使用MMDetection训练自己的数据集(COCO)

import os
import shutil
 
this_dir_path = './train_labelme/'
destination_directory = './train_new/'
if not os.path.exists(destination_directory):
    os.makedirs(destination_directory)
 
subdirectories = []  # 存储子目录名称的列表
# 遍历目录
for root, dirs, files in os.walk(this_dir_path):
    for dir_name in dirs:
        subdirectories.append(dir_name)
 
# for directory in os.listdir(this_dir_path):
for directory in subdirectories:
    for file in os.listdir(os.path.join(this_dir_path, directory)):
        source_file = os.path.join(os.path.join(this_dir_path, directory), file)
        if directory == '211029-009C':
            # 这四个文件采用的是"line strip"无法转化为coco
            if os.path.splitext(file)[0] == '107_1_590_345_0.513' or \
                    os.path.splitext(file)[0] == '129_3_688_378_0.848' or \
                    os.path.splitext(file)[0] == '10_1_737_180_0.571' or \
                    os.path.splitext(file)[0] == '127_2_590_378_0.492':
                continue
        elif os.path.splitext(source_file)[-1] == '.png':
            new_file_path = destination_directory + file
            print(source_file + '---->' + new_file_path)
            destination_file = os.path.join(destination_directory, file)
            shutil.copy2(source_file, destination_file)
        elif os.path.splitext(source_file)[-1] == '.json':
            new_file_path = destination_directory + file
            print(source_file + '---->' + new_file_path)
            destination_file = os.path.join(destination_directory, file)
            shutil.copy2(source_file, destination_file)

import json
import cv2
import os
import shutil
 
 
def copy2dataset(file_src, annotation, file_dir):
    with open(annotation, 'r', encoding='utf-8') as f:
        file_json = json.load(f)
    for img in file_json['images']:
        img_name = img['file_name']
        for file in os.listdir(file_src):
            if file.endswith(img_name):
                img_name = file
                break
        print(img_name)
        shutil.copyfile(os.path.join(file_src, img_name), os.path.join(file_dir, img_name))
 
 
if __name__ == '__main__':
    # labelme_path = "./train_new"
    file_root = './train_new/'
    saved_coco_path = "./"
    # 创建文件
    if not os.path.exists("%scoco/annotations/" % saved_coco_path):
        os.makedirs("%scoco/annotations/" % saved_coco_path)
    if not os.path.exists("%scoco/train2017/" % saved_coco_path):
        os.makedirs("%scoco/train2017" % saved_coco_path)
    if not os.path.exists("%scoco/val2017/" % saved_coco_path):
        os.makedirs("%scoco/val2017" % saved_coco_path)
 
    annotation_train = './runs/labelme2coco/train.json'
    annotation_val = './runs/labelme2coco/val.json'
    file_dest_train = './coco/train2017/'
    file_dest_val = 'coco/val2017/'
    shutil.copyfile(annotation_train, os.path.join("%scoco/annotations/instances_train2017.json" % saved_coco_path))
    shutil.copyfile(annotation_val, os.path.join("%scoco/annotations/instances_val2017.json" % saved_coco_path))
    copy2dataset(file_root, annotation_train, file_dest_train)
    copy2dataset(file_root, annotation_val, file_dest_val)
 

from pycocotools.coco import COCO
import numpy as np
from matplotlib import pyplot as plt
import cv2 as cv
 
# 加载COCO格式的标注文件
coco = COCO('./runs/labelme2coco/train.json')
 
imgIds = coco.getImgIds()  # 获取所有的image id，可以选择参数 coco.getImgIds(imgIds=[], catIds=[])
imgIds = coco.getImgIds(imgIds=[0, 1, 2])  # 获得image id 为 0，1，2的图像的id
imgIds = coco.getImgIds(catIds=[0, 1, 2])  # 获得包含类别 id 为0，1，2的图像
 
annIds = coco.getAnnIds(catIds=[0, 1, 2])  # 获得类别id为0，1，2的标签
annIds = coco.getAnnIds(imgIds=imgIds[0])  # 获得和image id对应的标签
 
catIds = coco.getCatIds(catNms=['0'])  # 通过类别名筛选
catIds = coco.getCatIds(catIds=[0, 1, 2])  # 通过id筛选
catIds = coco.getCatIds(supNms=[])  # 通过父类的名筛选
 
print('类别信息')
cats_name = coco.loadCats(ids=catIds)
print(cats_name)
 
print('\n标签信息:')
anns = coco.loadAnns(annIds)
bboxes = np.array([i['bbox'] for i in anns]).astype(np.int32)
cats = np.array([i['category_id'] for i in anns])
print(anns)
print('\n从标签中提取的Bounding box:')
print(bboxes)
 
print('图像')
imgIdx = imgIds[0]
img = coco.loadImgs([imgIdx])  # 读取图片信息
img = cv.imread('./train_new/' + img[0]['file_name'])
# 绘制bounding box
for i in range(len(bboxes)):
    p1 = bboxes[i][0:2]
    p2 = bboxes[i][0:2] + bboxes[i][2:4]
 
    cv.rectangle(img, (p1[0], p1[1]), (p2[0], p2[1]), (255, 0, 0))
plt.figure(figsize=(8, 8))
plt.imshow(img)
plt.show()

python  tools/train.py configs/mask_rcnn/mask-rcnn_r50_fpn_1x_coco.py   --work-dir run_workstation

python  tools/train.py run_workstation/mask-rcnn_r50_fpn_1x_coco.py

python  tools/train.py run_workstation/mask-rcnn_r50_fpn_1x_coco.py

python tools/analysis_tools/analyze_logs.py plot_curve  run_workstation/20230602_155324/vis_data/20230602_155324.json --keys acc 

python tools/analysis_tools/analyze_logs.py plot_curve  run_workstation/20230602_155324/vis_data/20230602_155324.json --keys loss_cls loss_bbox loss_mask

python tools/analysis_tools/analyze_logs.py plot_curve  run_workstation/20230602_155324/vis_data/20230602_155324.json --keys acc --out out.pdf

python tools/test.py run_workstation/mask-rcnn_r50_fpn_1x_coco.py run_workstation/epoch_12.pth --show