Future Support for 'Acceleration','mex' When Running ONNX Networks - RF-DETR

I am working with an ONNX model imported into MATLAB using importNetworkFromONNX.

The model is a RF-DETR object detection network exported from Python.

I would like to use MEX acceleration for inference, but it currently does not seem to be supported for this workflow.

oNet = importNetworkFromONNX("rfdetr.onnx");

sizeImg = 384;

% Load an image

tImg = imread("img.png");

tImgSmall = imresize(tImg,[sizeImg sizeImg]);

mImg = single(tImgSmall);

dlImg = dlarray(mImg,"SSC");

mGPUImg = gpuArray(dlImg);

[dlBoxes, dlLabels] = oNet.predict(mGPUImg, 'Acceleration', 'mex');

I get this error message:

error using dlnetwork/predict (line 700)

Option 'Acceleration', 'mex' is not supported for this network. Use option 'Acceleration', 'auto' instead.

Caused by:

Error using nnet.internal.cnn.coder.MexNetworkConfig/mustBeSupportedNetwork

Layer hyper-parameters for custom layer 'ConstantOfShape_To_ReshapeLayer1102' must be numeric scalar, scalar logical, character or string array, or a matrix of type double or single.

in the future it will be possible that you will support this network?

thanks

[Ram Kokku]

Hi Ilan,

MEX acceleration uses code generation internally. It appears that the layer 'ConstantOfShape_To_ReshapeLayer1102' contains unsupported constructs. Use analyzeNetworkForCodegen(dlnet, TargetLibrary = 'cudnn') to find more details about this error and other blocking issues. While replacing these constructs with supported ones could be a workaround, it may not be straightforward. Since your goal is to improve model inference in MATLAB ( based on : https://www.mathworks.com/matlabcentral/answers/2181894-onnx-model-import-and-inference-performance-in-matlab-is-significantly-slower-than-python-cpu-gpu?s_tid=prof_contriblnk), I recommend waiting for a response to that thread before modifying the model.