- New features
- Add ir_optim, use_mkl(only for cpu version)argument
- Support custom DAG for prediction service
- HTTP service supports prediction with batch
- HTTP service supports startup by uwsgi
- Support model file monitoring, remote pull and hot loading
- Support ABTest
- Add image preprocessing, Chinese word segmentation preprocessing, Chinese sentiment analysis preprocessing module, and graphics segmentation postprocessing, image detection postprocessing module in paddle-serving-app
- Add pre-trained model and sample code acquisition in paddle-serving-app, integrated profile function
- Release Centos6 docker images for compile Paddle Serving
- Bug fixed
- New documents
- Performance optimization
- Optimized the time consumption of input and output memory copy in numpy.array format. When the client-side single concurrent batch size is 1 in the resnet50 imagenet classification task, qps is 100.38% higher than the 0.2.0 version.
- Compatibility optimization
- The client side removes the dependency on patchelf
- Released paddle-serving-client for python27, python36, and python37
- Server and client can be deployed in Centos6/7 and Ubuntu16/18 environments
- More demos
- Chinese sentiment analysis task : lac+senta
- Image segmentation task : deeplabv3、unet
- Image detection task : faster_rcnn
- Image classification task : mobilenet、resnet_v2_50