Among various AI conferences, Nvidia’s twice-a-year GTC conference is quite famous. This is of course not only because NVIDIA's GPU products play a pivotal role in the AI industry, but also because of the technical hotspots, industry insights and guidance and training brought by the GTC conference for many years, allowing developers to get practical help.
And after a year away from Suzhou last year, NVIDIA GTC 2020 has finally returned to us. This time GTC 2020 adopts the form of online live broadcast, but the content will not be changed in the slightest because of the change of form. The same splendor, the same dry goods and technology, Baidu Fei Pao , which contributed wonderful content last year , came back this year with four speeches.
Flying Paddle is based on Baidu's many years of deep learning technology research and business applications. It integrates the deep learning core framework, basic model library, end-to-end development kit, tool components and service platform. It is fully open source, technologically advanced, and fully functional. industry-level deep learning platform. Last year at GTC, PaddlePaddle upheld the concept of parallel technical practice and theoretical knowledge, and brought a fruitful open technical class to developers .
This year, the four speeches delivered by Paddle at GTC 2020 are also full of sincerity , covering Paddle inference engine, large-scale distributed training , Easy DL zero threshold high-precision AI service customization , machine learning development environment BML CodeLab , etc. Lots of content. Four speeches will take turns at GTC 2020, let's take a look!
Lecture 1. Performance optimization of flying paddle inference engine
At the GTC 2020 "Deep Learning Platform and Application" sub-forum at 17:00 pm on December 15th, Shang Zhizhou, a senior R&D engineer of Baidu, will give us the first speech on the performance optimization of the flying paddle inference engine.
This speech will introduce Paddle I nference , the native inference engine of Paddle Inference , and the optimization work in this engine for GPU inference. The Flying Paddle inference engine is an important foundation for the Flying Paddle model inference and deployment, and has been fully verified in Baidu's internal core business lines and AI services delivered by many ToBs. We first introduce the general optimization measures made by the flying paddle inference engine for GPU inference, such as OP fusion, video memory reuse, TensorRT integration, mixed-precision inference, etc. At the same time, we will take the inference optimization of the ERNIE model as an example to illustrate the Inference optimization methods and effects for specific models.
The content of this speech will focus on the optimization and deployment of deep learning inference technology. At the same time as the event, the lecturer will interact with the participants online to answer questions for the audience.
Lecture 2. Progress and application of large-scale distributed training
In the sub-forum of "Cloud Computing and Consumer Internet" on December 17, Dong Daxiang, chief R&D architect of Baidu, will reveal the secrets of the large-scale distributed training progress and application of the Paddle Framework. The theme of this speech will focus on the large-scale training and application of the paddle framework , and introduce the function, performance, and practice of large-scale distributed training in the industry. The current version of Paddle Framework 2.0RC has been released and users are invited to try it out.
Speech 3. Customize high-precision AI services with EasyDL zero threshold
In the sub-forum of "Deep Learning Platform and Application" on December 15, Baidu senior R&D engineer Hu Mingren, Baidu senior R&D engineer Liu Jie, and the two teachers will work together to talk about the creation of high-precision AI services .
This speech will introduce EasyDL products and core technical features from the background of actual needs ; introduce various implementations of EasyDL effect optimization and the optimization brought by the combination of NVIDIA GPU and EasyDL ; the speech will also analyze the deployment of EasyDL models based on GPU at the end of the speech and practical.
This course will pay more attention to the fine-grained content. Based on deep learning technology use cases and successful cases, it will explain to the participants how to realize optimization and deployment from scratch, so that everyone has the opportunity to contact and master relevant technical knowledge.
Lecture 4. Introduction to BML CodeLab, a Machine Learning-Oriented Development Environment
In the "GPU Development and Tools" sub-forum on December 15, Ma Ruyue, chief architect of Baidu's AI development platform, will share the introduction of BML CodeLab, a development environment for machine learning, for developers.
This speech will focus on the development environment BML CodeLab to help developers achieve zero-threshold machine learning development. The content of the speech is divided into three parts:
- The background and features of the interactive development environment BML CodeLab, and quick to install and use . BML CodeLab is improved and optimized based on JupyterLab , and can be flexibly deployed to developers' local single machines, IDC machines, and hosted resources on the cloud. It has been highly optimized in performance, added many enterprise-level features, and seamlessly extended to cloud clusters when single-machine resources are limited.
- The principle of a high-performance data science engine to improve the speed of analysis and modeling. Using GPU and CPU many-core parallel acceleration and hybrid computing, large data processing, efficient data storage and other technologies, data science development not only maintains the simplicity and ease of use of a single machine, but also rivals the processing power of distributed systems. The BML CodeLab with built-in high-performance engine has nearly ten times higher performance than open source products.
- The built-in easy-to-use development plug-ins are used to improve development efficiency. Based on the open source Jupyterlab extension mechanism, BML CodeLab integrates many feature-rich and easy-to-use development tools. Such as : lightweight machine learning. The application development applet plug-in, through simple Python code, publishes the analysis and training results into high-performance applications; AI workflow plug-in, manages workflow arrangement and tracking experiments, and improves iterative efficiency.
With the theme of endless innovation, this year's GTC 2020 will officially meet with developers and friends from December 15th to 19th. Baidu Flying Paddle has prepared the content of dry goods, waiting to meet with old friends again. If you are also interested in the content of the flying paddle speech, don't forget to click to read the original text and sign up to participate!
Registration link: https://www.gtcevent.cn/Reg/?code=NVBAIY
About the paddle
PaddlePaddle is based on Baidu's many years of deep learning technology research and business applications. It is China's first open source, technology-leading, and fully functional industrial-level deep learning platform, including PaddlePaddle open source platform and PaddlePaddle Enterprise Edition. Flying Paddle open source platform includes core framework, basic model library, end-to-end development kit and tool components. It continues to open source core capabilities and provides a foundation for industry, academia, and scientific research innovation. The Flying Paddle Enterprise Edition is based on the Flying Paddle open source platform, and has enhanced corresponding features for enterprise-level needs, including the zero-threshold AI development platform EasyDL and the full-featured AI development platform BML. EasyDL is mainly for small and medium-sized enterprises, providing zero threshold, preset rich networks and models, convenient and efficient development platform; BML is a development platform with comprehensive functions, flexible customization and deep integration for large enterprises.
Download and install command ## CPU version installation command pip install -f https://paddlepaddle.org.cn/pip / oschina /cpu paddlepaddle ## GPU version installation command pip install -f https://paddlepaddle.org.cn/pip / oschina /gpu paddlepaddle-gpu