How to export trained Faster RCNN to another hardware platform

Question

Alberto Tellaeche on 19 Oct 2019

0
Link

Direct link to this question

https://se.mathworks.com/matlabcentral/answers/486299-how-to-export-trained-faster-rcnn-to-another-hardware-platform

Commented: Abdussalam Elhanashi on 12 Feb 2020

Hello all,

I have used MATLAB with the deep learning toolbox to train my own Faster RCNN object detector.

Either if I use a predefined CNN (squeezenet for example) or my designed CNN, I want to export the trained Faster RCNN to use it on other embedded platforms.

exportONNXNetwork gives errors with Faster RCNN architectures no matter the CNN used.

How can I export my work to use it in another HW? I have worked many hours and now I can not deploy my design !!

Thank you all in advance,

Alberto

3 Comments
Show 1 older commentHide 1 older comment

Alberto Tellaeche on 24 Oct 2019

Open in MATLAB Online

Hi,

I can provide an example. Ihave a fasterRCNNObjectDetector trained using the squeezenet network model.

Here is the output of MATLAB 2019b when trying to export. It only happens with the faster RCNN models. It seems to work with FastRCNN or RCNN.

Thank you all in advance,

Alberto

>> load('ResultsFasterRCNN_squeezenet.mat')
>> exportONNXNetwork(detector.Network,'test.onnx')
Warning: ONNX does not support layer 'nnet.cnn.layer.RPNSoftmaxLayer'. Exporting to ONNX operator
'com.MathWorks.Placeholder'. 
> In nnet.internal.cnn.onnx/NNTLayerConverter/makeLayerConverter (line 212)
  In nnet.internal.cnn.onnx/ConverterForNetwork/networkToGraphProto (line 100)
  In nnet.internal.cnn.onnx/ConverterForNetwork/toOnnx (line 44)
  In nnet.internal.cnn.onnx.exportONNXNetwork (line 35)
  In exportONNXNetwork (line 40) 
Warning: ONNX does not support layer 'nnet.cnn.layer.RPNClassificationLayer'. Exporting to ONNX operator
'com.MathWorks.Placeholder'. 
> In nnet.internal.cnn.onnx/NNTLayerConverter/makeLayerConverter (line 212)
  In nnet.internal.cnn.onnx/ConverterForNetwork/networkToGraphProto (line 100)
  In nnet.internal.cnn.onnx/ConverterForNetwork/toOnnx (line 44)
  In nnet.internal.cnn.onnx.exportONNXNetwork (line 35)
  In exportONNXNetwork (line 40) 
Warning: ONNX does not support layer 'nnet.cnn.layer.RegionProposalLayer'. Exporting to ONNX operator
'com.MathWorks.Placeholder'. 
> In nnet.internal.cnn.onnx/NNTLayerConverter/makeLayerConverter (line 212)
  In nnet.internal.cnn.onnx/ConverterForNetwork/networkToGraphProto (line 100)
  In nnet.internal.cnn.onnx/ConverterForNetwork/toOnnx (line 44)
  In nnet.internal.cnn.onnx.exportONNXNetwork (line 35)
  In exportONNXNetwork (line 40) 
Error using nnet.internal.cnn.onnx.ConverterForClassificationOutputLayer/toOnnx (line 28)
Assertion failed.
Error in nnet.internal.cnn.onnx.ConverterForNetwork/networkToGraphProto (line 102)
                    = toOnnx(layerConverter, nodeProtos, TensorNameMap, TensorLayoutMap);
Error in nnet.internal.cnn.onnx.ConverterForNetwork/toOnnx (line 44)
            modelProto.graph            = networkToGraphProto(this);
Error in nnet.internal.cnn.onnx.exportONNXNetwork (line 35)
modelProto  = toOnnx(converter);
Error in exportONNXNetwork (line 40)
nnet.internal.cnn.onnx.exportONNXNetwork(Network, filename, varargin{:});
 
>> 