Model Conversion¶

Commands for converting model files from various frameworks into the SoftNeuro DNN file format.

About preprocess, postprocess¶

In SoftNeuro preprocess and postprocess are separate networks attached to the beginning and to the end of the main network, respectively. These networks can be attached as a json file when converting a model. There's a preset json file, but it can be changed for custom settings.

Currently, adding preprocess and postprocess networks is only available for import-onnx and import-tensorflow, and not available for import-ver3.

Preprocess specification¶

The preprocess network can be used for resizing images, specifying resize mode, specifying color format, normalizing data, and so on. Descriptions of each element in preprocess is below.

layers¶

Each layer consists of four elements: type, params, weights, and attrs.

Properties	Required	Description	Type	Example
type	Mandatory	String specifying the layer type. source, sink, madd, permute, can be specified.	String	"madd"
params	Optional	JSON object specifying layer parameters.	Object	example
weights	Optional	JSON object specifying leyer weights.	Object	example
attrs	Optional	JSON object specifying layer attributes.	Object	example

Source Layer¶

Input layer of the network.

params

Properties	Required	Description	Type	Example
shape	Mandatory	Array of integer specifying the shape of the input data.	Array of Integer	[224, 224, 3]

weights: None
attrs

Properties	Required	Description	Type	Example
format	Optional	String specifying the color format of the input image. Only "rgb" can be specified.	String	"rgb"
resize_mode	Optional	String specifying the resize mode. "bilinear" (default) or "nearest" can be specified.	String	"bilinear"
keep_ar	Optional	Boolean specifying whether or not to keep aspect ratio when resizing. Default is false.	Boolean	true
padding_color	Optional	Array of numbers specifying the padding color filling the margin space when resizing.	Array of Numbers	[255, 255, 255]

Sink Layer¶

Output layer of the network.

params: None
weights: None
attrs: None

Permute Layer¶

Permutes the data order in the specified axis. Commonly used for channel-swapping.

params

Properties	Required	Description	Type	Example
axis	Mandatory	Integer specifying the axis to be permuted.	Integer	2
order	Mandatory	Array of integers containing the original positions ordered by their new indexes.	Array of Integer	[2, 1, 0]

weights: None
attrs: None

Madd Layer¶

Multiply add (scale * x + bias) layer. Commonly used for normalization of input data, subtraction of mean, and so on.

params

Properties	Required	Description	Type	Example
has_relu	Optional	Boolean specifying whether or not to execute Relu activation. Default is False.	Boolean	False
relu_max_value	Optional	Max output value after Relu activation. Only valid when has_relu is True.	Number	6.0

weights

Properties	Required	Description	Type	Example
scale	Mandatory	Number or array of numbers by which the input data is multiplied. Must be array if different value is specified for each channel. Must have the same number of elements as bias.	Number or Array of Number	[0.01712475383, 0.0175070028, 0.01742919389]
bias	Mandatory	Number or array of numbers, to be added to the input data. Must be array if different value is specified for each channel. Must have the same number of elements as scale.	Number or Array of Number	[2.11790393013, 2.03571428571, 1.80444444444]

attrs: None

preprocess example¶

[
    {
        "name": "preprocess",
        "layers": [
            {
                "type": "source",
                "params": {
                    "shape": [
                        1,
                        320,
                        320,
                        3
                    ]
                },
                "attrs": {
                    "format": "rgb",
                    "attrs": {
                        "format": "rgb",
                        "resize_mode": "bilinear",
                        "keep_ar": true,
                        "padding_clor": [
                            0,
                            0,
                            0
                        ]
                    }
                }
            },
            {
                "type": "permute",
                "params": {
                    "axis": 3,
                    "dims": [
                        2,
                        1,
                        0
                    ]
                }
            },
            {
                "type": "madd",
                "weights": {
                    "scale": [
                        0.0135694,
                        0.0143123,
                        0.0141064
                    ],
                    "bias": [
                        -1.41176,
                        -1.63139,
                        -1.69065
                    ]
                }
            },
            {
                "type": "sink"
            }
        ]
    }
]

Postprocess specification¶

The postprocess network can be used for labeling outputs, specifying the decoding type for tasks such as Object detection, and so on. Descriptions of each element in postprocess is below.

layers¶

Each layer consists of four elements, type, params, weights, and attrs.

Properties	Required	Description	Type	Example
type	Mandatory	String specifying the layer type. source, sink, decode_centernet, decode_pelee, decode_ssd, decode_yolov3, decode_yolov4 can be specified.	String	"decode_ssd"
params	Optional	JSON object specifying layer parameters.	Object	example
weights	Optional	JSON object specifying leyer weights.	Object	example
attrs	Optional	JSON object specifying layer attributes.	Object	example

Source Layer¶

Input layer of the network.

params

Properties	Required	Description	Type	Example
shape	Mandatory	Array of integers specifying the shape of the input data.	Array of Integer	[224, 224, 3]

weights: None
attrs: None

Sink Layer¶

Output layer of the network.

params: None
weights: None
attrs

Properties	Required	Description	Type	Example
label_list	Optional	Array of strings representing output labels.	Array of String	["cat", "dog", "bird"]

Decode Layer¶

Executes decode operation for Object Detection task. decode_centernet, decode_pelee, decode_ssd, decode_yolov3, decode_yolov4 can be specified for type.

params

Properties	Required	Description	Type	Example
keep_top_k	centernet, pelee, ssd, yolov3, yolov4	Integer specifying the number of bounding boxes. Bounding box is selected in descending order of score. Default is 300.	Integer	300
do_nms	centernet, pelee, ssd, yolov3, yolov4	Boolean specifying whether or not to execute NMS (Non-Maximum Suppression). Default is true.	Boolean	true
nms_thresh	centernet, pelee, ssd, yolov3, yolov4	IoU threshold for NMS.	Float	0.5
conf_thresh	centernet, pelee, ssd, yolov3, yolov4	Confidence threshold for detecting bounding boxes.	Float	0.3
background_label_id	pelee, ssd	Label number of background. Default is 0.	Integer	0
img_width	yolov3	Width of input image.	Integer	416
img_height	yolov3	Height of input image.	Integer	416
num_anchors_per_out	yolov3	Number of anchor boxes for each grid cell.	Integer	3
anchors	yolov3	Array of numbers representing the size of anchor boxes. [w1, h1, w2, w3, ...]	Float array	[10,13, 16,30]

weights: None
attrs: None

postprocess example¶

[
  {
      "name": "postprocess",
      "layers": [
        {
            "name": "input",
            "type": "source",
        "params": {
            "shape":[1, 300, 15]
        }
        },
        {
            "name": "decode",
            "type": "decode_ssd",
            "params": {
          "keep_top_k" : 300,
          "background_label_id" : 0,
          "nms_thresh" : 0.45,
          "conf_thresh" : 0.01
        }
        },
        {
              "name": "output",
              "type": "sink"
        }
      ]
  }
]

import-onnx¶

Convert an ONNX model to DNN.

Usage

softneuro import-onnx [--naive] [--extract]
                       [--preprocess_json PREPROCESS_JSON]
                       [--postprocess_json POSTPROCESS_JSON] [--help]
                       INPUT OUTPUT

Arguments

Argument	Description
INPUT	ONNX model file to be converted.
OUTPUT	Resulting DNN file path. If the extract flag is used, this is the folder where the model JSON will be saved.
--preprocess_json PREPROCESS_JSON	A json file specified when preprocess is added to DNN model.
--postprocess_json POSTPROCESS_JSON	A json file specified when postprocess is added to DNN model.

Flags

Flag	Description
--naive	Save the DNN file without FuseRelu-type optimization.
--extract	Save the model information as a JSON file.
--help	Shows the command help.

Example
After the model is converted the mobilenet_v2.dnn file will be created.

$ softneuro import-onnx mobilenet_v2.onnx mobilenet_v2.dnn
import model
converting initializers: done
converting nodes: done
rectify model
optimize model
save model

import-tensorflow¶

Convert a TensorFlow model to DNN.

Usage

softneuro import-tensorflow [-h] [--keras] [--list-keras] [--naive] [--extract] [--preprocess_json PREPROCESS_JSON] [--postprocess_json POSTPROCESS_JSON] [--output_attrs OUTPUT_ATTRS]
                           [--input_node [INPUT_NODE_NAME]] [--input_shape [INPUT_SHAPE]] [--output_node [OUTPUT_NODE_NAME]] [--fix_shape_inf]
                           INPUT OUTPUT

Arguments

Argument	Description
INPUT	Protocol buffer(.pb) or hdf5 format model file. For hdf5, not only weights but also the model structure is needed.
OUTPUT	Resulting DNN file path. If the extract flag is used, this is the folder where the model JSON will be saved.

Flags

Flag	Description
--keras	It's possible to convert a model from tf.keras.applications pretrained models by inputting the model name.
--list-keras	Show the list of avalilable model names for `--keras` option.
--naive	Save the DNN file without FuseRelu-type optimization.
--extract	Save the model information as a JSON file.
--preprocess_json PREPROCESS_JSON	A json file containing a preprocess network definition to be added to the DNN model.
--postprocess_json POSTPROCESS_JSON	A json file containing a postprocess network definition to be added to the DNN model.
--output_attrs OUTPUT_ATTRS	A JSON containing the output attributes (output labels), or a preset name.
--input_node [INPUT_NODE]	Model input node name.
--input_shape [SHAPE]	Model input shape.
--output_node [OUTPUT_NODE]	Model output node name.
--fix_shape_inf	turn off shape inference
-h, --help	Shows the command help.

Example
After the model is converted the vgg16.dnn file will be created.

softneuro import-tensorflow vgg16.pb vgg16.dnn

import-ver3¶

Convert a SoftNeuro V3 DNN model file to a SoftNeuro V5 DNN model file.

Usage
softneuro import-ver3 [--naive] [--extract] [--help] INPUT OUTPUT

Arguments

Argument	Description
INPUT	V3 DNN file.
OUTPUT	Resulting DNN file path. If the extract flag is used, this is the folder where the model JSON will be saved.

Flags

Flag	Description
--naive	Save the DNN file without FuseRelu-type optimization.
--extract	Save the model information as a JSON file.
--help	Shows the command help.

Example After the model is converted the vgg16_v5 file will be created.

softneuro import-ver3 vgg16_v3.dnn vgg16_v5.dnn
import model
rectify model
optimize model
save model