Implementation of yolov3 and yolov3-tiny by pytorch.
most of codes are copied from here: https://github.com/TencentYoutuResearch/ObjectDetection-OneStageDet/tree/master/yolo
The original code was hard to understand, so I reorganized the code structure and added new features.
Additional functions
1. use sigmoid predict(replace softmax) class prob
2. support train 1 class dataset
3. Multi GPU training
Usage method
Environment: ubuntu, python3.6, pytroch 0.4
Prepare dataset:
1. donwnload dataset:
wget https://pjreddie.com/media/files/VOCtrainval_11-May-2012.tar
wget https://pjreddie.com/media/files/VOCtrainval_06-Nov-2007.tar
wget https://pjreddie.com/media/files/VOCtest_06-Nov-2007.tar
tar xf VOCtrainval_11-May-2012.tar
tar xf VOCtrainval_06-Nov-2007.tar
tar xf VOCtest_06-Nov-2007.tar
2. modify "ROOT" to VOCdevkit folder in labels_voc.py, such as: ROOT = '/media/data/VOCdevkit'
3. python3.6 labels_voc.py
4. the new dataset annotations will be saved in ROOT+"/onedet_cache" folder
How to train:
train yolov3-tiny:
1. modify "data_root_dir" to your dataset folder in "cfg/tiny_yolov3.yml", such as: '/media/data/VOCdevkit/onedet_cache'
2. modify "model_name" to "TinyYolov3" in train_my.py, such as: model_name = "TinyYolov3"
3. python3.6 train_my.py # train from scratch
train yolov3:
1. download init weight: wget https://pjreddie.com/media/files/darknet53.conv.74
2. modify "data_root_dir" to your dataset folder in "cfg/yolov3.yml", such as: '/media/data/VOCdevkit/onedet_cache'
3. modify "model_name" to "Yolov3" in train_my.py, such as: model_name = "Yolov3"
4. python3.6 train_my.py
the trained model will be saved in folder "output"
Detect a image:
python3.6 predict_cv.py
the images in folder "test_img" will be used, and the results are saved in folder "result"