【人工智能笔记】第二十八节 tensorflow 2.0 实现YoloV3与YoloV4

基于tensorflow2的YoloV3与YoloV4实现

yolov3源码地址：
https://github.com/tfwcn/tensorflow2-machine-vision/tree/master/AIServer/ai_api/ai_models/yolo_v3

yolov4源码地址：
https://github.com/tfwcn/tensorflow2-machine-vision/tree/master/AIServer/ai_api/ai_models/yolo_v4

实现内容

• CSPDarknet53 + SPP + PAN (已实现)
• CIoU-loss (已实现)
• Mish activation (已实现)
• Crossstage partial connections(CSP) (已实现)
• DIoU-NMS (已实现)
• CutMix (未实现)
• Mosaic data augmentation (未实现)
• DropBlock regularization (效果较差，暂不发布)
• Class label smoothing (未实现)
• Multiinput weighted residual connections(MiWRC) (未实现)
• CmBN (未实现)
• Self Adversarial Training (未实现)
• Eliminate gridsensitivity (未实现)
• Using multiple anchors for a single groundtruth (未实现)
• Cosine annealing scheduler (未实现)
• Optimal hyper-parameters (未实现)
• Random training shapes (未实现)
• SAM-block (未实现)

素材标注

素材标注使用Labelme
https://github.com/tfwcn/labelme

转换权重

yolo权重转h5，yolo权重请在官网下载：

https://raw.githubusercontent.com/AlexeyAB/darknet/master/cfg/yolov4.cfg

https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v3_optimal/yolov4.weights

python ./ai_api/ai_models/yolo_v4/convert.py ./data/yolo_v4_weights/yolov3.cfg ./data/yolo_v4_weights/yolov3.weights ./data/yolo_v4_weights/yolo.h5 --plot_model --weights_only

h5转tf，如果类别数有变化，需重新转换

python ./ai_api/ai_models/yolo_v4/convert_tf2.py --old_weights_path ./data/yolo_v4_weights/yolo.h5 --output_path ./data/yolo_v4_weights/tf2_weights/tf2_weights.ckpt --classes_num 80

coco2017标签转换

python ./ai_api/ai_models/utils/coco.py --dataDir ./coco2017 --dataType train2017 --outDir ./data

训练

python ./ai_api/ai_models/yolo_v4/train.py --trainData "./coco2017/train2017" --valData "./coco2017/val2017" --trainLabels "./data/coco_train_labels.txt" --valLabels "./data/coco_test_labels.txt" --classesFile "./data/coco_classes.txt" --anchorsFile "./data/coco_anchors.txt" --batchSize 4

测试

python ./ai_api/ai_models/yolo_v4/test.py --imageFile "./data/img.png" --anchorsFile "./data/coco_anchors.txt" --classesFile "./data/coco_classes.txt" --model_path "./data/yolo_v4_weights/tf2_weights/"

测试

启动服务：

python manage.py runserver 0.0.0.0:8080

浏览器打开：

http://127.0.0.1:8080/static/object_detection/predict_image_read.html

参考资料：

• https://github.com/AlexeyAB/darknet
• https://github.com/qqwweee/keras-yolo3