- Win10,
tensorflow-cpu 1.9.0,keras 2.1.5,python 3 - Ubuntu 14.04,
tensorflow-gpu 1.10.1,keras 2.2.4,python3,cuda-9.0onGTX1080
This repo is intended for purely training your own dataset of detection. And this repo is mainly modified from qqwweee/keras-yolo3.
Training on own dataset is quite simple, first download (choose one)
- China Jianguoyun yolo_weights.h5
- Google Drive yolo_weights.h5
into model_data/
And provide model_data/own_classes.txt which contains the class name of detecting objects, here I provide a example for only one class detection task.
And then all you need is to prepare own_train.txt in the same directory with train.py, each line of train.txt is of this format: /pat/to/img1 xmin,y_min,x_max,y_max,id, remember no <space> before/after , and an <space> between /path/to/img1 and xmin. Here I provide a example own_train.txt, remember id starts from 0.
like imgs/1.jpg 147,30,437,215,0 147,30,437,215,1 for two objects labelled in one image, <space> shall be inserted between two labels.
python train.py will start training, I recommend reading train.py carefully before starting. If memory out, change batchsize in train.py
The model trained will be stored in logs/000/ , please check.
After training, yolo_detect.py will detect objects on image while you type the image path.
Usage:
python yolo_detect.py --model_path /path/to/models, remember to use the path to models trained stored in logs/000/ mentioned above.
I trained my model on detecting box of input image. My dataset size is 500. And I got loss of 10 after training. The result is shown below. You can download my trained model in (choose one)
- China Jianguoyun trained_weights_final.h5
- Google Drive trained_weights_final.h5
and put it into model_data/
