关于mmdection一些问题

不同数据集的evaluation配置不同
# evaluation = dict(interval=1, metric='bbox') //coco格式
# evaluation = dict(interval=1, metric='mAP') //voc格式
数据集的图片格式问题

在mmdet/datasets/xml_style.py下更改jpg或者png等格式

数据集labels_name

mmdet/datasets/voc.py文件下更改成自定义数据集的label_name

模型的配置文件中更改类别数量

mmdet/core/evaluation/class_names.py文件中对应的数据格式，改成自己数据集的label_name

tools工具箱下的可视化使用方法

python analyze_logs.py plot_curve log.json --keys loss_cls --legend loss_cls

python analyze_logs.py plot_curve log.json --keys loss_cls loss_bbox --out losses.pdf

python analyze_logs.py plot_curve log.json log2.json --keys loss_bbox --legend run1 run2

python analyze_logs.py plot_curve log.json log2.json --keys bbox_mAP --legend run1 run2

python analyze_logs.py cal_train_time log.json --include-outliers

此外，必须先test_robustness.py，然后才能robustness_eval.py

使用tools下的数据集转换工具

mmdet/core/evaluation/class_names.py中voc_name()修改为自己数据集的labels_name

下面这段代码是用来生成train、val

import os

# 定义VOC文件存放的路径
xml_dir = r'annotations'
img_dir = r'images'
train_txt_path = r'train.txt'
valid_txt_path = r'val.txt'
label_list_txt_path = r'label_list.txt'
# 读取标注文件
xml_files = os.listdir(xml_dir)
print('数据集含有 ' + str(len(xml_files)) + '张图片')
# 训练集和测试集占比
radio = 0.8
trainset = xml_files[:int(radio*len(xml_files))]
validset = xml_files[int(radio*len(xml_files)):]

with open(train_txt_path, 'w') as f:
    #for i in trainset:
      #  img_path = './images/'+ i[:-4] + '.png'
        #xml_path = './annotations/'+ i 
        #text = img_path + ' ' + xml_path + '\n'
        #f.write(text)
    for i in trainset:
        #img_path = i[:-4]+ '\n'
        #xml_path = './annotations/'+ i 
        xml_path = i+'\n'
        text = xml_path
        f.write(text)
f.close()
print('生成train.txt完毕...')

with open(valid_txt_path, 'w') as f:
    for i in validset:
        img_path = i[:-4]+ '\n'
        #xml_path = './annotations/'+ i 
        xml_path = i+'\n'
        text = xml_path
        f.write(text)
    #for i in validset:
        #img_path = './images/'+ i[:-4] + '.png'
       # xml_path = './annotations/'+ i 
        #text = img_path + ' ' + xml_path + '\n'
        #f.write(text)
f.close()
print('生成valid.txt完毕...')


labels = ['helmet', 'head']
with open(label_list_txt_path, 'w') as f:
    for label in labels:
        f.write(label+'\n')
f.close()
print('生成label_list.txt完毕...')

使用tools/benchmark.py开启分布式出错，在py文件开头加入：

os.environ['RANK']='0'
# os.environ['local_rank']='0'
os.environ['WORLD_SIZE']='1'
os.environ['MASTER_ADDR'] = 'localhost'
os.environ['MASTER_PORT'] = '5678'
# dist.init_process_group(backend='nccl', init_method='env://', rank = 0, world_size = 1)

深度学习专栏

关于mmdection一些问题

相关

标签