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。
For deep learning tasks, Tegu uses the Network_Model class to manage training tasks, and uses the Network_Dataloader class to manage data sets.
- 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.- First import the API library of the SSD300 image detection model.
from Network.SSD300.API import SSD_Model, SSD_DataLoader- 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)- Set the Dataset for SSD_Model.
m.set_dataset(d)- 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)- After training, use Tegu-core to predict image. First create SSD_Model .
m = SSD_Model(class_count=2)- 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.
- Install Python3 and pip3
- Install CUDA and cuDNN
- If you would like to use the Facial Recognition feature, please compile and install OpenCV. Please make sure the DNN module is installed.
- Use the following command to install requirements:
pip3 install -r requirements.txt- For the usage of Tegu-core API, see
ExampleFolder.
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.
- Network/: Neural Networks, called by scripts in GUI/
- Example/: Samples for how to use Tegu-core.
- Util/: Utility used in the project
- 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
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.
- 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)
Project initialized by Generalized Intelligence Distributed under the BSD 3-Clause license. See LICENSE for more information.
Please follow CONTRIBUTING.md
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。
- 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.- 首先 import SSD300 图像检测模型的 API 库。
from Network.SSD300.API import SSD_Model, SSD_DataLoader- 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)- 为 SSD_Model 设置 Dataset。
m.set_dataset(d)- 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)- 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)- 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。
- 首先请安装 Python3 以及 pip3。
- 请安装 CUDA 和 cuDNN。
- 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。
- 使用如下命令安装所需的 Python 框架:
pip3 install -r requirements.txt- 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。
- 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。
- 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
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
- 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）
泛化智能 Generalized Intelligence 出品。 本项目通过 BSD 3-Clause 协议发布，详情见 LICENSE。
请参阅 CONTRIBUTING.md