克隆/下载
贡献代码
同步代码
取消
提示: 由于 Git 不支持空文件夾,创建文件夹后会生成空的 .keep 文件
Loading...
README
MIT

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
  1. 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.
  2. 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
  1. 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.
  2. 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:
    original picture 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
MIT License Copyright (c) 2019 calmisential Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

简介

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.). 展开 收起
MIT
取消

发行版

暂无发行版

贡献者

全部

近期动态

不能加载更多了
马建仓 AI 助手
尝试更多
代码解读
代码找茬
代码优化