yolo-tf2：keras和tensorflow 2.4中的yolo（所有版本）实现

[![Contributors][contributors-shield]][contributors-url] [![Forks][forks-shield]][forks-url] [![Stargazers][stars-shield]][stars-url] [![Issues][issues-shield]][issues-url] [![MIT License][license-shield]][license-url] <p> <a href="https://github.com/emadboctorx/yolov3-keras-tf2/"> </a> <h3 align="center">Yolo (all versions) Real Time Object Detector in tensorflow 2.4</h3> . <a href="https://github.com/emadboctorx/yolov3-keras-tf2/issues">Report Bug</a> · <a href="https://github.com/emadboctorx/yolov3-keras-tf2/issues">Request Feature</a> </p> <!-- TODO --> ## **TODO** * [ ] Transfer learning * [x] YoloV4 configuration * [x] YoloV4 training * [ ] YoloV4 loss function adjustments. * [ ] Live plot losses * [x] Command line options * [x] YoloV3 tiny * [ ] Rasberry Pi support <!-- TABLE OF CONTENTS --> ## **Table of Contents** * [Getting Started](#getting-started) * [Installation](#installation) * [Description](#description) * [Updates](#updates) * [Features](#features) * [(new) Command line options](#new-command-line-options) * [General](#general-options) * [Training](#training-options) * [Evaluation](#evaluation-options) * [Detection](#detection-options) * [DarkNet models loaded directly from .cfg files](#darknet-models-loaded-directly-from-cfg-files) * [YoloV4 support](#yolov4-support) * [tensorflow-2.4--keras-functional-api](#tensorflow-24--keras-functional-api) * [cpu-gpu support](#cpu--gpu-support) * [Random weights and DarkNet weights support](#random-weights-and-darknet-weights-support) * [csv-xml annotation parsers.](#csv-xml-annotation-parsers) * [Anchor generator.](#anchor-generation) * [matplotlib visualization of all stages.](#matplotlib-visualization-of-all-stages) * [tf.data input pipeline.](#tfdata-input-pipeline) * [imgaug augmentation pipeline(customizable).](#imgaug-augmentation-pipelinecustomizable) * [Logging](#logging) * [All-in-1 custom trainer.](#all-in-1-custom-trainer-class) * [Stop and resume training support.](#stop-and-resume-training) * [Fully vectorized mAP evaluation.](#fully-vectorized-map-evaluation) * [labelpix support.](#labelpix-support) * [Photo & video detection](#photo--video-detection) * [Usage](#usage) * [Training](#training) * [Terminal command](#training-command-line-equivalent) * [Augmentation](#augmentation) * [Evaluation](#evaluation) * [Terminal command](#evaluation-command-line-equivalent) * [Detection](#detection) * [Terminal command (photos)](#detection-photo-command-line-equivalent) * [Terminal command (video)](#detection-video-command-line-equivalent) * [Contributing](#contributing) * [Issues](#issue-policy) * [What will be addressed](#relevant-issues) * [What will not be adressed](#irrelevant-issues) * [License](#license) * [Show your support](#show-your-support) * [Contact](#contact)  <!-- GETTING STARTED --> ## **Getting started** ### **Installation** 1. **Clone the repo** ```sh git clone https://github.com/emadboctorx/yolo-tf2 ``` 2. **Install** **Note:** If you have a cuda-compatible GPU, uncomment `tensorflow-gpu` in `requirements.txt` ```sh cd yolo-tf2 pip install . ``` 3. **Verify installation** ```sh yolotf2 ``` **OUT:** Yolo-tf2 1.5 Usage: yolotf2 <command> [options] [args] Available commands: train Create new or use existing dataset and train a model evaluate Evaluate a trained model detect Detect a folder of images or a video <!-- DESCRIPTION --> ## **Description** yolo-tf2 was initially an implementation of [yolov3](https://pjreddie.com/darknet/yolo/) (you only look once)(training & inference) and support for all yolo versions was added in V1.1 Yolo's original repo is [here](https://github.com/AlexeyAB/darknet) (written in C/C++/Cu). Yolo is a state-of-the-art, real-time object detection system that is extremely fast and accurate. There are many implementations that support tensorflow, only a few that support tensorflow v2 and as I did not find versions that suit my needs so, I decided to create this version which is very flexible and customizable. It requires python 3.8+, is not platform specific and is MIT licensed. <!-- Updates --> ## **Updates** ### [1.5] - 2021-01-21 - Fix bug that creates out of image bounding boxes after augmentation. - Fix bug with display commands - Update dependencies to most recent versions. ### [1.4] - 2020-11-30 - Fix a bug that draws extra irrelevant boxes over photos for V4 configuration - Fix a bug that causes shape incompatibility issues - Remove extra required parameters `image_width`, `image_height` which are currently measured during runtime. - Fix a bug that prevents output from saving due to path mishandling - Unify all IO operations to go through `yolo_tf2.utils.common.get_abs_path()` - All commands are currently supported from any working directory, so users shouldn't worry about creating folders and matching names, which is handled automatically. - Upgrade all dependencies to latest versions ### [1.3] - 2020-10-20 - Add command line full support for training, evaluation and detection ### [1.2] - 2020-10-18 - Add setup.py, direct installation through python3 setup.py install - Restructure package folders ### [1.1] - 2020-06-02 - Add yolov4 support ### [1.0] - 2020-05-29 - Models are loaded directly from DarkNet .cfg files - YoloV4 is currently supported(inference only, no training) - Mish activation function(YoloV4) <!-- FEATURES --> ## **Features** ## **(new) Command line options** ### **General options** | flags | help | required | default | |:---------------------|:------------------------------------------------------------------|:-----------|:--------------| | --input-shape | Input shape ex: (416, 416, 3) | - | (416, 416, 3) | | --classes | Path to classes .txt file | True | - | | --model-cfg | Yolo DarkNet configuration .cfg file | True | - | | --max-boxes | Maximum boxes per image | - | 100 | | --iou-threshold | IOU(intersection over union threshold) | - | 0.5 | | --score-threshold | Confidence score threshold | - | 0.5 | | --workers | Concurrent tasks(in areas where that is possible) | - | 16 | | --process-batch-size | Batch size of operations that needs batching (excluding training) | - | 32 | ### **Training options** | flags | help | required | default | |:----------------------|:-------------------------------------------------------------|:-----------|:----------| | --weights | Path to trained weights .tf or .weights file | - | - | | --epochs | Number of training epochs | - | 100 | | --batch-size | Training batch size | - | 8 | | --learning-rate | Training learning rate | - | 0.001 | | --dataset-name | Name of the checkpoint | True | - | | --test-size | test dataset relative size (a value between 0 and 1) | - | 0.1 | | --evaluate | If True, evaluation will be conducted after training | - | - | | --merge-evaluation | If False, evaluate training an