Skip to content

Latest commit

 

History

History
229 lines (182 loc) · 9.24 KB

README.md

File metadata and controls

229 lines (182 loc) · 9.24 KB
Raw

Tegu Logo

Tegu

Join the chat at https://gitter.im/Tegutalk/community

Use Machine Learning Right Now In Your Project

即刻在你的项目中使用机器学习

Tegu-core is the core component of Tegu, which provides an encapsulation of some state-of-the-art deep learning models of computer vision, and its APIs are called by the Tegu GUI components. You can use Tegu-core to provide some deep learning functions in your own Python projects with only a few codes.

Tegu-core 是 Tegu 的核心组件，提供了对一些当前最佳实践（State of the Art）的计算机视觉深度学习模型的封装，与Tegu GUI配合使用。若您需要在Python项目中引入深度学习功能，只需要很少的代码，您就可以将Tegu-core与您的项目一起配合使用。

30s to start training car detection model

For deep learning tasks, Tegu uses the Network_Model class to manage training tasks, and uses the Network_Dataloader class to manage data sets.

  1. Download car detection dataset and unzip, the dataset contains:
|CarDataset/
    |--trainset/  #Images to train the model.
        |--xxx.jpg
        ...
    |--testset/  #Images to test your model.
        |--xxx.jpg
        ...
    |car_annotation.serval  #The annotation file contains the annotations for the images in the trainset directory.
  1. First import the API library of the SSD300 image detection model.
from Network.SSD300.API import SSD_Model, SSD_DataLoader
  1. Create SSD_Model and SSD_Dataloader instances.
m = SSD_Model(class_count=2, base_lr=0.0004)
d = SSD_DataLoader(anno_path=r"dataset/save/path/car_annotation.serval", data_path=r"dataset/save/path/trainset",batch_size=16)
  1. Set the Dataset for SSD_Model.
m.set_dataset(d)
  1. Start training, set the epoch to 100 rounds, and save every 20 rounds.
epoch = 100
for i in range(epoch):
    train_info = m.train()
    print(train_info)   #{'loss':[3.3914230046448886], 'val_loss':[3.8560243606567384]}
    if (i+1)%20==0 or i+1==epoch:
        save_path = "ssd_model{}.h5".format(str(i).zfill(3))
        m.model.save(save_path)
  1. After training, use Tegu-core to predict image. First create SSD_Model .
m = SSD_Model(class_count=2)
  1. Start predicting image using the model trained in the first few step.
m.predict(img_path=r"dataset/save/path/trainset/xxx.jpg", model_path=r"ssd_modelXXX.h5", anno_path=r"dataset/save/path/car_annotation.serval")
#[[label:int, class_name:str, score:double, (xmin, ymin), (xmax, ymax)]]

For more usage, see Example.

Quick Start

  1. Install Python3 and pip3
  2. Install CUDA and cuDNN
  3. If you would like to use the Facial Recognition feature, please compile and install OpenCV. Please make sure the DNN module is installed.
  4. Use the following command to install requirements:
pip3 install -r requirements.txt
  1. For the usage of Tegu-core API, see Example Folder.

For Image Recognition and Video Classification, we have developed a set of tools to process and clean up the datasets. You may use Tegu Image Annotation and Tegu Video Annotation to process your dataset.

File Structure

  • Network/: Neural Networks, called by scripts in GUI/
  • Example/: Samples for how to use Tegu-core.
  • Util/: Utility used in the project

Existing Features

  • Video Classification (Long Video Classification)
  • Image Recognition
  • Facial Recognition (build feature library, and recognize)
  • License Plate Recognition

HTTP API is available for Image Recognition, Facial Recognition and License Plate Recognition

Relevant Projects

For Image Recognition and Video Classification, we have developed a set of tools to process and clean up the datasets. You may use Tegu Image Annotation and Tegu Video Annotation to process your dataset.

Existing Neural Networks

TODO

  • Optimization of interprocess communication
  • Adding common neural networks, such as passenger detection, vehicle detection, and cat/dog detection
  • Abnormality Detection in Restricted Zone (Training Free)
  • Hand Gesture Recognition, Pose Detection based on Feature Point Paring, Facial Feature Point Detection (Training Free)
  • Image Segmentation
  • General OCR
  • Video Tracking (Work-In-Progress)

Meta

Project initialized by Generalized Intelligence Distributed under the BSD 3-Clause license. See LICENSE for more information.

Contribute

Please follow CONTRIBUTING.md

30秒开始训练车辆检测模型

对于深度学习任务，Tegu 使用 网络名_Model 类来管理训练任务，使用 网络名_Dataloader 类来管理数据集。

  1. 下载 车辆检测数据集 并解压缩。
|CarDataset/
    |--trainset/  #Images to train the model.
        |--xxx.jpg
        ...
    |--testset/  #Images to test your model.
        |--xxx.jpg
        ...
    |car_annotation.serval  #The annotation file contains the annotations for the images in the trainset directory.
  1. 首先 import SSD300 图像检测模型的 API 库。
from Network.SSD300.API import SSD_Model, SSD_DataLoader
  1. 创建 SSD_Model 和 SSD_Dataloader 实例。
m = SSD_Model(class_count=2, base_lr=0.0004)
d = SSD_DataLoader(anno_path=r"dataset/save/path/car_annotation.serval", data_path=r"dataset/save/path/trainset",batch_size=16)
  1. 为 SSD_Model 设置 Dataset。
m.set_dataset(d)
  1. 开始训练，设置 epoch 为 100 轮，并且每 20 轮保存一次。
epoch = 100
for i in range(epoch):
    train_info = m.train()
    print(train_info)   #{'loss':[3.3914230046448886],'val_loss':[3.8560243606567384]}
    if (i+1)%20==0 or i+1==epoch:
        save_path = "ssd_model{}.h5".format(str(i).zfill(3))
        m.model.save(save_path)
  1. 完成训练以后，使用 Tegu-core 预测图片。首先创建 SSD_Model 实例。
m = SSD_Model(class_count=2)
  1. 使用前几步训练出的模型预测图片
m.predict(img_path=r"dataset/save/path/trainset/xxx.jpg", model_path=r"ssd_modelXXX.h5", anno_path=r"dataset/save/path/car_annotation.serval")
#[[label:int, class_name:str, score:double, (xmin, ymin), (xmax, ymax)]]

查看更多应用，请参见 Example。

快速开始

  1. 首先请安装 Python3 以及 pip3。
  2. 请安装 CUDA 和 cuDNN。
  3. 如果需要使用人脸检测功能， 请编译安装 OpenCV, 并确保安装其 DNN 模块。
  4. 使用如下命令安装所需的 Python 框架:
pip3 install -r requirements.txt
  1. Tegu-core 模型API的用法，可以参照Example文件夹中的示例代码。

对于图像检测和视频分类任务，我们使用自己的格式处理数据集，您可以使用 Tegu 图像标注软件，和 Tegu 视频标注软件 来制作数据集。

文件结构

  • Network/: 项目所用到的网络部分，可供外部的Python文件调用。
  • Example/：项目的示例代码。
  • Util/: 项目使用的一些工具脚本。

现有功能

  • 视频分类（长视频分类）
  • 图像检测
  • 人脸识别（建立特征库，并识别人脸）
  • 车牌识别

相关项目

对于图像检测和视频分类的数据处理及标注，我们提供了 Tegu 图像标注软件，和 Tegu 视频标注软件

已有网络

待办事项

  • 修改进程通信模块
  • 建立行人，车辆，猫狗等常见检测模型
  • 禁入区等功能（免训练）
  • 手势识别，人体姿态特征点检测，人脸特征点检测（免训练）
  • 图像分割，要有相应的常见模型
  • 通用OCR。
  • 视频追踪（Work-In-Progress）

版权信息

泛化智能 Generalized Intelligence 出品。 本项目通过 BSD 3-Clause 协议发布，详情见 LICENSE。

贡献

请参阅 CONTRIBUTING.md