November 19, plan initiated by Alibaba, Baidu, Huawei, Tencent, Netease, and other well-known domestic enterprises, 2019 Soft Alliance Developers Conference was officially opened in Beijing National Convention Center. Baidu Ma Yanjun, director of learning technology platform depth, shared the latest technological advances deep learning platform open source fly paddle for large-scale industrial applications and 300+ partners and more than 2,000 developers present.
Depth study and development of breakthrough technologies, artificial intelligence and effectively makes the driving force at the core of all walks of life intelligent upgrade. The deck application, then under the chip depth learning framework can be described as the era of intelligent "operating system." Based on open source and open deep learning platform, can effectively reduce the depth of learning technology barriers for developers and businesses to avoid duplication, "create the wheel", which is the current industry mainstream development approach.
Fly to Baidu paddle years of deep learning technology research and business applications, based on a core set of deep learning framework, basic model-base, end to end development kit, tools, components and services in one platform, 2016 officially open source, open source is fully open, leading technology, complete industrial grade deep learning platform, and has a complete independent intellectual property rights. The event, Ni Guangnan also called channel entitled "Eco-innovation software and hardware to meet the new trend of" speech, China should accelerate the letter network technology self-control, and point out a way to promote open source software innovation software industry.
Current, AI + 5G + rapid development of IOT is opening the chi of life of the whole scene, all walks of life are being applied artificial intelligence technology for intelligent upgrade, build a pan-end software and hardware ecological self-control as an important issue. Face the challenges of large-scale industrial applications, on-site Ma Yanjun focused on sharing the latest achievements of side inference engine Paddle Lite 2.0 and four industrial end-stage two development kits to fly paddle.
Pan-end mobile, such as search software level sorting, the recommended video applications such as ultra-large-scale thinning characteristics, one trillion model parameters online continuous learning, challenging lifelong learning, as well as the hardware level heterogeneous computing challenges, Baidu launched a fly paddle lightweight side inference engine Paddle Lite 2.0.
据介绍，Paddle Lite 2.0版本拥有三大特点，包括：高易用性，提供大量的示例代码，以及操作的指南，方便快速实现在不同的设备场景上进行部署；广泛的硬件支持，Paddle Lite目前已经支持了8种主流的硬件，同时支持华为NPU和边缘设备的FPGA，且设计架构非常便于新硬件的拓展；以及性能优势，Paddle Lite性能优势不仅在FP32场景下，尤其是在移动端已广泛普及的INT8上领先优势会更加明显。性能优势对于实际的应用是至关重要的，时延上的优势直接关系到最终开发产品的用户体验。
值得一提的是，Paddle Lite与华为HiAI Foundation进行了深度联合优化取得了很好的进展。大会当天下午，华为正式对外发布了HiAI3.0版本。发布会上，百度深度学习技术平台部技术经理高铁柱，介绍了百度Paddle Lite端侧推理框特性及与华为进行的深度适配。双方在新一代人工智能产业技术创新战略联盟启动端测AI软件标准研究工作，通过标准化推理接口与标准化应用接口，促进端侧AI产业规模化有序发展，更好地助力开发者使用端侧AI进行应用创新。
除互联网领域外，当前深度学习也大量在传统行业中落地。针对落地过程中，问题定义技术选型、验证成本高研发周期长、设备成本限制及特定性能要求等真实挑战，飞桨推出了4大面向应用任务的产业级开发套件，包括：ERNIE语义理解，PaddleDetection目标检测、PaddleSeg图像分割，以及推荐方向的ElasticCTR点击率预估，降低开发门槛，满足低成本和快速集成需求。
语义理解方面，ERNIE开发套件是基于持续学习语义理解框架ERNIE 2.0升级而来，在16个中英文任务全面超越对标产品，具备轻量方案、能力全面、极速预测、部署灵活、平台赋能等五大特色。
PaddleSeg产业级图像分割库，覆盖了DeepLabv3+、PSPNet、U-Net、ICNet四大主流的分割 模型。通过统一的配置，帮助用户更便捷地完成从训练到部署的全流程图像分割应用。
Target detection aspects, PaddleDetection, 60+ pre-integrated training model and contains Baidu international model contest winner. The purpose is to provide easy to use for both industry and academia target detection model. In this library, also it provides many fly paddle small target detection model, easy to use the mobile terminal device.
ELASTIC CTR for personalized recommendation, Baidu's practice derived from the industry. CTR estimates can be achieved distributed training mission and Serving a key deployment process, CTR provides training and secondary development of end to end solution.
Today, the national economy and an important industry involving intelligent speed is constantly upgrading, industry head enterprise of artificial intelligence application depth and breadth has been considerable. Baidu fly paddle is a variety of fields, bringing tangible value. It is reported that fly paddle deep learning platform service has accumulated more than 1.5 million developers, only on customized training platform will have more than 65,000 business users, announced the 169,000 models.
