前言

去年打比赛期间，深入学习了mmdetection框架，今年有个项目着急看效果，在A100服务器上重新安装了mm，发现已经是3.X版本了，改动了一些函数，故重新记录一下。

安装mmcv

pip install -U openmim

mim install mmengine

mim install "mmcv>=2.0.0"

安装mmdetection

git clone https://github.com/open-mmlab/mmdetection.git
cd mmdetection
pip install -v -e .
# "-v" means verbose, or more output
# "-e" means installing a project in editable mode,
# thus any local modifications made to the code will take effect without reinstallation.

验证安装

mim download mmdet --config rtmdet_tiny_8xb32-300e_coco --dest .

python demo/image_demo.py demo/demo.jpg rtmdet_tiny_8xb32-300e_coco.py --weights rtmdet_tiny_8xb32-300e_coco_20220902_112414-78e30dcc.pth --device cpu

出现下图，则证明安装成功

数据集

数据集采用的是布匹瑕疵检测的数据集，

转为COCO

首先将下载后的标签文件转为COCO格式

import json
from tqdm import tqdm
import cv2, os
from glob import glob
from collections import defaultdictbase_dirs = ['/home/xray/guangdong1_round1/', ]mp = {"破洞": 1, "水渍": 2, "油渍": 2, "污渍": 2,  "三丝": 3, "结头": 4, "花板跳": 5, "百脚": 6, "毛粒": 7,"粗经": 8, "松经": 9, "断经": 10, "吊经": 11, "粗维": 12, "纬缩": 13, "浆斑": 14, "整经结": 15, "星跳": 16, "跳花": 16,"断氨纶": 17, "稀密档": 18, "浪纹档": 18, "色差档": 18, "磨痕": 19, "轧痕": 19, "修痕":19, "烧毛痕": 19, "死皱": 20,"云织": 20, "双纬": 20, "双经": 20, "跳纱": 20, "筘路": 20, "纬纱不良": 20}def make_coco_traindataset(images2annos, name='train'):idx = 1image_id = 20190000000images = []annotations = []for im_name in tqdm(images2annos):#         im = cv2.imread(base_dir + 'defect_Images/' + im_name)
#         h, w, _ = im.shapeh, w = 1000, 2446image_id += 1image = {'file_name': im_name, 'width': w, 'height': h, 'id': image_id}images.append(image)annos = images2annos[im_name]for anno in annos:bbox = anno[:-1]seg = [bbox[0], bbox[1], bbox[0], bbox[3],bbox[2], bbox[3], bbox[2], bbox[1]]bbox = [bbox[0], bbox[1], bbox[2] - bbox[0], bbox[3] - bbox[1]]anno_ = {'segmentation': [seg], 'area': bbox[2] * bbox[3], 'iscrowd': 0, 'image_id': image_id,'bbox': bbox, 'category_id': anno[-1], 'id': idx, 'ignore': 0}idx += 1annotations.append(anno_)ann = {}ann['type'] = 'instances'ann['images'] = imagesann['annotations'] = annotationscategory = [{'supercategory':'none', 'id': id, 'name': str(id)} for id in range(1, 21)]ann['categories'] = categoryjson.dump(ann, open(base_dir + '{}.json'.format(name),'w'))for idx, base_dir in enumerate(base_dirs, 1):annos = json.load(open(base_dir + 'Annotations/anno_train.json'))images2annos = defaultdict(list)for anno in annos:images2annos[anno['name']].append(anno['bbox'] + [mp[anno['defect_name']]])make_coco_traindataset(images2annos, 'train' + str(idx))

划分训练集、验证集及测试集

采用的是paddleX工具进行划分，比例为7：2：1

安装PaddlePaddle

python -m pip install paddlepaddle-gpu==2.2.2 -i https://mirror.baidu.com/pypi/simple

安装PaddleX

pip install paddlex -i https://mirror.baidu.com/pypi/simple

划分数据集

paddlex --split_dataset --format COCO --dataset_dir D:/MyDataset --val_value 0.2 --test_value 0.1

修改对应文件

修改coco.py

路径 /mmdetection/mmdet/dataset/coco.py
将类别修改为自己的类别名称。

重新安装

重新执行下述命令，修改才会生效。

pip install -v -e .

修改模型文件

ctrl+F 查找 num_classes, 改为自己数据集的类别数量，不需要加背景

训练

python tools/train.py \${CONFIG_FILE} \[optional arguments]

测试

测试带真值的图像

将测试结果保存到文件夹中

python tools/test.py \configs/rtmdet/rtmdet_l_8xb32-300e_coco.py \checkpoints/rtmdet_l_8xb32-300e_coco_20220719_112030-5a0be7c4.pth \--show-dir faster_rcnn_r50_fpn_1x_results

测试不带真值的图像

批量测试

Todo

错误集锦

ValueError: need at least one array to concatenate

解决方案： 修改coco.py文件对应的类别

Downgrade the protobuf package

If you cannot immediately regenerate your protos, some other possible workarounds are: 1. Downgrade the protobuf package to 3.20.x or lower. 2. Set PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python (but this will use pure-Python parsing and will be much slower).

解决方案：
卸载原有的
重新安装低版本的

pip uninstall protobuf

pip install protobuf==3.20.0