Tensorflow Lite source code analysis

        TensorFlow Lite It is specially re implemented for the characteristics of mobile and embedded devices TensorFlow edition . Compared with ordinary TensorFlow, Its functions are more streamlined , Model training is not supported , Distributed operation is not supported , There is not much cross platform logic , Supported by op It's also limited .TensorFlow Lite Use to run on the mobile end , Some devices even have very limited resources , Therefore, the overhead in memory and parsing must be reduced as much as possible .TensorFlow The exported model uses Protocol Buffer agreement , Therefore, it is necessary to convert it to a better performance FlatBuffer Format .

         You can go through more than one Delegate Use multiple accelerators to accelerate , As long as we are Interpreter After instantiation and calling Interpreter.Invoke() Initialize different types of Delegate, for example GPU、NNAPI Waiting Delegate, And use interpreter->ModifyGraphWithDelegate(delegate) For these Delegate Just register one by one .

TFLite Reasoning is mainly divided into the following steps ：

1.  Load model file .

2.  Create a class Interpreter Example . In the process , The inference engine will parse the model 、 Allocate the memory space needed in the running process of the model and determine the execution order between various layers in the model .

3.  adopt Delegate Use accelerator to accelerate .

4.  Ready to input data , This requires corresponding processing according to the input format required by the model ;

5. call Interpreter.Invoke() Reasoning ;

6.  Get reasoning results .

Pre training model ：

• inception v3： For target detection
• MobileNets： A model specifically for mobile terminals , With low latency , high velocity ,