mmdnn-graph-utils

This simple package allows importing mmdnn-ir files, onnx-files and stores them in ANNETTE format Network Architecture description files

Usage

• as submodule clone to your python module then see tests folder for examples, the init.py file ensures relative import so no adaptations should be necessary

import mmdnn-graph-utils as graph
print(graph.__dict__)

to see available classes

ANNETTE Format

Generally the Annette Format is derived from the MMDNN IR format in terms of attributes, and stored in json format.

Network header:

{
    "name" : *network_name*,
    "layers" : {
        "layer_0" : {
            "type" : "Conv",
            "parents" : [],
            "children" : ["layer_1"],
            ...Layer Attributes...
        },
        "layer_1" : {
            "type" : "Relu",
            "parents" : ["layer_0"],
            "children" : [],
            ...Layer Attributes...
        }
    },
    "input_layers" : ["layer_0"],
    "output_layers" : ["layer_1"]
}

• "name" ... denotes the network name and should be identical to the filename
• "layers" ... contains a dictionary with all network layers with layer_names as keys and layer attributes as values
• "input_layers" ... list of the names of the input_layers
• "output_layers" ... list of the namos of the output_layers

Layers

• the key of each layer represents the layer name
• the value of each layer contains:
  • "parents" ... list of parent layers in the network graph
  • "children" ... list of cild layers in the network graph
  • "input_shape" ... a list of the input shape of the layer inputs
    • NHWC-format
    • if multiple inputs then as a list of lists, order identical to the "parents list"
  • "output_shape" ... a list of the input shape of the layer inputs
    • NHWC-format
    • if multiple outputs then as a list of lists, order identical to the "children list"
  • Any additional attribute ... cild layers in the network graph

Layer Examples

DataInput

"Placeholder": {
    "type": "DataInput",
    "parents": [],
    "children": [
        "Conv1"
        ],
    "output_shape": [
        1,
        224,
        224,
        3
    ]
}

Convolution

"resnet_v1_18/conv1/Conv2D": {
    "type": "Conv",
    "parents": [
        "resnet_v1_18/Pad"
    ],
    "children": [
        "resnet_v1_18/conv1/BatchNorm/FusedBatchNorm"
    ],
    "output_shape": [
        -1,
        112,
        112,
        64
    ],
    "input_shape": [
        -1,
        230,
        230,
        3
    ],
    "kernel_shape": [
        7,
        7,
        3,
        64
    ],
    "strides": [
        1,
        2,
        2,
        1
    ],
    "pads": [
        0,
        0,
        0,
        0,
        0,
        0,
        0,
        0
    ]
}

for more examples see tests/data/