ppljs-ppl-core/src/interface/README.md

ppljs network inference framework core module

@author xusiyu@sensetime.com

PPL.js json format follows thd definition of OpNode in interface.ts file:
the total model is representer by OpNode,and OpNode is composed by some attributes as follows:
/*****************************************************************************************/
1.props:OpProps
OpPros defines the Op's type and name.Common op types such as "Conv","Relu"...Now we only support the caffe op,
so you can search the op's specific definitions on the caffe official website.The name of op is a unique value, 
we can use this value to find and operate a specific op

2.attrs:OpAttrs

3.inputs:OpEdge[]
The model must have at least one inputs OpEdge whose name contains "input".for example,if model have 4 inputs,
you should name corresponding op's inputs's name "input0"/"input1"/"input2"/"input3"

4.outputs:OpEdge[]
The model must have at least one outputs OpEdge whose name contains "output".for example,if model have 4 outputs,you should name corresponding op's outputs's name "output0"/"output1"/"output2"/"output3"

5.hostDataOffset:number;

6.data:OpData
The data structure is composed of key-value pairs. The key is the name of the parameter with op, and the value is the tensorInfo interface.The tensorInfo is composed of {name,shape,precision,hostData,deviceData},we will
Initialize data such as weights/bias in hostData,If the backend is gpu,we will copy the hostData to deviceData
in the end.