The service platform

AI Studio

website ：https://aistudio.baidu.com/aistudio/index

AI Studio It's based on Baidu deep learning platform AI Development platform , Provide online programming environment 、 free GPU Calculate the force 、 Massive open source algorithms and open data , Help developers quickly create and deploy models .

EasyDL

website ：https://ai.baidu.com/easydl/

Zero algorithm training model , There is no need for machine learning expertise , Just upload and mark the sample data that needs to be identified to train the model with one click .EasyDL Baidu brain is the introduction of customized AI Training and service platform , Support customization for all walks of life AI Enterprise users and developers who need to use . Support from data management and data annotation 、 model training 、 One stop model deployment AI Development process , Through the original picture 、 Text 、 Audio 、 Video data through EasyDL machining 、 Study 、 Deployment can be published as a public cloud API 、 Device end SDK 、 Localization deployment and integrated software and hardware products .EasyDL The products are divided into classic versions from the perspective of target customers and application scenarios 、 pro 、 Two core products of Retail Edition .

• Classic version is oriented to zero algorithm basis or high efficiency development AI Enterprise users , Image classification is now supported 、 Object detection 、 Image segmentation 、 Text classification 、 Video classification 、 Voice classification six types of model customization .
• The professional edition is for AI A beginner or AI Launched by professional engineers AI Model training and service platform , At present, it supports vision and natural language processing , Built in Baidu mass data training pre training model , Flexible script parameter adjustment , Only a small amount of data is needed to achieve the optimal model effect .
• The retail version is dedicated to the retail scene ISV、 Retail industry service providers and other enterprise users provide 【 Commodity identification scenario 】 Of AI Service acquisition plan , Support shelf oriented inspection 、 Self service checkout desk 、 Unmanned Retail cabinets and other commodity detection scenarios provide customized commodity detection training platform and standard commodity detection API Two types of services .

EasyEdge

website ：http://ai.baidu.com/easyedge/

It can be based on a variety of deep learning frameworks 、 The model of network structure , Fast generation of end computing model and encapsulation SDK, Suitable for a variety of AI Chips and operating systems . be based on Paddle Lite R & D end computing model generation platform , It can help deep learning developers to quickly deploy their own models to the device side . Just upload the model , The fastest 2 The end calculation model can be generated and obtained by dividing SDK.

See the table below for details of platform support ：

• Upload model support framework ：Caffe (ssd) 、PyTorch (1.4) 、TensorFlow (1.14) 、PaddlePaddle (1.6.2)
• The upload model supports the network ：VGG16 、InceptionV3/V4 、MobilenetV1 、MobilenetV1-SSD 、YoloV3 etc. 20 Kind of （2020.1.17 New support YoloV3 Wait for the Internet 、NNIE chip ）
• AI Chip acceleration support ： Universal ARM chip 、 Universal x86 chip 、 Ying Wei Da GPU 、 qualcomm Snapdragon GPU/DSP 、 Intel Movidius VPU 、 Huawei HiSilicon NPU 、 Huawei Hisilicon NNIE 、 Apple A-Bionic

Tools

AutoDL

AutoDL Of Github link ：https://github.com/PaddlePaddle/AutoDL

An efficient automatic search method to build the best network structure , Through enhanced learning in the process of continuous training to get customized high-quality models . The system consists of two parts , The first part is the encoder of network structure , The second part is the network structure evaluator . Encoder usually uses RNN To code the network structure , Then the evaluator will take the results of the coding for training and evaluation , Get including accuracy 、 Some indicators including model size , Feedback to encoder , Encoder modification , Re code , So iterate . After several iterations , Finally, we get a designed model .

PaddleHub

PaddleHub Of Github link ：https://github.com/PaddlePaddle/PaddleHub

Easy access PaddlePaddle Pre training model in Ecology , Complete model management and one click Forecasting . In combination with Fine-tune API , Based on large-scale pre training model, transfer learning can be completed quickly , Let the pre training model better serve the application of user specific scenarios .PaddleHub The pre training model provided covers image classification 、 object detection 、 Lexical analysis 、 semantic model 、 Sentiment analysis 、 Video classification 、 Image generation 、 Image segmentation 、 Text review 、 Key point detection and other mainstream models .

PaddleHub Taking the application of pre training model as the core, it has the following characteristics ：

• Models are software , adopt Python API Or command line implementation model call , You can quickly experience or integrate the pre training model with the characteristics of the propeller .
• Easy to use transfer learning , adopt Fine-tune API , Built in multiple optimization strategies , Only a small amount of code is needed to complete the pre training model Fine-tuning .
• One click model to service , A simple command can build your own deep learning model API Service deployment complete .
• Automatic super parametric optimization , built-in AutoDL Finetuner Ability , One click start automatic super parameter search .

PARL

PARL Of Github link ：https://github.com/paddlepaddle/parl

A high performance 、 Flexible reinforcement learning framework .

• They provide mainstream reinforcement learning algorithm implementation , Strictly reproduce the corresponding index of the paper .
• Large scale parallel support . The framework can support up to ten thousand CPU Concurrent computing , And support more GPU Strengthen the training of learning model .
• High reusability . Users do not need to re implement the algorithm themselves , The algorithm provided by the reuse framework can easily apply the classical reinforcement learning algorithm to specific scenarios .
• Good scalability . When users want to research new algorithms , You can quickly implement your own reinforcement learning algorithm by inheriting the base class provided

ERNIE

Continuous learning semantic understanding framework Aini （ERNIE ） Using Baidu massive data and flying oars （PaddlePaddle ） Multi machine multi card efficient training advantages , Through deep neural network and multi task learning technology , Keep learning massive data and knowledge . Eni based on this framework （ERNIE ） Pre training model , Accumulated learning 10 More than 100 million knowledge , Help each other NLP The task is significantly improved .

PaddleX

website ：https://www.paddlepaddle.org.cn/paddle/paddleX

The whole process development client of the propeller , Set the core frame of the propeller 、 model base 、 Tools, components and other in-depth learning and development capabilities required for the whole process , It not only provides you with a one click installation client , The open source technology kernel is more convenient for you to directly call or secondary development according to the actual production needs , It is the best auxiliary tool to improve the development efficiency of deep learning projects .

Develop deep learning from data access 、 model training 、 Parameter tuning 、 Model to evaluate 、 Through the whole process of prediction and deployment , And provide a visual interface , It eliminates the code development and script calls between the links , Greatly improve the development efficiency .

Paddle Lite

Paddle Lite Of Github link ：https://github.com/PaddlePaddle/Paddle-Lite

file ：https://paddle-lite.readthedocs.io/zh/latest/

FPGA Deploy :https://paddle-lite.readthedocs.io/zh/latest/demo_guides/fpga.html

Paddle Lite Committed to providing a complete set of functions 、 Easy to use 、 High performance end-to-end reasoning engine , It is convenient for developers to deploy applications to any end-to-end devices . Compared to the original beta edition , The official version is being compiled 、 file 、 performance 、 Hardware support 、 Platform support and other aspects have been greatly improved . The core purpose is to quickly deploy the trained model in different hardware platform scenarios , According to the input data , Perform predictive reasoning to get the calculation result , Support practical business applications .

PaddleCV

PaddleCV Of Github link ：https://github.com/PaddlePaddle/models/tree/develop/PaddleCV

be based on PaddlePaddle Intelligent visual tools developed by deep learning framework , Algorithm , Open source projects for models and data . Baidu in CV The deep accumulation in the field for many years is PaddleCV Provides a strong core power .PaddleCV Integrated with rich CV Model , Covering image classification , object detection , Image segmentation , Video classification , Action positioning , Target tracking , Image generation , Character recognition , Measure learning , Key point detection ,3D Vision, etc CV technology . meanwhile ,PaddleCV It also provides practical tools ,PLSC Support super large-scale classification ,PaddleSlim and PaddleLite Support industrial deployment , as well as PaddleDetection 、PaddleSeg Industry oriented end-to-end Development Suite , Through the model development 、 Compress 、 Deploy the whole process .