Are you sure you want to delete this task? Once this task is deleted, it cannot be recovered.
You can not select more than 25 topics
Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.
肖雄
08190e7e88
|
1 year ago | |
|---|---|---|
| PaddleDetection | 1 year ago | |
| README.md | 1 year ago | |
| main.ipynb | 1 year ago | |
README.md
零.写在最前
该项目源于《飞桨学习赛：钢铁缺陷检测挑战赛》
基于飞桨目标检测开发套件PaddleDetection进行快速微调和预测，生成比赛结果文件。
一.安装PaddleDetection
PaddleDetection为基于飞桨PaddlePaddle的端到端目标检测套件，内置30+模型算法及250+预训练模型，覆盖目标检测、实例分割、跟踪、关键点检测等方向，其中包括服务器端和移动端高精度、轻量级产业级SOTA模型、冠军方案和学术前沿算法，并提供配置化的网络模块组件、十余种数据增强策略和损失函数等高阶优化支持和多种部署方案，在打通数据处理、模型开发、训练、压缩、部署全流程的基础上，提供丰富的案例及教程，加速算法产业落地应用。
二.数据预处理
将数据集生成为符合PaddleDetection的格式,这里是VOC格式
格式如下：
# 由于云脑创建调试任务比较慢,我们先将克隆项目到本地
git clone https://openi.pcl.ac.cn/xiaoxiong/Steel-defect-detection.git
# 安装paddle环境
pip install paddlepaddle-gpu -i https://mirror.baidu.com/pypi/simple
# 安装PaddleDetection及相关依赖
pip install paddledetection
cd PaddleDetection
pip install --upgrade -r requirements.txt -i https://mirror.baidu.com/pypi/simple
打开本地的notebook环境
下载数据集到本地:https://openi.pcl.ac.cn/attachments/bfc6715a-686e-4c0a-890d-18f2aea826e9?type=0
# 解压数据集
!unzip -o 钢铁缺陷检测数据集.zip
%cd 钢铁缺陷检测数据集
# 重命名
! mv train/ANNOTATIONS/ train/annotations
! mv train/IMAGES/ train/images
! mv test/IMAGES/ test/images
import random
class Pre(object):
def __init__(self):
self.categorys = {
'crazing':0,
'inclusion':1,
'pitted_surface':2,
'scratches':3,
'patches':4,
'rolled-in_scale':5
}
self.trainDataRatio = 0.9
def run(self):
# 生成label_list.txt
labelFile = open('train/label_list.txt', 'w')
for key, value in self.categorys.items():
labelFile.write(key)
labelFile.write('\n')
# 生成train.txt和valid.txt
trainFile = open('train/train.txt', 'w')
vaildFile = open('train/valid.txt', 'w')
trainNums = int(1400*self.trainDataRatio)
alls = range(0,1400)
trainIndexs = random.sample(alls, trainNums)
vaildIndexs = ret = list(set(alls) ^ set(trainIndexs))
for k in trainIndexs:
trainFile.write('./images/'+str(k)+'.jpg ./annotations/'+str(k)+'.xml')
trainFile.write('\n')
for k in vaildIndexs:
vaildFile.write('./images/'+str(k)+'.jpg ./annotations/'+str(k)+'.xml')
vaildFile.write('\n')
preObj = Pre()
preObj.run()
# 拷贝数据集到 PadddleDetection 对应目录下
!mkdir -p PaddleDetection/dataset/mine_voc/
!cp -r train/* PaddleDetection/dataset/mine_voc/
!cp -r test PaddleDetection/dataset/mine_voc/
三.训练
训练需要配置文件。
1.入口配置文件:
PaddleDetection/configs/ppyoloe/ppyoloe_plus_crn_x_80e_coco.yml
_BASE_: [
'../datasets/mine_voc.yml',
'../runtime.yml',
'./_base_/optimizer_80e.yml',
'./_base_/ppyoloe_plus_crn.yml',
'./_base_/ppyoloe_plus_reader.yml',
]
log_iter: 5
snapshot_epoch: 1
weights: output/ppyoloe_plus_crn_x_80e_coco/model_final
pretrain_weights: https://bj.bcebos.com/v1/paddledet/models/pretrained/ppyoloe_crn_x_obj365_pretrained.pdparams
mine_voc.yml 主要说明了训练数据和验证数据的路径
runtime.yml 主要说明了公共的运行参数，比如说是否使用GPU、每多少个epoch存储checkpoint等
optimizer_80e.yml 主要说明了学习率和优化器的配置。
ppyoloe_plus_crn.yml 主要说明模型、和主干网络的情况。
ppyoloe_plus_reader.yml 主要说明数据读取器配置，如batch size，并发加载子进程数等，同时包含读取后预处理操作，如resize、数据增强等等
# 单卡GPU 训练
%cd PaddleDetection/
!export CUDA_VISIBLE_DEVICES=0
# 训练，--eval 边训练边评估
!python tools/train.py \
-c configs/ppyoloe/ppyoloe_plus_crn_x_80e_coco.yml \
--use_vdl=true \
--vdl_log_dir=vdl_dir/scalar \
--eval
# [11/09 20:15:02] ppdet.engine INFO: Epoch: [53] [ 0/157] learning_rate: 0.000043 loss: 1.640873 loss_cls: 0.768857 loss_iou: 0.159550 loss_dfl: 0.975133 loss_l1: 0.374083 eta: 0:33:40 batch_cost: 0.4564 data_cost: 0.0005 ips: 17.5284 images/s
# [11/09 20:15:14] ppdet.engine INFO: Epoch: [53] [ 20/157] learning_rate: 0.000042 loss: 1.623144 loss_cls: 0.784792 loss_iou: 0.148097 loss_dfl: 0.928553 loss_l1: 0.350931 eta: 0:33:30 batch_cost: 0.4790 data_cost: 0.0003 ips: 16.7022 images/s
# [11/09 20:15:27] ppdet.engine INFO: Epoch: [53] [ 40/157] learning_rate: 0.000042 loss: 1.667645 loss_cls: 0.778825 loss_iou: 0.157998 loss_dfl: 0.953017 loss_l1: 0.396211 eta: 0:33:21 batch_cost: 0.4920 data_cost: 0.0003 ips: 16.2617 images/s
# [11/09 20:15:39] ppdet.engine INFO: Epoch: [53] [ 60/157] learning_rate: 0.000042 loss: 1.702080 loss_cls: 0.819409 loss_iou: 0.166717 loss_dfl: 0.961283 loss_l1: 0.377798 eta: 0:33:11 batch_cost: 0.4605 data_cost: 0.0003 ips: 17.3721 images/s
# [11/09 20:15:50] ppdet.engine INFO: Epoch: [53] [ 80/157] learning_rate: 0.000042 loss: 1.685796 loss_cls: 0.820941 loss_iou: 0.159150 loss_dfl: 0.994252 loss_l1: 0.389399 eta: 0:33:01 batch_cost: 0.4379 data_cost: 0.0003 ips: 18.2674 images/s
# [11/09 20:16:02] ppdet.engine INFO: Epoch: [53] [100/157] learning_rate: 0.000042 loss: 1.617593 loss_cls: 0.744304 loss_iou: 0.155046 loss_dfl: 0.947161 loss_l1: 0.374160 eta: 0:32:52 batch_cost: 0.4626 data_cost: 0.0003 ips: 17.2947 images/s
# [11/09 20:16:14] ppdet.engine INFO: Epoch: [53] [120/157] learning_rate: 0.000041 loss: 1.675176 loss_cls: 0.807137 loss_iou: 0.160434 loss_dfl: 0.964645 loss_l1: 0.397004 eta: 0:32:42 batch_cost: 0.4787 data_cost: 0.0003 ips: 16.7128 images/s
# [11/09 20:16:26] ppdet.engine INFO: Epoch: [53] [140/157] learning_rate: 0.000041 loss: 1.648713 loss_cls: 0.788865 loss_iou: 0.151504 loss_dfl: 0.928811 loss_l1: 0.361151 eta: 0:32:33 batch_cost: 0.4757 data_cost: 0.0003 ips: 16.8159 images/s
# [11/09 20:16:41] ppdet.utils.checkpoint INFO: Save checkpoint: output/ppyoloe_plus_crn_x_80e_coco
# [11/09 20:16:41] ppdet.engine INFO: Eval iter: 0
# [11/09 20:16:54] ppdet.metrics.metrics INFO: Accumulating evaluatation results...
# [11/09 20:16:54] ppdet.metrics.metrics INFO: mAP(0.50, integral) = 79.64%
# [11/09 20:16:54] ppdet.engine INFO: Total sample number: 140, averge FPS: 10.8917303840318
# [11/09 20:16:54] ppdet.engine INFO: Best test bbox ap is 0.796.
# [11/09 20:17:06] ppdet.utils.checkpoint INFO: Save checkpoint: output/ppyoloe_plus_crn_x_80e_coco
四.评估模型
可以在output/ppyoloe_plus_crn_x_80e_coco/目录下指定某个epoch的参数，或者直接用最终的参数,或者指定最优权重
%cd PaddleDetection/
!python tools/eval.py \
-c configs/ppyoloe/ppyoloe_plus_crn_x_80e_coco.yml \
-o weights=output/ppyoloe_plus_crn_x_80e_coco/best_model.pdparams
五.预测
--infer_dir 表示预测的图片文件夹
--output_dir 表示输出的图片文件夹
--draw_threshold 可视化阀值
-o weights 权重路径
--use_vdl 可视化
--save_results 保存结果，这里设置为True
%cd PaddleDetection/
!python tools/infer.py -c configs/ppyoloe/ppyoloe_plus_crn_x_80e_coco.yml \
--infer_dir=dataset/mine_voc/test/images \
--output_dir=mine_output/ \
--draw_threshold=0.1 \
-o weights=output/ppyoloe_plus_crn_x_80e_coco/best_model.pdparams \
--use_vdl=True \
--save_results=True
结果保存在mine_output目录下
特别注意。在aistudio 平台上运行infer.py，得到的测试集图片列表不一定是按照数字大小排序，所以修改了infer.py源代码，修改部分如下图
六.生成比赛结果
import glob
import os
import json
import pandas as pd
class Result(object):
def __init__(self):
self.imagesPath = 'PaddleDetection/mine_output/'
self.bboxPath = 'PaddleDetection/mine_output/bbox.json'
self.submissionPath = 'submission.csv'
def run(self):
images = self.get_image_ids()
bbox = self.get_bbox()
results = []
for i in range(len(images)):
image_id = images[i]
for j in range(len(bbox['bbox'][i])):
bbox_ = [int(i) for i in bbox['bbox'][i][j]]
item = [
image_id,
bbox_,
int(bbox['label'][i][j]),
round(bbox['score'][i][j],2)
]
results.append(item)
submit = pd.DataFrame(results, columns=['image_id', 'bbox','category_id','confidence'])
submit[['image_id', 'bbox','category_id','confidence']].to_csv(self.submissionPath, index=False)
def get_image_ids(self):
images = set()
exts = ['jpg']
for ext in exts:
images.update(glob.glob('{}/*.{}'.format(self.imagesPath, ext)))
images = list(images)
lists = []
for item in images:
item = item.replace(self.imagesPath,'')
lists.append(item)
lists.sort(key=lambda x:int(x.split('.')[0]))
ids = []
for fname in lists:
ids.append(fname.replace('.jpg',''))
return ids
def get_bbox(self):
with open(self.bboxPath, 'r', encoding='utf-8') as bbox:
bbox = json.load(bbox)
return bbox
resultObj = Result()
resultObj.run()
最终生成的结果保存在submission.csv 下，将其压缩后进行提交为11月榜5
7.总结
通过PaddleDetection可以快速实现检测目标任务
优化方向
1.本项目用的是ppyoloe_x_80e模型, 可以尝试换下学术界霸榜的transformer系列模型
2.诸如学习率，数据增强等参数还有调优的可能
基于飞桨目标检测开发套件PaddleDetection进行快速微调训练，评估和预测，一键生成比赛结果文件。模型采用ppyoloe_x_80e，成绩为11月榜5：45.358
Python Markdown C++ Text Shell other
Contributors (1)
