Advanced transfer learning

Time arrangement

2022 year 05 month 27 Japan — 2022 year 05 month 30 Japan   

No.1 The first day

  One 、​ ​ machine learning ​​ Introduction and classical machine learning algorithms

What is machine learning ？

Machine learning framework and basic components

The training steps of machine learning

Classification of machine learning problems

Classic machine learning algorithm

Chapter goal ： Machine learning is ​ ​ Artificial intelligence ​​ It's one of the most important technologies , Understand the principle of machine learning in detail 、 Mechanism and method , Lay a solid foundation for deep learning and transfer learning .

  Two 、 Introduction to deep learning and classic network structure

Introduction to neural networks

Introduction to neural network components

Neural network training method

Convolution neural network introduction

Introduction to classic network structure

Chapter goal ： In depth understanding of the composition of neural networks 、 Train and achieve , Master key concepts such as depth spatial feature distribution , Lay a knowledge foundation for deep transfer learning

3、 ... and 、 Transfer learning Basics

Introduction to transfer learning

Sample based transfer learning

Feature based transfer learning

Transfer learning based on classifier adaptation

Chapter goal ： Master the thought and basic form of transfer learning , Understand the basic methods of traditional transfer learning , Compare the advantages and disadvantages of various methods , Master the applicable scope of transfer learning .

Four 、 Introduction to deep transfer learning

Overview of deep transfer learning

Deep transfer learning based on distance function

Deep transfer learning based on confrontation network

Introduction to deep heterogeneous transfer learning methods

Introduction to deep domain generalization learning

Chapter goal ： Master the ideas and modules of deep transfer learning , Learn various methods of deep transfer learning , Compare the advantages and disadvantages of various methods , Master the scope of application of deep transfer learning .

No.2 the second day

5、 ... and 、 Introduction to frontier methods of transfer learning  

Deep migration network structure design

Design of objective function for deep transfer learning

Transfer learning in a new scenario

Chapter goal ： Master the network structure design of deep transfer learning 、 Frontier methods of objective function design , Understand that transfer learning is PDA、Source-Free DA Application on .

6、 ... and 、 Transfer learning frontier applications  

Application of transfer learning in semantic segmentation

Application of transfer learning in target detection

Application of transfer learning in pedestrian recognition

Image and video style migration

Chapter goal ： Master the application of deep transfer learning in semantic segmentation 、 object detection 、 Application in tasks such as pedestrian re recognition , Learning images / Video style migration method , Understand the application of style transfer in real life .

7、 ... and 、 Small sample learning 、Transformer And other cutting-edge methods and applications   

Introduction to the concept and basic methods of small sample learning

Small sample learning application

Transformer Introduction to concepts and basic methods

Transformer Application in the field of image

Chapter goal ： Master small sample learning 、Transformer And other cutting-edge methods and basic ideas , Learn about small sample learning 、Transformer And other applications in actual scenarios .

No.3 On the third day

8、 ... and 、 Construction of practical operation environment for experimental operation  

Hardware preparation ：GPU memory 11GB above

Software preparation ：Linux​ ​ operating system ​​（Ubuntu16.04 above ）, Video card driver installation （512.54）,CUDA Toolkit（10.1） and cuDNN Acceleration Library （7.6.4）,VS Code Editor installation ,Jupyter Notebook

Programming languages and frameworks ：Python3.8.5、torch==1..07、torchvision==0.8.2、mmcv-full==1.3.7、opencv-python==4.4.0、matplotlib==3.4.2、numpy==1.19.2、Pillow==8.3.1、scikit-learn==1.0.2

Data set preparation ：Office-31、IRVI、GTA5、Cityscapes、Foggy cityscapes etc.

notes ： The hardware preparation is provided by the sponsor

Nine 、 Deep transfer learning practice of experimental practice  

master PyTorch Basic principles and programming ideas in .

Understand under a new scene or data set , When and how to transfer learning .

utilize PyTorch Load data 、 Build a model 、 Train the network and fine tune the network .

Given migration scenario , utilize daib Library and generation countermeasure technology independently complete the domain adaptation in image classification .

Visualization of migration effects , Using machine learning library scikit-learn Medium t-SNE Visualize the migrated high-dimensional data .

Ten 、 Experimental practice of image and video style transfer  

Master the style migration technology based on generative confrontation network .

Images / Construction of video style migration network , Focus on mastering ​ ​ Encoder ​​ And the internal logic of the decoder and the application of different loss functions .

practice ​ ​ infrared ​​ Style transfer from video to visible video .

11、 ... and 、 The practice of cross domain semantic segmentation in autopilot   

Master the development status and representative work of semantic segmentation , Such as FCN,DeepLab Series etc. .

Understand the commonly used semantic segmentation evaluation indicators （PA、mPA、mIoU、FWIoU） And common data sets （PASCAL VOC2012,ADE20K、BDD100K、Cityscapes、GTA5、Dark Zurich）.

Semantic segmentation toolbox MMSegmentaion Understanding and use of .

Designing a segmentation model can start from ​ ​ Simulation ​​ The data obtained in the environment is migrated to the data generated in the real scene .

Twelve 、 Target detection practice of experimental operation  

Master the basic target detection algorithm in the classical target detection framework , Such as R-CNN Two stage detection model and YOLO A series of single-stage detection models .

Master the evaluation indicators of target detection model （IOU and mAP）、 Standard evaluation data set （Pascal VOC,MS COCO and Cityscapes） And some training skills in the detection model , Such as data enhancement 、 Multiscale training /​ ​ test ​​、 Prediction box fine tuning / laws and regulations governing balloting 、 Online hard case mining 、 Softening does not greatly inhibit 、RoI Alignment and integration .

Practice based on Transformer Of ​ ​ End to end ​​ Construction of target detection framework , And based on the new data set CNN Compare the performance of the network .

【 WeChat ID ：AI_longteng, WeChat official account ：​ ​ artificial ​​ Intelligent technology and consulting 】 Welcome to add consultation ！