv.weights'),
data_format=data_format, data_format=data_format,
bias_attr=False) bias_attr=False)
self.batch_norm = batch_norm( self.batch_norm = batch_norm(
......
ppdet/modeling/backbones/mobilenet_v1.py
浏览文件 @ 91609c99
...@@ -55,14 +55,11 @@ class ConvBNLayer(nn.Layer): ...@@ -55,14 +55,11 @@ class ConvBNLayer(nn.Layer):
weight_attr=ParamAttr( weight_attr=ParamAttr(
learning_rate=conv_lr, learning_rate=conv_lr,
initializer=KaimingNormal(), initializer=KaimingNormal(),
regularizer=L2Decay(conv_decay), regularizer=L2Decay(conv_decay)),
name=name + "_weights"),
bias_attr=False) bias_attr=False)
param_attr = ParamAttr( param_attr = ParamAttr(regularizer=L2Decay(norm_decay))
name=name + "_bn_scale", regularizer=L2Decay(norm_decay)) bias_attr = ParamAttr(regularizer=L2Decay(norm_decay))
bias_attr = ParamAttr(
name=name + "_bn_offset", regularizer=L2Decay(norm_decay))
if norm_type == 'sync_bn': if norm_type == 'sync_bn':
self._batch_norm = nn.SyncBatchNorm( self._batch_norm = nn.SyncBatchNorm(
out_channels, weight_attr=param_attr, bias_attr=bias_attr) out_channels, weight_attr=param_attr, bias_attr=bias_attr)
...@@ -72,9 +69,7 @@ class ConvBNLayer(nn.Layer): ...@@ -72,9 +69,7 @@ class ConvBNLayer(nn.Layer):
act=None, act=None,
param_attr=param_attr, param_attr=param_attr,
bias_attr=bias_attr, bias_attr=bias_attr,
use_global_stats=False, use_global_stats=False)
moving_mean_name=name + '_bn_mean',
moving_variance_name=name + '_bn_variance')
def forward(self, x): def forward(self, x):
x = self._conv(x) x = self._conv(x)
......
ppdet/modeling/backbones/resnet.py
浏览文件 @ 91609c99
...@@ -64,7 +64,7 @@ class ConvNormLayer(nn.Layer): ...@@ -64,7 +64,7 @@ class ConvNormLayer(nn.Layer):
padding=(filter_size - 1) // 2, padding=(filter_size - 1) // 2,
groups=groups, groups=groups,
weight_attr=paddle.ParamAttr( weight_attr=paddle.ParamAttr(
learning_rate=lr, name=name + "_weights"), learning_rate=lr, ),
bias_attr=False) bias_attr=False)
else: else:
self.conv = DeformableConvV2( self.conv = DeformableConvV2(
...@@ -75,7 +75,7 @@ class ConvNormLayer(nn.Layer): ...@@ -75,7 +75,7 @@ class ConvNormLayer(nn.Layer):
padding=(filter_size - 1) // 2, padding=(filter_size - 1) // 2,
groups=groups, groups=groups,
weight_attr=paddle.ParamAttr( weight_attr=paddle.ParamAttr(
learning_rate=lr, name=name + '_weights'), learning_rate=lr, ),
bias_attr=False, bias_attr=False,
name=name) name=name)
...@@ -84,12 +84,10 @@ class ConvNormLayer(nn.Layer): ...@@ -84,12 +84,10 @@ class ConvNormLayer(nn.Layer):
param_attr = paddle.ParamAttr( param_attr = paddle.ParamAttr(
learning_rate=norm_lr, learning_rate=norm_lr,
regularizer=L2Decay(norm_decay), regularizer=L2Decay(norm_decay),
name=bn_name + "_scale",
trainable=False if freeze_norm else True) trainable=False if freeze_norm else True)
bias_attr = paddle.ParamAttr( bias_attr = paddle.ParamAttr(
learning_rate=norm_lr, learning_rate=norm_lr,
regularizer=L2Decay(norm_decay), regularizer=L2Decay(norm_decay),
name=bn_name + "_offset",
trainable=False if freeze_norm else True) trainable=False if freeze_norm else True)
global_stats = True if freeze_norm else False global_stats = True if freeze_norm else False
...@@ -102,9 +100,7 @@ class ConvNormLayer(nn.Layer): ...@@ -102,9 +100,7 @@ class ConvNormLayer(nn.Layer):
act=None, act=None,
param_attr=param_attr, param_attr=param_attr,
bias_attr=bias_attr, bias_attr=bias_attr,
use_global_stats=global_stats, use_global_stats=global_stats)
moving_mean_name=bn_name + '_mean',
moving_variance_name=bn_name + '_variance')
norm_params = self.norm.parameters() norm_params = self.norm.parameters()
if freeze_norm: if freeze_norm:
......
ppdet/modeling/heads/yolo_head.py
浏览文件 @ 91609c99
...@@ -56,9 +56,7 @@ class YOLOv3Head(nn.Layer): ...@@ -56,9 +56,7 @@ class YOLOv3Head(nn.Layer):
stride=1, stride=1,
padding=0, padding=0,
data_format=data_format, data_format=data_format,
weight_attr=ParamAttr(name=name + '.conv.weights'), bias_attr=ParamAttr(regularizer=L2Decay(0.))))
bias_attr=ParamAttr(
name=name + '.conv.bias', regularizer=L2Decay(0.))))
self.yolo_outputs.append(yolo_output) self.yolo_outputs.append(yolo_output)
def parse_anchor(self, anchors, anchor_masks): def parse_anchor(self, anchors, anchor_masks):
......
ppdet/modeling/losses/fcos_loss.py
浏览文件 @ 91609c99
...@@ -20,6 +20,7 @@ import paddle ...@@ -20,6 +20,7 @@ import paddle
import paddle.nn as nn import paddle.nn as nn
import paddle.nn.functional as F import paddle.nn.functional as F
from ppdet.core.workspace import register from ppdet.core.workspace import register
from ppdet.modeling import ops
INF = 1e8 INF = 1e8
__all__ = ['FCOSLoss'] __all__ = ['FCOSLoss']
...@@ -44,19 +45,6 @@ def flatten_tensor(inputs, channel_first=False): ...@@ -44,19 +45,6 @@ def flatten_tensor(inputs, channel_first=False):
return output_channel_last return output_channel_last
def sigmoid_cross_entropy_with_logits_loss(inputs,
label,
ignore_index=-100,
normalize=False):
output = F.binary_cross_entropy_with_logits(inputs, label, reduction='none')
mask_tensor = paddle.cast(label != ignore_index, 'float32')
output = paddle.multiply(output, mask_tensor)
if normalize:
sum_valid_mask = paddle.sum(mask_tensor)
output = output / sum_valid_mask
return output
@register @register
class FCOSLoss(nn.Layer): class FCOSLoss(nn.Layer):
""" """
...@@ -226,7 +214,7 @@ class FCOSLoss(nn.Layer): ...@@ -226,7 +214,7 @@ class FCOSLoss(nn.Layer):
# 3. centerness: sigmoid_cross_entropy_with_logits_loss # 3. centerness: sigmoid_cross_entropy_with_logits_loss
centerness_flatten = paddle.squeeze(centerness_flatten, axis=-1) centerness_flatten = paddle.squeeze(centerness_flatten, axis=-1)
ctn_loss = sigmoid_cross_entropy_with_logits_loss(centerness_flatten, ctn_loss = ops.sigmoid_cross_entropy_with_logits(centerness_flatten,
tag_center_flatten) tag_center_flatten)
ctn_loss = ctn_loss * mask_positive_float / num_positive_fp32 ctn_loss = ctn_loss * mask_positive_float / num_positive_fp32
......
ppdet/modeling/losses/yolo_loss.py
浏览文件 @ 91609c99
...@@ -54,6 +54,7 @@ class YOLOv3Loss(nn.Layer): ...@@ -54,6 +54,7 @@ class YOLOv3Loss(nn.Layer):
self.scale_x_y = scale_x_y self.scale_x_y = scale_x_y
self.iou_loss = iou_loss self.iou_loss = iou_loss
self.iou_aware_loss = iou_aware_loss self.iou_aware_loss = iou_aware_loss
self.distill_pairs = []
def obj_loss(self, pbox, gbox, pobj, tobj, anchor, downsample): def obj_loss(self, pbox, gbox, pobj, tobj, anchor, downsample):
# pbox # pbox
...@@ -108,6 +109,7 @@ class YOLOv3Loss(nn.Layer): ...@@ -108,6 +109,7 @@ class YOLOv3Loss(nn.Layer):
x, y = p[:, :, :, :, 0:1], p[:, :, :, :, 1:2] x, y = p[:, :, :, :, 0:1], p[:, :, :, :, 1:2]
w, h = p[:, :, :, :, 2:3], p[:, :, :, :, 3:4] w, h = p[:, :, :, :, 2:3], p[:, :, :, :, 3:4]
obj, pcls = p[:, :, :, :, 4:5], p[:, :, :, :, 5:] obj, pcls = p[:, :, :, :, 4:5], p[:, :, :, :, 5:]
self.distill_pairs.append([x, y, w, h, obj, pcls])
t = t.transpose((0, 1, 3, 4, 2)) t = t.transpose((0, 1, 3, 4, 2))
tx, ty = t[:, :, :, :, 0:1], t[:, :, :, :, 1:2] tx, ty = t[:, :, :, :, 0:1], t[:, :, :, :, 1:2]
...@@ -173,6 +175,7 @@ class YOLOv3Loss(nn.Layer): ...@@ -173,6 +175,7 @@ class YOLOv3Loss(nn.Layer):
gt_targets = [targets['target{}'.format(i)] for i in range(np)] gt_targets = [targets['target{}'.format(i)] for i in range(np)]
gt_box = targets['gt_bbox'] gt_box = targets['gt_bbox']
yolo_losses = dict() yolo_losses = dict()
self.distill_pairs.clear()
for x, t, anchor, downsample in zip(inputs, gt_targets, anchors, for x, t, anchor, downsample in zip(inputs, gt_targets, anchors,
self.downsample): self.downsample):
yolo_loss = self.yolov3_loss(x, t, gt_box, anchor, downsample, yolo_loss = self.yolov3_loss(x, t, gt_box, anchor, downsample,
......
ppdet/modeling/ops.py
浏览文件 @ 91609c99
...@@ -59,11 +59,8 @@ def batch_norm(ch, ...@@ -59,11 +59,8 @@ def batch_norm(ch,
return batch_norm( return batch_norm(
ch, ch,
weight_attr=ParamAttr( weight_attr=ParamAttr(
name=bn_name + '.scale', initializer=initializer, regularizer=L2Decay(norm_decay)),
initializer=initializer, bias_attr=ParamAttr(regularizer=L2Decay(norm_decay)),
regularizer=L2Decay(norm_decay)),
bias_attr=ParamAttr(
name=bn_name + '.offset', regularizer=L2Decay(norm_decay)),
data_format=data_format) data_format=data_format)
...@@ -1558,7 +1555,6 @@ def sigmoid_cross_entropy_with_logits(input, ...@@ -1558,7 +1555,6 @@ def sigmoid_cross_entropy_with_logits(input,
output = F.binary_cross_entropy_with_logits(input, label, reduction='none') output = F.binary_cross_entropy_with_logits(input, label, reduction='none')
mask_tensor = paddle.cast(label != ignore_index, 'float32') mask_tensor = paddle.cast(label != ignore_index, 'float32')
output = paddle.multiply(output, mask_tensor) output = paddle.multiply(output, mask_tensor)
output = paddle.reshape(output, shape=[output.shape[0], -1])
if normalize: if normalize:
sum_valid_mask = paddle.sum(mask_tensor) sum_valid_mask = paddle.sum(mask_tensor)
output = output / sum_valid_mask output = output / sum_valid_mask
......
ppdet/slim/__init__.py
浏览文件 @ 91609c99
...@@ -14,6 +14,47 @@ ...@@ -14,6 +14,47 @@
from . import prune from . import prune
from . import quant from . import quant
from . import distill
from .prune import * from .prune import *
from .quant import * from .quant import *
from .distill import *
import yaml
from ppdet.core.workspace import load_config
from ppdet.utils.checkpoint import load_pretrain_weight
def build_slim_model(cfg, slim_cfg, mode='train'):
with open(slim_cfg) as f:
slim_load_cfg = yaml.load(f, Loader=yaml.Loader)
if mode != 'train' and slim_load_cfg['slim'] == 'Distill':
return cfg
if slim_load_cfg['slim'] == 'Distill':
model = DistillModel(cfg, slim_cfg)
cfg['model'] = model
elif slim_load_cfg['slim'] == 'DistillPrune':
if mode == 'train':
model = DistillModel(cfg, slim_cfg)
pruner = create(cfg.pruner)
pruner(model.student_model)
else:
model = create(cfg.architecture)
weights = cfg.weights
load_config(slim_cfg)
pruner = create(cfg.pruner)
model = pruner(model)
load_pretrain_weight(model, weights)
cfg['model'] = model
else:
load_config(slim_cfg)
model = create(cfg.architecture)
if mode == 'train':
load_pretrain_weight(model, cfg.pretrain_weights)
slim = create(cfg.slim)
cfg['model'] = slim(model)
if mode != 'train':
load_pretrain_weight(cfg['model'], cfg.weights)
return cfg
ppdet/slim/distill.py 0 → 100644
浏览文件 @ 91609c99
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import paddle
import paddle.nn as nn
import paddle.nn.functional as F
from ppdet.core.workspace import register, serializable, load_config
from ppdet.core.workspace import create
from ppdet.utils.logger import setup_logger
from ppdet.modeling import ops
from ppdet.utils.checkpoint import load_pretrain_weight
from ppdet.modeling.losses import YOLOv3Loss
logger = setup_logger(__name__)
class DistillModel(nn.Layer):
def __init__(self, cfg, slim_cfg):
super(DistillModel, self).__init__()
self.student_model = create(cfg.architecture)
logger.debug('Load student model pretrain_weights:{}'.format(
cfg.pretrain_weights))
load_pretrain_weight(self.student_model, cfg.pretrain_weights)
slim_cfg = load_config(slim_cfg)
self.teacher_model = create(slim_cfg.architecture)
self.distill_loss = create(slim_cfg.distill_loss)
logger.debug('Load teacher model pretrain_weights:{}'.format(
slim_cfg.pretrain_weights))
load_pretrain_weight(self.teacher_model, slim_cfg.pretrain_weights)
for param in self.teacher_model.parameters():
param.trainable = False
def parameters(self):
return self.student_model.parameters()
def forward(self, inputs):
if self.training:
teacher_loss = self.teacher_model(inputs)
student_loss = self.student_model(inputs)
loss = self.distill_loss(self.teacher_model, self.student_model)
student_loss['distill_loss'] = loss
student_loss['teacher_loss'] = teacher_loss['loss']
student_loss['loss'] += student_loss['distill_loss']
return student_loss
else:
return self.student_model(inputs)
@register
class DistillYOLOv3Loss(nn.Layer):
def __init__(self, weight=1000):
super(DistillYOLOv3Loss, self).__init__()
self.weight = weight
def obj_weighted_reg(self, sx, sy, sw, sh, tx, ty, tw, th, tobj):
loss_x = ops.sigmoid_cross_entropy_with_logits(sx, F.sigmoid(tx))
loss_y = ops.sigmoid_cross_entropy_with_logits(sy, F.sigmoid(ty))
loss_w = paddle.abs(sw - tw)
loss_h = paddle.abs(sh - th)
loss = paddle.add_n([loss_x, loss_y, loss_w, loss_h])
weighted_loss = paddle.mean(loss * F.sigmoid(tobj))
return weighted_loss
def obj_weighted_cls(self, scls, tcls, tobj):
loss = ops.sigmoid_cross_entropy_with_logits(scls, F.sigmoid(tcls))
weighted_loss = paddle.mean(paddle.multiply(loss, F.sigmoid(tobj)))
return weighted_loss
def obj_loss(self, sobj, tobj):
obj_mask = paddle.cast(tobj > 0., dtype="float32")
obj_mask.stop_gradient = True
loss = paddle.mean(
ops.sigmoid_cross_entropy_with_logits(sobj, obj_mask))
return loss
def forward(self, teacher_model, student_model):
teacher_distill_pairs = teacher_model.yolo_head.loss.distill_pairs
student_distill_pairs = student_model.yolo_head.loss.distill_pairs
distill_reg_loss, distill_cls_loss, distill_obj_loss = [], [], []
for s_pair, t_pair in zip(student_distill_pairs, teacher_distill_pairs):
distill_reg_loss.append(
self.obj_weighted_reg(s_pair[0], s_pair[1], s_pair[2], s_pair[
3], t_pair[0], t_pair[1], t_pair[2], t_pair[3], t_pair[4]))
distill_cls_loss.append(
self.obj_weighted_cls(s_pair[5], t_pair[5], t_pair[4]))
distill_obj_loss.append(self.obj_loss(s_pair[4], t_pair[4]))
distill_reg_loss = paddle.add_n(distill_reg_loss)
distill_cls_loss = paddle.add_n(distill_cls_loss)
distill_obj_loss = paddle.add_n(distill_obj_loss)
loss = (distill_reg_loss + distill_cls_loss + distill_obj_loss
) * self.weight
return loss
ppdet/utils/checkpoint.py
浏览文件 @ 91609c99
...@@ -178,6 +178,38 @@ def load_pretrain_weight(model, pretrain_weight): ...@@ -178,6 +178,38 @@ def load_pretrain_weight(model, pretrain_weight):
logger.info('Finish loading model weights: {}'.format(weights_path)) logger.info('Finish loading model weights: {}'.format(weights_path))
def load_static_pretrain_weight(model, pretrain_weight):
if is_url(pretrain_weight):
pretrain_weight = get_weights_path_dist(pretrain_weight)
path = _strip_postfix(pretrain_weight)
if not (os.path.isdir(path) or os.path.isfile(path) or
os.path.exists(path + '.pdparams')):
raise ValueError("Model pretrain path `{}` does not exists. "
"If you don't want to load pretrain model, "
"please delete `pretrain_weights` field in "
"config file.".format(path))
model_dict = model.state_dict()
pre_state_dict = paddle.static.load_program_state(path)
param_state_dict = {}
for key in model_dict.keys():
weight_name = model_dict[key].name
if weight_name in pre_state_dict.keys():
logger.info('Load weight: {}, shape: {}'.format(
weight_name, pre_state_dict[weight_name].shape))
param_state_dict[key] = pre_state_dict[weight_name]
else:
if 'backbone' in key:
logger.info('Lack weight: {}, structure name: {}'.format(
weight_name, key))
param_state_dict[key] = model_dict[key]
model.set_dict(param_state_dict)
return
def save_model(model, optimizer, save_dir, save_name, last_epoch): def save_model(model, optimizer, save_dir, save_name, last_epoch):
""" """
save model into disk. save model into disk.
......
tools/eval.py
浏览文件 @ 91609c99
...@@ -33,6 +33,7 @@ from ppdet.utils.check import check_gpu, check_version, check_config ...@@ -33,6 +33,7 @@ from ppdet.utils.check import check_gpu, check_version, check_config
from ppdet.utils.cli import ArgsParser from ppdet.utils.cli import ArgsParser
from ppdet.engine import Trainer, init_parallel_env from ppdet.engine import Trainer, init_parallel_env
from ppdet.metrics.coco_utils import json_eval_results from ppdet.metrics.coco_utils import json_eval_results
from ppdet.slim import build_slim_model
from ppdet.utils.logger import setup_logger from ppdet.utils.logger import setup_logger
logger = setup_logger('eval') logger = setup_logger('eval')
...@@ -100,22 +101,22 @@ def run(FLAGS, cfg): ...@@ -100,22 +101,22 @@ def run(FLAGS, cfg):
def main(): def main():
FLAGS = parse_args() FLAGS = parse_args()
cfg = load_config(FLAGS.config) cfg = load_config(FLAGS.config)
# TODO: bias should be unified # TODO: bias should be unified
cfg['bias'] = 1 if FLAGS.bias else 0 cfg['bias'] = 1 if FLAGS.bias else 0
cfg['classwise'] = True if FLAGS.classwise else False cfg['classwise'] = True if FLAGS.classwise else False
cfg['output_eval'] = FLAGS.output_eval cfg['output_eval'] = FLAGS.output_eval
merge_config(FLAGS.opt) merge_config(FLAGS.opt)
place = paddle.set_device('gpu' if cfg.use_gpu else 'cpu')
if FLAGS.slim_config: if FLAGS.slim_config:
slim_cfg = load_config(FLAGS.slim_config) cfg = build_slim_model(cfg, FLAGS.slim_config, mode='eval')
merge_config(slim_cfg)
check_config(cfg) check_config(cfg)
check_gpu(cfg.use_gpu) check_gpu(cfg.use_gpu)
check_version() check_version()
place = paddle.set_device('gpu' if cfg.use_gpu else 'cpu')
run(FLAGS, cfg) run(FLAGS, cfg)
......
tools/export_model.py
浏览文件 @ 91609c99
...@@ -31,6 +31,7 @@ from ppdet.core.workspace import load_config, merge_config ...@@ -31,6 +31,7 @@ from ppdet.core.workspace import load_config, merge_config
from ppdet.utils.check import check_gpu, check_version, check_config from ppdet.utils.check import check_gpu, check_version, check_config
from ppdet.utils.cli import ArgsParser from ppdet.utils.cli import ArgsParser
from ppdet.engine import Trainer from ppdet.engine import Trainer
from ppdet.slim import build_slim_model
from ppdet.utils.logger import setup_logger from ppdet.utils.logger import setup_logger
logger = setup_logger('export_model') logger = setup_logger('export_model')
...@@ -84,15 +85,15 @@ def run(FLAGS, cfg): ...@@ -84,15 +85,15 @@ def run(FLAGS, cfg):
def main(): def main():
paddle.set_device("cpu") paddle.set_device("cpu")
FLAGS = parse_args() FLAGS = parse_args()
cfg = load_config(FLAGS.config) cfg = load_config(FLAGS.config)
# TODO: to be refined in the future # TODO: to be refined in the future
if 'norm_type' in cfg and cfg['norm_type'] == 'sync_bn': if 'norm_type' in cfg and cfg['norm_type'] == 'sync_bn':
FLAGS.opt['norm_type'] = 'bn' FLAGS.opt['norm_type'] = 'bn'
merge_config(FLAGS.opt) merge_config(FLAGS.opt)
if FLAGS.slim_config: if FLAGS.slim_config:
slim_cfg = load_config(FLAGS.slim_config) cfg = build_slim_model(cfg, FLAGS.slim_config, mode='test')
merge_config(slim_cfg)
check_config(cfg) check_config(cfg)
check_gpu(cfg.use_gpu) check_gpu(cfg.use_gpu)
check_version() check_version()
......
tools/infer.py
浏览文件 @ 91609c99
...@@ -31,6 +31,7 @@ from ppdet.core.workspace import load_config, merge_config ...@@ -31,6 +31,7 @@ from ppdet.core.workspace import load_config, merge_config
from ppdet.engine import Trainer from ppdet.engine import Trainer
from ppdet.utils.check import check_gpu, check_version, check_config from ppdet.utils.check import check_gpu, check_version, check_config
from ppdet.utils.cli import ArgsParser from ppdet.utils.cli import ArgsParser
from ppdet.slim import build_slim_model
from ppdet.utils.logger import setup_logger from ppdet.utils.logger import setup_logger
logger = setup_logger('train') logger = setup_logger('train')
...@@ -127,19 +128,20 @@ def run(FLAGS, cfg): ...@@ -127,19 +128,20 @@ def run(FLAGS, cfg):
def main(): def main():
FLAGS = parse_args() FLAGS = parse_args()
cfg = load_config(FLAGS.config) cfg = load_config(FLAGS.config)
cfg['use_vdl'] = FLAGS.use_vdl cfg['use_vdl'] = FLAGS.use_vdl
cfg['vdl_log_dir'] = FLAGS.vdl_log_dir cfg['vdl_log_dir'] = FLAGS.vdl_log_dir
merge_config(FLAGS.opt) merge_config(FLAGS.opt)
place = paddle.set_device('gpu' if cfg.use_gpu else 'cpu')
if FLAGS.slim_config: if FLAGS.slim_config:
slim_cfg = load_config(FLAGS.slim_config) cfg = build_slim_model(cfg, FLAGS.slim_config, mode='test')
merge_config(slim_cfg)
check_config(cfg) check_config(cfg)
check_gpu(cfg.use_gpu) check_gpu(cfg.use_gpu)
check_version() check_version()
place = paddle.set_device('gpu' if cfg.use_gpu else 'cpu')
run(FLAGS, cfg) run(FLAGS, cfg)
......
tools/train.py
浏览文件 @ 91609c99
...@@ -33,6 +33,7 @@ import paddle ...@@ -33,6 +33,7 @@ import paddle
from ppdet.core.workspace import load_config, merge_config, create from ppdet.core.workspace import load_config, merge_config, create
from ppdet.utils.checkpoint import load_weight, load_pretrain_weight from ppdet.utils.checkpoint import load_weight, load_pretrain_weight
from ppdet.engine import Trainer, init_parallel_env, set_random_seed, init_fleet_env from ppdet.engine import Trainer, init_parallel_env, set_random_seed, init_fleet_env
from ppdet.slim import build_slim_model
import ppdet.utils.cli as cli import ppdet.utils.cli as cli
import ppdet.utils.check as check import ppdet.utils.check as check
...@@ -98,7 +99,7 @@ def run(FLAGS, cfg): ...@@ -98,7 +99,7 @@ def run(FLAGS, cfg):
# load weights # load weights
if FLAGS.resume is not None: if FLAGS.resume is not None:
trainer.resume_weights(FLAGS.resume) trainer.resume_weights(FLAGS.resume)
elif not FLAGS.slim_config and 'pretrain_weights' in cfg and cfg.pretrain_weights: elif 'pretrain_weights' in cfg and cfg.pretrain_weights:
trainer.load_weights(cfg.pretrain_weights) trainer.load_weights(cfg.pretrain_weights)
# training # training
...@@ -107,22 +108,22 @@ def run(FLAGS, cfg): ...@@ -107,22 +108,22 @@ def run(FLAGS, cfg):
def main(): def main():
FLAGS = parse_args() FLAGS = parse_args()
cfg = load_config(FLAGS.config) cfg = load_config(FLAGS.config)
cfg['fp16'] = FLAGS.fp16 cfg['fp16'] = FLAGS.fp16
cfg['fleet'] = FLAGS.fleet cfg['fleet'] = FLAGS.fleet
cfg['use_vdl'] = FLAGS.use_vdl cfg['use_vdl'] = FLAGS.use_vdl
cfg['vdl_log_dir'] = FLAGS.vdl_log_dir cfg['vdl_log_dir'] = FLAGS.vdl_log_dir
merge_config(FLAGS.opt) merge_config(FLAGS.opt)
place = paddle.set_device('gpu' if cfg.use_gpu else 'cpu')
if FLAGS.slim_config: if FLAGS.slim_config:
slim_cfg = load_config(FLAGS.slim_config) cfg = build_slim_model(cfg, FLAGS.slim_config)
merge_config(slim_cfg)
check.check_config(cfg) check.check_config(cfg)
check.check_gpu(cfg.use_gpu) check.check_gpu(cfg.use_gpu)
check.check_version() check.check_version()
place = paddle.set_device('gpu' if cfg.use_gpu else 'cpu')
run(FLAGS, cfg) run(FLAGS, cfg)
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或