MATLAB: Does the function ‘exportONNXNetwork’ reports warnings about incompatibility while trying to convert the pre-trained yolov3 to onnx

Question

I would like to deploy the pre-trained yolov3 detector to an ONNX model for inference. The function 'exportONNXNetwork' reports warnings that some layers may change the behavior. I tried the following example to download the yolov3 pretrained network:

https://www.mathworks.com/help/vision/ug/object-detection-using-yolo-v3-deep-learning.html

After downloading the mat file in the yolov3 example, below is the step to reproduce the warning:

>> load('yolov3SqueezeNetVehicleExample_21aSPKG.mat')
>> exportONNXNetwork(detector.Network, 'test.onnx')
Warning: Cannot exactly export Layer 'featureResize2' with 'GeometricTransformMode' set to
'half-pixel' for 'OpsetVersion' '8'. Exported network may produce different results.
> In nnet.internal.cnn.onnx/ConverterForResizeLayers/toOnnx (line 57)
In nnet.internal.cnn.onnx/ConverterForNetwork/networkToGraphProto (line 114)
In nnet.internal.cnn.onnx/ConverterForNetwork/toOnnx (line 44)
In nnet.internal.cnn.onnx.exportONNXNetwork (line 36)
In exportONNXNetwork (line 38)
Warning: Cannot exactly export Layer 'featureResize2' with 'NearestRoundingMode' set to
'round' for 'OpsetVersion' '8'. Exported network may produce different results.
> In nnet.internal.cnn.onnx/ConverterForResizeLayers/toOnnx (line 60)
In nnet.internal.cnn.onnx/ConverterForNetwork/networkToGraphProto (line 114)
In nnet.internal.cnn.onnx/ConverterForNetwork/toOnnx (line 44)
In nnet.internal.cnn.onnx.exportONNXNetwork (line 36)

Is there a workaround for this problem? It seems like 'featureResize2' is not compatible with onnx format.

Answer 1

Currently, MATLAB export support is limited to opset version 9, but the parameters used in the 'resize2Dlayer '(featureResize2) of 'yolov3ObjectDetector' requires export support of opset >=10. As a workaround, you can replace the 'featureResize2' with a new 'resize2Dlayer' that can be exported even with opset version 9. Here is an example:

lgraph= layerGraph(detector.Network);
layer = resize2dLayer('OutputSize',[28,28],'Name','resize2DLayer','GeometricTransformMode','asymmetric','NearestRoundingMode','onnx-10');
lgraph = replaceLayer(lgraph,'featureResize2',layer);
net = dlnetwork(lgraph);

The above net is a yolov3 network that can be exported without warnings.

To verify the mAP, you can create a detector as below and follow the steps mentioned in Evaluate Model section of the "Object Detection Using YOLO v3 Deep Learning" example.

yolov3Detector = yolov3ObjectDetector(net, detector.ClassNames, detector.AnchorBoxes);