MMocr模型训练报错记录

模型训练的教程也很简单，按照以下的官网教程就能实现：

欢迎来到 MMOCR 的中文文档! — MMOCR 1.0.1 文档

在这里提醒一下：文字检测的模型目前除咯db,mask_rcnn能跑，其余的drrg,textsnake模型我都跑不不通，大家如果遇到报错,就别跑咯。我搜集咯资料，这个原因是标注的信息报错，但是我把数据全部审查咯好几遍，还运用官网的代码转换还是报错，后面就没办法咯。（如果你能跑出来，希望你能给我支招）

assert point.shape[0]>=4

不过之前作为小白的我，就一直卡在数据集着一个步骤，现在详细讲一下：

路径就是mmocr-main/date/brand_data/这样的一个根目录，在根目录下有：train/imgaes和json文件（图中的txt文件就是我的paddleocr标注文件，还有output都是不相干的内容不用在意）

一般就是在ctw这个配置文件修改：

修改_base_四个配置文件，除咯dataset，其他的配置文件基本上不需要修改：

python tools/train.py configs/textdet/maskrcnn/mask-rcnn_resnet50_fpn_160e_ctw1500.py

顺便给一下paddleocr格式转换成mmocr格式的脚本：

import json
import numpy as np
import cv2
import os

# mode = 'train'    
mode = 'test'
# f = open(f'/home/lixinru/Vir_env/mmocr-main/data/brand_data/brand_images/textdet_train.txt', 'r', encoding='utf-8').readlines()
f = open(f'/home/lixinru/Vir_env/mmocr-main/data/brand_data/brand_images/textdet_test.txt', 'r', encoding='utf-8').readlines()
# filename='/home/lixinru/Vir_env/mmocr-main/data/brand_data/brand_images/train/'
filename='/home/lixinru/Vir_env/mmocr-main/data/brand_data/brand_images/test/'


mmocr_label = {"metainfo": {"dataset_type": "TextDetDataset", "task_name": "textdet", "category": [{"id": 0, "name": "text"}]}, "data_list": []}
data_list = []


for line in f:
    instances = {}
    per_img_anno = []
    anno = line.split('\t')

    img_path = os.path.join(filename,anno[0])
    print((img_path))
    images_path=anno[0]

    paddle_label = anno[1]
    paddle_label = json.loads(paddle_label)

    for i in range(len(paddle_label)):
        per_box = {"polygon": [], "bbox": [], "bbox_label": 0}
        polygon = []
        points = paddle_label[i]['points']
        bounding = list(cv2.boundingRect(np.array(points)))
        if len(points) == 4:
            for j in range(len(points)):
                x, y = points[j][0], points[j][1]
                polygon.append(int(x))
                polygon.append(int(y))
                # print()
        else:
            x1, y1 = bounding[0], bounding[1]
            x2, y2 = bounding[0]+bounding[2]-1, bounding[1]
            x3, y3 = bounding[0]+bounding[2]-1, bounding[1]+bounding[3]-1
            x4, y4 = bounding[0], bounding[1]+bounding[3]-1
            polygon.append(int(x1))
            polygon.append(int(y1))
            polygon.append(int(x2))
            polygon.append(int(y2))
            polygon.append(int(x3))
            polygon.append(int(y3))
            polygon.append(int(x4))
            polygon.append(int(y4))

        bbox = [int(bounding[0]), int(bounding[1]), int(bounding[0]+bounding[2]-1), int(bounding[1]+bounding[3]-1)]

        per_box.update({"polygon": polygon})
        per_box.update({"bbox": bbox})
        per_box.update({"bbox_label": 0})
        if paddle_label[i]["transcription"] == "###":
            mark = bool(1)
            per_box.update({"ignore": mark})
        else:
            mark = bool(0)
            per_box.update({"ignore": mark})
        per_img_anno.append(per_box)
        # print()
    instances.update({"instances": per_img_anno})
    instances.update({"img_path": f'{mode}/{images_path}'})
    # instances.update({"img_path": f'{images_path}'})
    img = cv2.imread(img_path)
    instances.update({"height": img.shape[0]})
    instances.update({"weight": img.shape[1]})
    instances.update({"seg_map": f'{mode}/gt_{img_path.split(".")[0]}.txt'})
    # instances.update({"seg_map": f'gt_{img_path.split(".")[0]}.txt'})

    data_list.append(instances)

mmocr_label['data_list'] = data_list
file_name = f"textdet_{mode}.json"
# if not os.path.exists(file_name):
#     os.makedirs(file_name)
with open(f'/home/lixinru/Vir_env/mmocr-main/data/brand_data/brand_images/output/{file_name}', 'w') as f:
    str = json.dumps(mmocr_label)
    f.write(str)
    f.close()

print("close")

判断paddleocr的标注是否是四个标注点的，如果报错我一开始说的问题，就用这个检查一下：

import json


def check_paddleocr_txt_annotations(txt_file_path):
    try:
        with open(txt_file_path, 'r', encoding='utf-8') as file:
            lines = file.readlines()

            for idx, line in enumerate(lines):
                try:
                    line = line.strip().split('\t')
                    file_path = line[0]
                    # print(line[0])
                    annotation_str = line[1]
                    # print(line[1])
                    annotation_data = json.loads(annotation_str)  # 使用json.loads将JSON字符串转换为Python对象

                    # 检查标注点数量
                    count=0
                    for annotation in annotation_data:
                        transcription = annotation["transcription"]
                        points = annotation["points"]
                        if len(points) != 4:
                            print(f"Invalid annotation found in file {file_path}:")
                            count+=1
                            print(f"Transcription: {transcription}")
                            print(f"Points: {points}")
                            print("------")
                except Exception as e:
                    print(f"Error occurred at line {idx + 1}: {str(e)}")
    except Exception as e:
        print(f"Error occurred while checking annotation file: {str(e)}")
        print('number:', count)


# 替换为你的txt标注文件路径
txt_file_path = '/home/lixinru/Vir_env/mmocr-main/data/brand_data/Label_all.txt'
check_paddleocr_txt_annotations(txt_file_path)