飞桨PaddlePaddle

飞桨PaddlePaddle

关注
发信
关注(28)粉丝(399)

用Paddle实现高性能目标检测网络：ShuffleNetV2-YOLOv3

项目简介

**本项目旨在设计以YOLOv3为主体框架的高性能目标检测网络.主要思路为：利用YOLOv3的结构，将主干网络换成现阶段比较高效的ShuffleNetv2,进而提升网络的性能，相比较于原版YOLOv3预测，预测速度能提高10ms~20ms，模型大小不到原来的八分之一，相比较YOLO-tiny, 在同样的训练方式下map能够提升约20个百分点。

下载安装命令

## CPU版本安装命令
pip install -f https://paddlepaddle.org.cn/pip/oschina/cpu paddlepaddle

## GPU版本安装命令
pip install -f https://paddlepaddle.org.cn/pip/oschina/gpu paddlepaddle-gpu

ShuffleNetV2-YOLOv3模型结构

用Paddle实现高性能目标检测网络：ShuffleNetV2-YOLOv3-LMLPHP

主干网络为ShuffleNetV2, 其主要结构Inverted_residual_unit的组成单元如下图(c)、(d)所示:

用Paddle实现高性能目标检测网络：ShuffleNetV2-YOLOv3-LMLPHP
(a): the basic ShuffleNet-V1 unit; (b) the ShuffleNet-V1 unit for spatial down sampling (2×); (c) ShuffleNet-V2 basic unit; (d) ShuffleNet-V2 unit for spatial down sampling (2×)

ShuffleNetV2网络结构如下：

用Paddle实现高性能目标检测网络：ShuffleNetV2-YOLOv3-LMLPHP

参考链接

YOLOv3

论文原文：https://arxiv.org/pdf/1804.02767v1.pdf

参考链接：https://blog.csdn.net/litt1e/article/details/88907542

ShuffleNetV2

论文原文：ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design

参考链接：ShuffleNetV2：轻量级CNN网络中的桂冠

 

代码结构

code/
  ├── model
  	├── pretrained_model  	        预训练模型
  ├── YOLOv3.py   	        以darknet53为主干网络的yolov3(原版yolov3)
  ├── shufflenet_yolo3.py   	以ShuffleNetV2为主干网络的yolov3
  ├── reader.py   		数据读取脚本
  ├── train.py   		训练脚本
  ├── freeze.py   		模型固化脚本
  ├── infer.py      		预测脚本
  ├── config.py     		配置参数脚本
  ├── train_list.txt     	训练数据列表
  ├── val_list.txt     		验证数据列表
  ├── label_list      		label列表
 

数据集

本项目采用Pascal VOC数据集，数据集具体介绍可以参考https://blog.csdn.net/u013832707/article/details/80060327

本项目根据实际需要，将VOC数据集的Annotation文件做了进一步整理，写入在code/train_list.txt, code/val_list.txt方便reader调用

In[1] 
#解压数据集，并将voc数据集的训练和验证文件拷贝到相应位置
!cd data/data4379/ && unzip -qo pascalvoc.zip
!cd data/data4379/pascalvoc/ && rm label_list
!cp code/train_list.txt data/data4379/pascalvoc/
!cp code/val_list.txt data/data4379/pascalvoc/
!cp code/label_list data/data4379/pascalvoc/
In[2] 
#模型训练
!cd code && python train.py
{'aeroplane': 0, 'bicycle': 1, 'bird': 2, 'boat': 3, 'bottle': 4, 'bus': 5, 'car': 6, 'cat': 7, 'chair': 8, 'cow': 9, 'diningtable': 10, 'dog': 11, 'horse': 12, 'motorbike': 13, 'person': 14, 'pottedplant': 15, 'sheep': 16, 'sofa': 17, 'train': 18, 'tvmonitor': 19}
2020-02-27 14:29:34,172-INFO: start train YOLOv3, train params:{'data_dir': '../data/data4379/pascalvoc', 'train_list': 'train_list.txt', 'eval_list': 'val_list.txt', 'use_filter': False, 'class_dim': 20, 'label_dict': {'aeroplane': 0, 'bicycle': 1, 'bird': 2, 'boat': 3, 'bottle': 4, 'bus': 5, 'car': 6, 'cat': 7, 'chair': 8, 'cow': 9, 'diningtable': 10, 'dog': 11, 'horse': 12, 'motorbike': 13, 'person': 14, 'pottedplant': 15, 'sheep': 16, 'sofa': 17, 'train': 18, 'tvmonitor': 19}, 'num_dict': {0: 'aeroplane', 1: 'bicycle', 2: 'bird', 3: 'boat', 4: 'bottle', 5: 'bus', 6: 'car', 7: 'cat', 8: 'chair', 9: 'cow', 10: 'diningtable', 11: 'dog', 12: 'horse', 13: 'motorbike', 14: 'person', 15: 'pottedplant', 16: 'sheep', 17: 'sofa', 18: 'train', 19: 'tvmonitor'}, 'image_count': 16551, 'continue_train': True, 'pretrained': True, 'pretrained_model_dir': 'model/pretrained_model', 'save_model_dir': 'model/model', 'model_prefix': 'yolo-v3', 'freeze_dir': 'model/freeze_model', 'use_tiny': False, 'max_box_num': 50, 'num_epochs': 1, 'train_batch_size': 16, 'use_gpu': True, 'yolo_cfg': {'input_size': [3, 384, 384], 'anchors': [9, 12, 15, 28, 30, 21, 28, 56, 57, 42, 54, 110, 107, 83, 144, 183, 344, 301], 'anchor_mask': [[6, 7, 8], [3, 4, 5], [0, 1, 2]]}, 'yolo_tiny_cfg': {'input_size': [3, 384, 384], 'anchors': [9, 12, 20, 23, 33, 51, 72, 75, 122, 150, 308, 287], 'anchor_mask': [[3, 4, 5], [0, 1, 2]]}, 'ignore_thresh': 0.7, 'mean_rgb': [127.5, 127.5, 127.5], 'mode': 'train', 'multi_data_reader_count': 4, 'apply_distort': True, 'nms_top_k': 400, 'nms_pos_k': 100, 'valid_thresh': 0.005, 'nms_thresh': 0.45, 'image_distort_strategy': {'expand_prob': 0.5, 'expand_max_ratio': 4, 'hue_prob': 0.5, 'hue_delta': 18, 'contrast_prob': 0.5, 'contrast_delta': 0.5, 'saturation_prob': 0.5, 'saturation_delta': 0.5, 'brightness_prob': 0.5, 'brightness_delta': 0.5}, 'sgd_strategy': {'learning_rate': 0.002, 'lr_epochs': [10, 45, 80, 110, 135, 160, 180], 'lr_decay': [1, 0.5, 0.25, 0.1, 0.025, 0.004, 0.001, 0.0005]}, 'early_stop': {'sample_frequency': 50, 'rise_limit': 10, 'min_loss': 5e-08, 'min_curr_map': 0.84}}
2020-02-27 14:29:34,172 - train.py[line:429] - INFO: start train YOLOv3, train params:{'data_dir': '../data/data4379/pascalvoc', 'train_list': 'train_list.txt', 'eval_list': 'val_list.txt', 'use_filter': False, 'class_dim': 20, 'label_dict': {'aeroplane': 0, 'bicycle': 1, 'bird': 2, 'boat': 3, 'bottle': 4, 'bus': 5, 'car': 6, 'cat': 7, 'chair': 8, 'cow': 9, 'diningtable': 10, 'dog': 11, 'horse': 12, 'motorbike': 13, 'person': 14, 'pottedplant': 15, 'sheep': 16, 'sofa': 17, 'train': 18, 'tvmonitor': 19}, 'num_dict': {0: 'aeroplane', 1: 'bicycle', 2: 'bird', 3: 'boat', 4: 'bottle', 5: 'bus', 6: 'car', 7: 'cat', 8: 'chair', 9: 'cow', 10: 'diningtable', 11: 'dog', 12: 'horse', 13: 'motorbike', 14: 'person', 15: 'pottedplant', 16: 'sheep', 17: 'sofa', 18: 'train', 19: 'tvmonitor'}, 'image_count': 16551, 'continue_train': True, 'pretrained': True, 'pretrained_model_dir': 'model/pretrained_model', 'save_model_dir': 'model/model', 'model_prefix': 'yolo-v3', 'freeze_dir': 'model/freeze_model', 'use_tiny': False, 'max_box_num': 50, 'num_epochs': 1, 'train_batch_size': 16, 'use_gpu': True, 'yolo_cfg': {'input_size': [3, 384, 384], 'anchors': [9, 12, 15, 28, 30, 21, 28, 56, 57, 42, 54, 110, 107, 83, 144, 183, 344, 301], 'anchor_mask': [[6, 7, 8], [3, 4, 5], [0, 1, 2]]}, 'yolo_tiny_cfg': {'input_size': [3, 384, 384], 'anchors': [9, 12, 20, 23, 33, 51, 72, 75, 122, 150, 308, 287], 'anchor_mask': [[3, 4, 5], [0, 1, 2]]}, 'ignore_thresh': 0.7, 'mean_rgb': [127.5, 127.5, 127.5], 'mode': 'train', 'multi_data_reader_count': 4, 'apply_distort': True, 'nms_top_k': 400, 'nms_pos_k': 100, 'valid_thresh': 0.005, 'nms_thresh': 0.45, 'image_distort_strategy': {'expand_prob': 0.5, 'expand_max_ratio': 4, 'hue_prob': 0.5, 'hue_delta': 18, 'contrast_prob': 0.5, 'contrast_delta': 0.5, 'saturation_prob': 0.5, 'saturation_delta': 0.5, 'brightness_prob': 0.5, 'brightness_delta': 0.5}, 'sgd_strategy': {'learning_rate': 0.002, 'lr_epochs': [10, 45, 80, 110, 135, 160, 180], 'lr_decay': [1, 0.5, 0.25, 0.1, 0.025, 0.004, 0.001, 0.0005]}, 'early_stop': {'sample_frequency': 50, 'rise_limit': 10, 'min_loss': 5e-08, 'min_curr_map': 0.84}}
2020-02-27 14:29:34,173-INFO: create place, use gpu:True
2020-02-27 14:29:34,173 - train.py[line:431] - INFO: create place, use gpu:True
2020-02-27 14:29:34,173-INFO: build network and program
2020-02-27 14:29:34,173 - train.py[line:434] - INFO: build network and program
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/nn.py:10251: UserWarning: actual_shape will be deprecated, it is recommended to use out_shape instead of actual_shape to specify output shape dynamically.
  "actual_shape will be deprecated, it is recommended to use "
2020-02-27 14:29:34,586-INFO: origin learning rate: 0.002 boundaries: [10340, 46530, 82720, 113740, 139590, 165440, 186120]  values: [0.002, 0.001, 0.0005, 0.0002, 5e-05, 8e-06, 2e-06, 1e-06]
2020-02-27 14:29:34,586 - train.py[line:255] - INFO: origin learning rate: 0.002 boundaries: [10340, 46530, 82720, 113740, 139590, 165440, 186120]  values: [0.002, 0.001, 0.0005, 0.0002, 5e-05, 8e-06, 2e-06, 1e-06]
2020-02-27 14:29:36,470-INFO: build executor and init params
2020-02-27 14:29:36,470 - train.py[line:447] - INFO: build executor and init params
W0227 14:29:37.427084   115 device_context.cc:236] Please NOTE: device: 0, CUDA Capability: 70, Driver API Version: 9.2, Runtime API Version: 9.0
W0227 14:29:37.430891   115 device_context.cc:244] device: 0, cuDNN Version: 7.3.
2020-02-27 14:29:38,703-INFO: load param from pretrained model
2020-02-27 14:29:38,703 - train.py[line:415] - INFO: load param from pretrained model
2020-02-27 14:29:38,804-INFO: current pass: 0, start read image
2020-02-27 14:29:38,804 - train.py[line:466] - INFO: current pass: 0, start read image
2020-02-27 14:32:15,026-INFO: pass 0, trainbatch 200, loss 13.787325859069824, time 0.23 sec
2020-02-27 14:32:15,026 - train.py[line:481] - INFO: pass 0, trainbatch 200, loss 13.787325859069824, time 0.23 sec
2020-02-27 14:34:54,558-INFO: pass 0, trainbatch 400, loss 32.522918701171875, time 0.22 sec
2020-02-27 14:34:54,558 - train.py[line:481] - INFO: pass 0, trainbatch 400, loss 32.522918701171875, time 0.22 sec
2020-02-27 14:37:32,941-INFO: pass 0, trainbatch 600, loss 35.67811584472656, time 0.22 sec
2020-02-27 14:37:32,941 - train.py[line:481] - INFO: pass 0, trainbatch 600, loss 35.67811584472656, time 0.22 sec
2020-02-27 14:40:11,286-INFO: pass 0, trainbatch 800, loss 21.77606964111328, time 0.41 sec
2020-02-27 14:40:11,286 - train.py[line:481] - INFO: pass 0, trainbatch 800, loss 21.77606964111328, time 0.41 sec
2020-02-27 14:42:47,528-INFO: pass 0, trainbatch 1000, loss 26.96040916442871, time 0.22 sec
2020-02-27 14:42:47,528 - train.py[line:481] - INFO: pass 0, trainbatch 1000, loss 26.96040916442871, time 0.22 sec
2020-02-27 14:43:14,532-INFO: pass 0 train result, current pass mean loss: 23.394970231355675
2020-02-27 14:43:14,532 - train.py[line:483] - INFO: pass 0 train result, current pass mean loss: 23.394970231355675
pred (160135, 6)
2020-02-27 14:47:48,308-INFO: 0 epoch current pass map is 0.15955796837806702, accum_map is 0.15955796837806702
2020-02-27 14:47:48,308 - train.py[line:489] - INFO: 0 epoch current pass map is 0.15955796837806702, accum_map is 0.15955796837806702
2020-02-27 14:47:48,310-INFO: model save 0 epcho train result, current best pass MAP 0.15955796837806702
2020-02-27 14:47:48,310 - train.py[line:495] - INFO: model save 0 epcho train result, current best pass MAP 0.15955796837806702
2020-02-27 14:47:48,859-INFO: best pass 0 current best pass MAP is 0.15955796837806702
2020-02-27 14:47:48,859 - train.py[line:499] - INFO: best pass 0 current best pass MAP is 0.15955796837806702
2020-02-27 14:47:48,861-INFO: end training
2020-02-27 14:47:48,861 - train.py[line:516] - INFO: end training
In[3] 
#模型固化
!cd code && python freeze.py
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/nn.py:10251: UserWarning: actual_shape will be deprecated, it is recommended to use out_shape instead of actual_shape to specify output shape dynamically.
  "actual_shape will be deprecated, it is recommended to use "
freeze end
In[4] 
#模型预测
!cd code && python infer.py

######检测结果可视化
%matplotlib inline
import matplotlib.pyplot as plt
import numpy as np
import cv2

detect_img= cv2.imread('code/result.jpg')

plt.imshow(detect_img)
plt.show()
W0227 14:47:55.996620   265 device_context.cc:236] Please NOTE: device: 0, CUDA Capability: 70, Driver API Version: 9.2, Runtime API Version: 9.0
W0227 14:47:56.000547   265 device_context.cc:244] device: 0, cuDNN Version: 7.3.
predict cost time:38.32 ms
检测到目标
检测结果保存为result.jpg
infer one picture cost 38.31744194030762 ms
用Paddle实现高性能目标检测网络：ShuffleNetV2-YOLOv3-LMLPHP

点击链接，使用AI Studio一键上手实践项目吧：https://aistudio.baidu.com/aistudio/projectdetail/273746 

下载安装命令

## CPU版本安装命令
pip install -f https://paddlepaddle.org.cn/pip/oschina/cpu paddlepaddle

## GPU版本安装命令
pip install -f https://paddlepaddle.org.cn/pip/oschina/gpu paddlepaddle-gpu

>> 访问 PaddlePaddle 官网，了解更多相关内容。

09-04 15:24
Powered by LMLPHP ©2024 飞桨PaddlePaddle 0.090279