calmisential / YOLOv3_TensorFlow2

基于TensorFlow2.x实现的YOLOv3，支持在自定义数据集上训练，支持保存为TFLite模型。A tensorflow2 implementation of YOLO_V3(Supports training on custom dataset and saving as tflite models.).

YOLOv3_TensorFlow2

A tensorflow2 implementation of YOLO_V3.

Requirements:

• Python == 3.7
• TensorFlow == 2.1.0
• numpy == 1.17.0
• opencv-python == 4.1.0

Usage

Train on PASCAL VOC 2012

1. Download the PASCAL VOC 2012 dataset.
2. Unzip the file and place it in the 'dataset' folder, make sure the directory is like this :

|——dataset
    |——VOCdevkit
        |——VOC2012
            |——Annotations
            |——ImageSets
            |——JPEGImages
            |——SegmentationClass
            |——SegmentationObject

3. Change the parameters in configuration.py according to the specific situation. Specially, you can set "load_weights_before_training" to True if you would like to restore training from saved weights. You can also set "test_images_during_training" to True, so that the detect results will be show after each epoch.
4. Run write_voc_to_txt.py to generate data.txt, and then run train_from_scratch.py to start training.

Train on COCO2017

1. Download the COCO2017 dataset.
2. Unzip the train2017.zip, annotations_trainval2017.zip and place them in the 'dataset' folder, make sure the directory is like this :

|——dataset
    |——COCO
        |——2017
            |——annotations
            |——train2017

3. Change the parameters in configuration.py according to the specific situation. Specially, you can set "load_weights_before_training" to True if you would like to restore training from saved weights. You can also set "test_images_during_training" to True, so that the detect results will be show after each epoch.
4. Run write_coco_to_txt.py to generate data.txt, and then run train_from_scratch.py to start training.

Train on custom dataset

1. Turn your custom dataset's labels into this form: xxx.jpg 100 200 300 400 1 300 600 500 800 2. The first position is the image name, and the next 5 elements are [xmin, ymin, xmax, ymax, class_id]. If there are multiple boxes, continue to add elements later.
   Considering that the image will be resized before it is entered into the network, the values of xmin, ymin, xmax, and ymax will also change accordingly.
   The example of original picture(from PASCAL VOC 2012 dataset) and resized picture:
   
   Create a new file data.txt in the data_process directory and write the label of each picture into it, each line is a label for an image.
2. Change the parameters CATEGORY_NUM, use_dataset, custom_dataset_dir, custom_dataset_classes in configuration.py.
3. Run write_to_txt.py to generate data.txt, and then run train_from_scratch.py to start training.

Test

1. Change "test_picture_dir" in configuration.py according to the specific situation.
2. Run test_on_single_image.py to test single picture.

Convert model to TensorFlow Lite format

1. Change the "TFLite_model_dir" in configuration.py according to the specific situation.
2. Run convert_to_tflite.py to generate TensorFlow Lite model.

References

1. YOLO_v3 paper: https://pjreddie.com/media/files/papers/YOLOv3.pdf or https://arxiv.org/abs/1804.02767
2. Keras implementation of YOLOV3: https://github.com/qqwweee/keras-yolo3
3. blog 1, blog 2, blog 3, blog 4, blog 5, blog 6, blog 7
4. 李金洪. 深度学习之TensorFlow工程化项目实战[M]. 北京: 电子工业出版社, 2019: 343-375
5. https://zhuanlan.zhihu.com/p/49556105