Intermediate JSON Schema

Within the Neural-Network-Translator the frontend-plugins transform a given network model created with a specific framework into an intermediate format, which can be translated into an executable form for a specific backend.
In the following, the intermediate format will be described in more detail to understand what happens in a frontend plugin.

The correctness of the generated network model is checked after the transformation by the respective frontend plugin during the program run of the Neural-Network-Translator. The program run is aborted if the result does not correspond to the defined schema.

The intermediate format was created as JSON and, therefore, can easily be generated and read in Python.

Structure of the Intermediate Format

The basic structure of the format looks like follows:

{
    "class_name" : <type of the network>,
    "config" : {
        "name" : <name of the network>,
        "layers" : [
            {
                "class_name" : <type of the layer>,
                "config" : <configuration of the layer>
            },
            ...
        ]
    }
}

Valid Attributes contained in the Layer Config Object

The configuration of the layer can contain multiple attributes which define the behavior of the layer.

The following attributes are valid entries in the configuration object:

Key	Value
batch_input_shape	Shape of the input values. For example: [null,8]
dtype	Type of the input values
units	Number of neurons in the layer
activation	Name of the used activation function
use_bias	Boolean indicating if bias values are used in the layer
kernel_values	Weight values for the calculation process
bias_values	Bias values for the calculation process if use_bias = True
data_format	channels_last or channels_first in order of the input format
pool_size	Array with the size of the pooling kernel. For example: [2,2]
padding	same if padding is used or valid if not
strides	Array with the stride size. For example: [2,2]

Example of the Intermediate Format

The following code is of an exemplary network with a dense layer:

{
    "class_name": "Sequential",
    "config": {
        "name": "sequential_1",
        "layers": [
            {
                "class_name": "Dense",
                "config": {
                    "name": "dense_1",
                    "batch_input_shape": [
                        null,
                        8
                    ],
                    "dtype": "float32",
                    "units": 8,
                    "activation": "sigmoid",
                    "use_bias": true
                },
                "kernel_values": [...],
                "bias_values": [...]
              }, ...
        ], ...
    }
}