Shulou(Shulou.com)05/31 Report--

In this issue, the editor will bring you what are the five advantages of PaddlePaddle. The article is rich in content and analyzes and narrates it from a professional point of view. I hope you can get something after reading this article.

PaddlePaddle is the only fully functional end-to-end open source deep learning platform in China, which integrates deep learning training and prediction framework, model base, tool components and service platform. It has flexible and efficient development mechanism, industrial application effect model, super-large-scale parallel deep learning ability, integrated design of reasoning engine and systematic service support. Committed to making the innovation and application of deep learning technology easier.

Since Paddle Fluid v1.0, Flying Propeller has devoted itself to creating a better user experience. Taking advantage of Baidu developer Conference, Flying Propeller has also carefully prepared a gift for users, carried out a new upgrade in the whole process of development, training and deployment, and released five major features of Flying Propeller. Next, the editor interprets it for you one by one.

1. Dynamic diagrams & static diagrams-both dynamic diagrams and static diagrams have the advantages of both

Starting from Padlde Fluid v1.5, the core framework of Flying Propeller, Flying Propeller provides users with both dynamic and static diagrams. The static diagram defines the network structure first and then runs it. Analyzing the defined graph structure can make the running speed faster and take up lower memory. It has great advantages in business deployment and provides efficient support for users'AI application landing. But the static graph networking and the execution phase are separated, which is not very friendly for new users to understand.

Since the latest version, Flying Propeller provides a more convenient dynamic graph mode, and all operations can get the execution results immediately, without having to wait for the execution phase to get the results, which makes it more convenient to debug the model. At the same time, it also reduces a lot of code used to build Executor, which makes the process of writing and debugging the network more convenient. Users can use the more convenient dynamic graph mode for debugging and training, and then the trained model can be transformed into the structure of static graph and quickly deployed online.

Second, the official model with the best application effect-covering three mainstream tasks

Based on Baidu's many years of industrial application experience and the artificial intelligence solution practice of Baidu's ecological partners, Flying Propeller provides users with 70 + selected official algorithm models that have been verified by real business scenarios and have the best application results, covering AI core technology areas such as vision, NLP, voice and recommendation.

Flying Propeller Natural language processing Model Library PaddleNLP: based on the industrial Chinese NLP open source toolset built by Flying Propeller, it has the best Chinese semantic representation model in the industry and a pre-training model based on 10 billion-level big data training, and a variety of models in the field of natural language processing are implemented with a set of shared skeleton code, which can greatly reduce users' repetitive work in the development process. Users can not only greatly reduce the cost of research and development, but also obtain better application results based on industrial practice. The release of PaddleNLP-Research, support for NLP cutting-edge research, has opened up the recent work of Baidu in the academic field of MRQA2019, such as Paddle Fluid baseline, DuConv (ACL2019), ARNOR (ACL2019), MMPMS (IJCAI2019), MPM (NAACL2019) and so on.

Flying oar visual model library PaddleCV: based on flying oars to create the industry's best CV open source tool set, and open source a number of Baidu self-developed, international competition winning solution model. In the basic task libraries of object detection unified framework, image classification library, image generation library and video recognition library, there are both high-precision model and high-speed reasoning model. Based on the easy-to-expand and easy-to-modular operation, users can efficiently complete the industrial applications of all kinds of visual tasks.

PaddleDetection object detection unified framework covers the mainstream detection algorithms, including high-precision model and high-speed reasoning model, including Faster-RCNN (supporting FPN), Mask-RCNN (supporting FPN), Cascade-RCNN, RetinaNet, Yolo v3, SSD algorithms and providing a series of pre-training models, which has the advantages of industrialization, modularization and high performance. The high-speed reasoning engine combined with the flying propeller core framework can connect seamlessly from training to deployment; provide modular design, model network structure and data processing can be customized; based on the efficient core framework, there are certain advantages in training speed and memory consumption. For example, the training speed of YOLO v3 is 6 times faster than that of similar frameworks. In addition, in addition to the unified detection framework, a series of pre-training models are also released, such as the detection model based on the improved version of ResNet, which generally improves the accuracy by about 1% without increasing the amount of calculation.

Nine new image classification models have been added to the image classification database, covering 10 and more than 25 ImageNet pre-training models so far, of which the ResNet model continues to improve and publishes improved models with a considerable amount of calculation. for example, the accuracy of ResNet50 Top1 has increased from 5% to 79.84% (+ 3.34%).

PaddleGAN provides users with easy-to-use and one-click runnable GAN models, covering mainstream GAN algorithms, including CGAN, DCGAN and Pix2Pix,CycleGAN,StarGAN,STGAN,ATTGAN. STGAN is a face attribute editing model developed by Baidu and published in CVPR 2019.

Following the release of the industry's first video identification and positioning tool set in April, PaddleVideo continues to optimize the training speed, and the speed of some models is 30% better than that of similar products. This new addition of C-TCN, Baidu's self-developed video action positioning model, is also the first time that ActivityNet won the championship in 2018.

Based on the pre-training model, users can more easily complete their own AI applications. Flying Propeller provides users with pre-training model management and transfer learning component PaddleHub, which can load the industrial pre-training model with one click. This new release of 29 pre-training models, a total of 40 + pre-training models for users, covering text, image, video three major areas of eight types of models. PaddleHub provides Fine-tune API,10 lines of code to complete the transfer learning of large-scale pre-training model. PaddleHub also introduces the concept of "model is software", through Python API or command line tools, one line of code to complete the prediction of the pre-training model.

Third, large-scale distributed training-the industry's strongest ultra-large-scale parallel deep learning ability.

The flying paddle supports ultra-large-scale deep learning parallel training of both dense and sparse parameter scenes, and supports efficient parallel training of hundreds of billions of parameters and hundreds of nodes. It is also the first deep learning platform to provide such a powerful deep learning parallel technology.

Flying Propeller provides a cost-effective multi-machine CPU parameter server solution. Based on the data verification of real recommendation scenarios, it can effectively solve the problems of very large-scale recommendation system, super-large-scale data, self-expanding massive features and high-frequency model iteration, and achieve high throughput and high speedup.

Based on Paddle Fluid v1.5, distributed training newly released High-level API Fleet, the cost of stand-alone-to-distributed training is significantly reduced; the performance of GPU multi-machine and multi-card is significantly improved, and the speed of 4x8 v100 configuration in ResNet50, BERT, ERNIE and other models is more than 50% faster than that of previously released Benchmark.

Fourth, end-to-end deployment-reasoning engine integrated design, training to multi-terminal reasoning seamless docking mobile terminal acceleration

Based on Paddle Fluid v1.5, Flying Propeller fully supports multi-framework, multi-platform and multi-operating system, providing users with high compatibility, high performance, multi-terminal deployment capability, comprehensive and leading underlying acceleration library and reasoning engines Paddle Mobile and Paddle Serving.

For developers, in addition to model training, in the production process will also encounter a variety of engineering problems. With the widespread use of mobile devices, the application of deep learning and neural network technology in mobile Internet products has become an inevitable trend. For example, when deployed on the mobile side, you need to face a lot of problems, such as installation package size, running memory footprint, reasoning speed and effect, and so on. Current mainstream models are difficult to deploy directly to mobile devices. In the release in April, PaddleSlim implemented three main compression strategies: network quantization, pruning and distillation, and can quickly configure a combination of multiple compression strategies. For the already small MobileNet model, the volume compression of more than 70% is achieved without losing the effect of the model.

This version of PaddleSlim is further upgraded with the addition of automatic pruning strategy based on simulated annealing and lightweight model structure automatic search function Light-NAS, which reduces FLOPS by 17% compared with MobileNet v2 in ImageNet 1000 classification tasks with lossless accuracy. And in Baidu's OCR recognition, human body detection, face key point detection and other business line applications, the accuracy is non-destructive or even improved, the speed has been improved by 30%.

Service support-the only platform that provides systematic deep learning technical services

Flying Propeller has achieved the stability and backward compatibility of API, providing users with complete bilingual documentation in both Chinese and English, from introductory tutorials to installation and compilation documents, user manuals, model documents, API interfaces and index documents. At the same time, provide a systematic service system to protect corporate partners, help universities and educational partners to build a sound system, and provide developers with different levels of training system.

These are the five advantages of PaddlePaddle shared by the editor. If you happen to have similar doubts, you might as well refer to the above analysis to understand. If you want to know more about it, you are welcome to follow the industry information channel.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.