paddle 1-advanced

Table of contents

Why You Should Master the Advanced Content of Deep Learning

What weapons are included in the premium content

1. Model resources

2. Design thinking and secondary research and development

3. Industrial deployment

4. Flying paddle full-process R&D tools

5. Industry application and project case

Overview of usage scenarios of Flying Paddle open source components

Frameworks and full-process tools

1. Model training component

2. Model deployment components

3. Other auxiliary tools for the whole R&D process

model resource

---

Why You Should Master the Advanced Content of Deep Learning

In the previous chapters, we first learned the basics of neural network models and how to write deep learning models using flying paddles , and then learned how to implement models for computer vision, natural language processing, and recommendation systems .

But it is not easy to win in the artificial intelligence battlefield. We will also face the following challenges:

• It is necessary to propose a modeling solution for the business scenario;
• Exploring which of the many complex models is more effective;
• Explore deploying models to various types of hardware.

What weapons are included in the premium content

1. Model resources

Nowadays, deep learning applications have been implemented in many fields, and the first choice of modeling for developers is often not to write by themselves, but to use off-the-shelf models, or to optimize on off-the-shelf models. On the one hand, this will greatly reduce the workload of R&D personnel. On the other hand, the ready-made models are generally refined in terms of accuracy and performance, and the effect is better.

So, where to find ready-made model resources?

Paddle provides three types of model resources:

• Pre-training model tool (PaddleHub);
• Scenario-specific development kits covering more than a dozen tasks in fields such as computer vision, natural language processing, speech, and recommendation systems (such as paddle image segmentation kit PaddleSeg, paddle semantic understanding kit ERNIE, etc.);
• Open source model library (Paddle Models)

2. Design thinking and secondary research and development

When readers challenge some of the latest models, in a few cases, they will encounter situations where the operator paddle required by the model has not been implemented. This chapter will tell you how to add custom operators to the flying paddle frame, and by describing the implementation principles of dynamic and static graphs, let you have a deeper understanding of the design ideas of the flying paddle frame .

3. Industrial deployment

Different from the scientific research and teaching of the model, the model in the industrial application needs to be deployed on a very rich hardware environment , such as embedding the model in a business system written in C++ language, using the model as a separate web service, or putting the model in Wait on camera. This chapter introduces Paddle Inference, Paddle Serving, and Paddle Lite to meet the above demand scenarios, and introduces the model compression tool Paddle Slim, which allows the model to run faster on limited hardware.

4. Flying paddle full-process R&D tools

Flying Paddle provides you with so many model resources and tool components, how to connect these components and develop a visual modeling tool suitable for the industry of the readers? This chapter will take an officially produced Demo as a case to show readers the full-process services that PaddleX can provide to users .

5. Industry application and project case

The authoritative consulting organization iResearch predicts that the growth rate of the industrial scale of artificial intelligence will reach 40% in the next ten years. Artificial intelligence is also a strategic focus of the country's new infrastructure. The State Council also has high growth expectations for the development plan of AI applications. Although artificial intelligence empowers all walks of life to flourish, there are still friends in traditional industries who have doubts:

"The industry I work in is too traditional, artificial intelligence is useless?"

This chapter takes the energy industry as an example to analyze the aspects of a typical electric power company that can be optimized with artificial intelligence, and demonstrates a real project based on flying paddle modeling.

Artificial intelligence and deep learning are practical sciences, and these weapons cannot be used on the battlefield without actual hands-on practice. Therefore, this chapter carefully selects 6 homework competitions, so that everyone can really master these weapons in interesting case practice, and compete with the top deep learning experts.

Overview of usage scenarios of Flying Paddle open source components

Next, let's review the full set of weapons provided by the flying paddle through an overview map. Based on Baidu's years of deep learning technology research and business applications, Flying Paddle integrates deep learning core framework, basic model library, end-to-end development kit, tool components and service platform, providing users with a variety of supporting service products , to help the application of deep learning technology. Paddle supports two development and deployment modes: local and cloud , and users can choose flexibly according to business needs.

Figure 1: Overview of usage scenarios of PaddlePaddle components

 

The upper part of the overview diagram is full-process tools from development, training to deployment, and the lower part is model resources such as pre-trained models, development kits in various fields, and model libraries.

Frameworks and full-process tools

In addition to providing the basic framework for model development, Paddle also introduces a series of tool components to support the entire process of deep learning models from training to deployment.

1. Model training component

Flying Paddle provides a distributed training framework FleetAPI , and also provides a tool to open cloud task submission PaddleCloud . At the same time, the flying paddle also supports multi-task training, and the multi-task learning framework PALM can be used .

2. Model deployment components

Paddle provides a wealth of support solutions for different hardware environments:

• Paddle Inference : Paddle's native inference library, used for server-side model deployment , supports languages ​​such as Python, C, C++, and Go, and is the first choice for integrating models into business systems.
• Paddle Serving : Paddle Serving Deployment Framework, used for cloud service deployment, can use the model as a separate web service.
• Paddle Lite : Paddle lightweight inference engine, used for the deployment of Mobile, IoT and other scenarios, with a wide range of hardware support.
• Paddle.js : Use JavaScript (Web) language to deploy models, which is used to quickly deploy models in environments such as browsers and applets.
• PaddleSlim : Model compression tool for smaller models and faster execution performance.
• X2Paddle : Flying paddle model conversion tool, convert other frame models into Paddle models, and the above 5 tools can be used conveniently after the converted format.

3. Other auxiliary tools for the whole R&D process

• AutoDL : Flying Paddle automated deep learning tool, aimed at automatic network structure design , the image classification network designed by the open source AutoDL has a correct rate of 98% in the CIFAR10 data set, and the effect is better than the network designed by 10 types of human experts that have been published so far. Industry leading position.
• VisualDL: A visual analysis tool for flying paddles , which presents key model information such as training parameter change trends, model structure, data samples, high-dimensional data distribution, and precision-recall curves in rich charts. Help users understand the deep learning model training process and model structure clearly and intuitively, and then realize efficient model tuning, and share the algorithm training process and results.
• PaddleFL: Paddle federated learning framework , researchers can easily use PaddleFL to copy and compare different federated learning algorithms, conveniently implement large-scale distributed cluster deployment, and provide rich horizontal and vertical federated learning strategies and their applications in computer vision , natural language processing, recommendation algorithms and other fields. In addition, relying on PaddlePaddle's large-scale distributed training and Kubernetes' flexible scheduling capabilities for training tasks, PaddleFL can be easily deployed based on full-stack open source software.

model resource

Paddle provides a rich end-to-end development kit, pre-trained models and model libraries.

PaddleHub : Paddle pre-training model application tool, covering more than 200 high-quality pre-training models in the four major fields of text, image, video, and voice. Developers can easily combine actual business scenarios, select high-quality pre-trained models and cooperate with Fine-tune API to quickly complete model verification and application deployment. It is suitable for various business scenarios such as individual developer learning, enterprise POC rapid verification, participation in AI competitions, teaching and scientific research, etc.

Development Kit : Provides a complete set of R&D tools for specific application scenarios. For example, in image detection scenarios, it not only provides pre-trained models, but also provides tools such as data enhancement. The development kit also covers mainstream areas such as computer vision, natural language processing, speech, recommendation, and even graph neural networks and reinforcement learning. Different from PaddleHub, the development kit can provide a state-of-the-art (State Of The Art) implementation. There was a domestic team that used the paddle development kit to win the grand prize in the international modeling competition. Some typical development kits include:

• PaddleClas: Paddle image classification suite , the purpose is to provide a convenient and easy-to-use image classification task pre-training model and tool set for industry and academia, open up the whole process of model development, training, compression, deployment, and assist other high-level vision task networking And improve the effect of the model to help developers train better image classification models and implement applications.
• PaddleDetection : Paddle target detection kit, designed to help developers complete the whole process of detection model training, accuracy speed optimization and deployment faster and better. PaddleDetection implements a variety of mainstream target detection algorithms with a modular design, and provides a wealth of modules such as data enhancement, network components, and loss functions, and integrates model compression and cross-platform high-performance deployment capabilities. At present, the projects based on Paddle Detection have been completed and implemented, involving industrial quality inspection, remote sensing image detection, unmanned inspection and other fields.
• PaddleSeg : Flying paddle image segmentation suite , covering a large number of high-quality segmentation models in different directions such as DeepLabv3+/OCRNet/BiseNetv2/Fast-SCNN with high precision and lightweight. Through the modular design, it provides two application methods, such as configuration driver and API call, to help developers complete the whole process of image segmentation application from training to deployment more conveniently.
• PaddleOCR: Paddle text recognition kit , aiming to create a rich, leading, and practical OCR tool library, open-sourced PPOCR-based practical ultra-lightweight Chinese and English OCR models, general Chinese and English OCR models, and German, French, Japanese, Korean, etc. Multilingual OCR model. It also provides the above-mentioned model training methods and multiple prediction deployment methods. At the same time, the open source text style data synthesis tool Style-Text and the semi-automatic text image annotation tool PPOCRLable.
• PaddleGAN : Paddle image generation development kit, integrating SOTA algorithms such as style transfer, super-resolution, animation generation, image coloring, face attribute editing, makeup transfer, and pre-training models. And the modular design is convenient for developers to carry out secondary research and development, or directly use the pre-trained model for application.
• PaddleX : Paddle scene application development kit, which supports developers to quickly realize deep learning algorithm development and industrial deployment in the form of low code. Provides two development modes: minimalist Python API and visual interface Demo, which can be installed with one click. It provides a complete deployment solution for general-purpose hardware such as CPU (OpenVINO), GPU, and Raspberry Pi, and can quickly complete integration and redevelopment through the RESTful API. Developers can complete the production and deployment of the whole process model without using different packages. The visual reasoning interface and rich industry cases provide developers with the best practices of the whole process of flying paddle development.
• PLSC : Flying Paddle Mass Classification Kit, which provides users with a full-process solution for large-scale classification tasks from training to deployment. Provides a simple and easy-to-use high-level API, which can realize the training of tens of millions of classification neural networks with a few lines of code, and provides the ability to quickly deploy models.
• ERNIE : Realization of knowledge-enhanced semantic understanding framework based on continuous learning, built-in industry-leading series of ERNIE pre-training models, capable of supporting fine-tuning of various NLP algorithm tasks, including Fast-inference API that guarantees extremely fast reasoning, flexible deployment of ERNIE Service and The lightweight solution ERNIE Slim, what you see is what you get during the training process, supports dynamic debugging, and facilitates secondary development.
• ElasticCTR : Paddle Elastic Computing Recommendation Suite, which can realize distributed training CTR estimation tasks and online personalized recommendation services based on Paddle Serving. The Paddle Serving service deployment framework has good ease of use, flexibility and high performance, and can provide end-to-end CTR training and deployment solutions. ElasticCTR has the characteristics of industrial practice foundation, flexible scheduling capability, high performance and industrial-grade deployment.
• Parakeet : Paddle Speech Synthesis Kit, which provides flexible, efficient, and advanced text-to-speech synthesis tools to help developers complete the development and application of speech synthesis models more conveniently and efficiently.
• PGL : Flying paddle graph learning framework, the industry's first industrial-grade graph learning framework that proposes a general message parallel transmission mechanism and supports tens of billions of large-scale graphs. PGL is a new upgrade based on the dynamic graph of the flying paddle, which greatly improves the ease of use. It natively supports heterogeneous graphs, supports distributed graph storage and distributed learning algorithms, and covers 30+ graph learning models, including the graph semantic understanding model ERNIESage, etc. After a large number of real industrial application verification, it can flexibly and efficiently build cutting-edge large-scale graph learning algorithms.
• PARL : Flying Paddle Deep Reinforcement Learning Framework, won three consecutive championships in the Reinforcement Learning Challenge in 2018, 2019, and 2020. It has the characteristics of high flexibility, scalability and high performance. Implemented more than ten examples of mainstream reinforcement learning algorithms, covering reinforcement learning algorithm support in different fields from single agent to multi-agent, discrete decision-making to continuous control. Realize high-performance parallelism of thousands of CPUs and GPUs based on the GRPC mechanism.
• Paddle Quantum : measuring paddle, flying paddle quantum machine learning framework, providing cutting-edge application tool sets such as quantum optimization and quantum chemistry, common quantum circuit models, and rich quantum machine learning cases, helping developers to easily build quantum neural networks and develop quantum AI applications.
• PaddleHelix : The flying paddle propeller biocomputing framework opens up AI capabilities that empower vaccine design, new drug development, and precision medicine. In terms of vaccine design, PaddleHelix’s LinearRNA series algorithms have improved the efficiency of RNA folding by hundreds or thousands of times compared with traditional methods; in the development of new drugs, PaddleHelix provides molecular representations based on large-scale data pre-training to help predict molecular properties , drug screening, drug design and other fields; in precision medicine, PaddleHelix provides a high-performance model that uses omics information to precisely locate drugs and improve the cure rate.

Model library : It contains a wealth of open source model codes in various fields. Not only can the model be run directly, but the original model code can also be modified according to the needs of the application scenario to obtain a new model implementation.

Comparing the three types of model resources, PaddleHub is the easiest to use, and the model library is the most customizable and covers the most extensive fields. Readers can refer to the order of "PaddleHub->Development kits in various fields->Model library" to find the required model resources, and optimize according to business needs on this basis to achieve twice the result with half the effort.