YOLO v4

YOLOv4, YOLOv4-tiny Implemented in Tensorflow 2.0.

COCO Dataset to YOLO format

• Please refer to this url.

data_root_path(default:./data)
ㄴdataset
	ㄴcoco_val2017.txt
	ㄴcoco_train2017.txt

in coco_%2017.txt
dataset_folder_path/images/%2017/000000000139.jpg
.....

dataset_folder_path
ㄴimages
	ㄴval2017
		ㄴ000000000139.jpg
		ㄴ....jpg
	ㄴtrain2017
		ㄴ....jpg
ㄴlabels
	ㄴval2017
		ㄴ000000000139.txt
		ㄴ....txt
	ㄴtrain2017
		ㄴ....txt

Train and Test(eval)

If you run the following, it will be executed as default, and the description of additional parameters is as follows.

python train.py
python eval.py

Train argument

• batch_size : Size of batch size. (But, The gradient update is performed every 64/batch size.) / default : 4 / 64 (YOLO v4/ YOLOv4 tiny)

• img_size : Size of input image. / default : 416

• data_root: Root path of class name file and coco_%2017.txt / default : './data'

• class_file : Class name file / default : 'coco.name'

• num_classes : Number of classes (in COCO 80) / default : 80

• augment : Flag of augmentation (hsv, flip, random affine) / default : true

• mosaic : Flag of mosaic augmentation / default : true

• is_shuffle : Flag of data shuffle / default : true

• train_by_steps : Flag for whether to proceed with the training by step or epoch. / default : false

• train_steps : This is the total iteration to be carried out based on 64 batches. So the total steps will be train_steps * 64 / batch_size. Used only when the train_by_steps flag is true. / default : 500500

• epochs : Total epochs to be trained. Used only when the train_by_steps flag is false. / default : 300

• warmup_by_steps : Flag for whether to proceed with the warm up by step or epoch. / default : false

• warmup_steps : This is the total iteration of warm up to be carried out based on 64 batches. So the steps will be warm up_steps * 64 / batch_size. Used only when the warmup_by_steps flag is true. / default : 1000

• warmup_epochs : Total epochs to warm up. Used only when the warmup_by_steps flag is false. / default : 3

• save_steps : Step cycle to store weight. /default : 1000

• is_tiny : Flag for using tiny. / default : false

• soft : This is a value for soft labeling, and soft/num_class becomes the label for negative class. / default : 0.0

• log_path : logdir path for Tensorboard

• weight_path : Path of weight file / default : ''

• weight_save_path : Path to store weights. / default : './wegiht'

Test argument

• img_size : Size of input image. / default : 416

• data_root: Root path of class name file and coco_%2017.txt / default : './data'

• class_file : Class name file / default : 'coco.name'

• num_classes : Number of classes (in COCO 80) / default : 80

• augment : Flag of augmentation (hsv, flip, random affine) / default : false

• mosaic : Flag of mosaic augmentation / default : false

• is_shuffle : Flag of data shuffle / default : true

• only_coco_eval : When the flag is true, only pycocotools is used to show the result. In this case, you must enter the path to the json file. / default : false

• out_json_path : Folder of output file (json) / default : './eval'

• json_path : Path of json result file. This flag only used when the only_coco_eval flag is true

• weight_path : path of weight

• annotation_path : COCO annotation file folder / default : './data/dataset/COCO/annotations'

• is_darknet_wegiht : If true, load the weight file used by the darknet framework.

• is_tiny : Flag for using tiny. / default : false

• is_padding : If true, padding is performed to maintain the ratio of the input image. / default : false

Convert to tflite

1. Download weights from darknet or yolov4.weights

2. Convert to keras model

   python save_model.py --data_root ./data --class_file coco.names --is_darknet_weight --weight_path ./weight/yolov4.weights --num_classes 80 --out_path ./saved_model/darknet_weight
   

3. Convert to tflite

   python convert_tflite.py --num_classes 80 --dataset './data/dataset/coco_val2017.txt' --weight_path ./saved_model/darknet_weight --out_model ./darknet.tflite --quantize_mode float16
   

4. Detect

   python detect.py --num_classes 80 --input_dir './data/dataset/COCO/images/val2017' --weight_path darknet.tflite --data_name cooc --is_tflite
   

References

• YOLOv4: Optimal Speed and Accuracy of Object Detection YOLOv4.
• darknet
• Pytorch YOLOv4 Implemenation

• tensorflow-yolov4-tflite