Yolov3 trains its own data set (mmdetection)

use FasterRcnn Trained the data set marked by myself Voc Format , Now I want to use yolo Let's train , Revised yolo The contents of the document , I'm going to use it directly yolo Training voc Formatted data , There's a problem , Because I'm in a hurry , There is no further detailed study .
MMDetection Most of the training models are coco Format design , Plan to put voc Format conversion to coco Format , It is also convenient for the training of other models in the future .

1.voc Format conversion to coco Format

May refer to ：

https://github.com/Stephenfang51/VOC_to_COCO

2.coco Modification of relevant documents

May refer to
【mmdetection】 Use custom coco Format data set for training and testing

（1） Define data types （mmdetection/mmdet/datasets/coco.py）, hold CLASSES the tuple Change to the category corresponding to your data set tuple that will do .

   CLASSES = ('Other Car', 'Taxi')

（2） modify coco_classes Dataset categories （mmdetection/mmdet/core/evaluation/class_names.py）

def coco_classes():
    return [
        'Other Car', 'Taxi'
    ]

（3） Modify usage model model In the dictionary num_classes

num_classes=2,# Number of categories 

3. Training

python tools/train.py configs/yolo/yolov3_d53_mstrain-608_273e_coco.py