Tengine, developed by Open AI lab, is a lite, high-performance, and modular inference engine for embedded device.

Tengine is composed of four modules: core/operator/serializer/executor.

• core provides the basic components and functionalities of the system.

• operator defines the schema of operators, such as convolution, relu, pooling, etc. al. Here are the current support operator list

• serializer is to load the saved model. The serializer framework is extensible to support different format, including the customized one. Current version only support caffe model. Tensorflow and MXNet support will be the next.

• executor implements the code to run graph and operators. Current version only provides a highly optimized implementation for single A72.

This version can load and run caffe model of mobilenet and squeezenet directly. For more details, please goto install.

NOTE: old caffe model has to be upgraded using upgrade_net_proto_binary/upgrade_net_proto_binary from caffe's package

The data is collected on single 1.8G A72 on chip RK3399, by repeating calling the forward interface to get the average time cost (ms) per run.

For details to run benchmark, please visit benchmark page.

please refer to the install page

It is easy to add new operator to Tengine. Here is the guide on new operator

Tengine can be extended to support new serialization format, by building new serializer module.

How to build new serializer module

Initial release of single A72 support