PaddlePaddle deep learning platform v2.6.2
2.6.2
PaddlePaddle (PArallel Distributed Deep LEarning Parallel Distributed Deep Learning) is a deep learning platform developed by Baidu. It is easy to use, efficient, flexible and scalable. It provides deep learning algorithm support for many internal products of Baidu.
PaddlePaddle is currently the only industrial-level deep learning platform in China that is independently developed, open source, and fully functional. It integrates the deep learning core framework, basic model library, end-to-end development kit, tool components, and service platform. Fei Paddle is derived from industrial practice and is committed to in-depth integration with the industry. It provides leading deep learning & machine learning task development, training, and deployment capabilities, accelerating the process of enterprises from algorithm research and development to industrial implementation. Currently, Flying Paddle has been widely used in industry, agriculture, service industries, etc., serving more than 1.5 million developers, and working with partners to help more and more industries complete AI empowerment.
1. Develop a convenient industrial-level deep learning framework
The Feipiao deep learning framework adopts a networking paradigm based on programming logic, which is easier for ordinary developers to get started and conforms to their development habits. It supports both declarative and imperative programming, combining development flexibility and high performance. The network structure is automatically designed, and the model effect exceeds that of human experts.
2. Support the training of ultra-large-scale deep learning models
Feipiao has broken through the training technology of ultra-large-scale deep learning models and implemented an open-source large-scale training platform that supports hundreds of billions of features, trillions of parameters, and hundreds of nodes. It has overcome the online learning problems of ultra-large-scale deep learning models and achieved trillions of scale. Real-time updates of parameter models.
3. High-performance inference engine deployed on multiple terminals and multiple platforms
Flying Paddle is not only compatible with models trained by other open source frameworks, but can also be easily deployed on platform devices of different architectures. At the same time, the reasoning speed of the flying paddle is also leading in all aspects. Especially after the integrated software and hardware optimization with Huawei's Kirin NPU, Fei Paddle's reasoning speed on the NPU has further broken through.
4. For industrial applications, open source and open industrial-grade model libraries covering multiple fields.
Flying Paddle officially supports more than 100 mainstream models that have been polished through long-term industrial practice, including models that have won championships in international competitions. At the same time, more than 200 pre-trained models are open source and open to facilitate rapid industrial applications.
