Skip to content

create-nodes.md

Latest commit

 

History

History
43 lines (34 loc) · 2.19 KB

create-nodes.md

File metadata and controls

43 lines (34 loc) · 2.19 KB

On creating new node types

Creating a model consists of to phases:

  1. Parsing the model-json and creating an abstract representation.
  2. Using the abstract representation to actually build the model. This bears the advantage, that the math-library used in the model can be changed, by only chaning the building process.

First Phase

ParsedNode

The base class ParsedNode can be found in HybridML.parsing.DataModel. Implement your CustomParsedNode in parsing/nodes/Custom.py, where Custom is the name of your new node. Abstract representation of the parsed node. Contains information that is filled in by the parser.

Mandatory implementation:

  • determine_output_sizes is needed to propagate the size of the in and outputs. It returns a list of the sizes of the outputs

Optional overwrites:

  • get_input_ids in the base class the value of the inputs attribute from the model-json is used. If another way of determining the ids of the inputs is wished this can be overwritten.
  • get_output_ids: similar to get_input_ids
  • __init__: in the base class only saves the json-data.

NodeParser

Can be found in HybridML.parsing.BaseParsers. Implement your CustomNodePaser in parsing/nodes/Custom.py. Class used to parse a model-json dict and create a ParsedNode. The parses_types attribute saves a list of strings, saying which types of nodes it can parse. The types are thoses, that are used in the model-json to type a node, eg. "nn", "ArithmeticExpression","linear_ode". Usually the parser does not do much apart from handing the information to the ParsedNode.

Second Phase

NodeBuilder

The base class NodeBuilder can be found in HybridML.building.BaseBuilders. Implement your CustomNodeBuilder in building/nodes/Custom.py This class creates a Node and puts it into a BuiltNodeContainer. If it is necessary to create a new custom keras layer, it can be saved in keras/layers/Custom.py

Add it to the Tool

In the file NodeRegistry.py, all parsers and builders are registered. In case a new Tensorflow Model or Layer has been implemented, it has to be added to the custom_objects dictionary.

Example

A simple implementation of a new node can be found in the BasicFunctionLayer