error Loading onnx model exported from pytorch to matlab

Hi!

I trained a neural network regression (numerical input/output data) model in Pytorch and I want to load it to MATLAB.

I followed the instructions to do so, firstly by converting the pretrained network from Pytorch to .onnx format.

However when I try to load it to MATLAB environment using the MATLAB Deep Learning Toolbox Converter for ONNX Model Format

network = importONNXNetwork('model.onnx', 'OutputLayerType', 'regression')

, I get the following error:

Only image inputs are supported for import

Does this toolbox converter for onnx, just support models with image inputs?

Answer 1

[Shashank Gupta]

Hi Yasser,

Yeah right now “importONNXNetwork” only support 3D-sequence-batch tensor or 4D image batch tensor. you can perhaps recode the 2D input as a 4D image batch tensor. In pytorch, an image batch tensor has the layout [batch, channel, height, width]. If the [h,w] input tensor could be recoded as a [h,w,1,1] tensor in pytorch, then that might be importable into a MATLAB “imageInputLayer”. (Or, you may need to omit the batch dimension in pytorch and declare it to be [h,w,1] )

I hope this helps.

[Jan Drgona]

Hi Shashank,

Does the same apply to importKerasNetwork?

If so is there any other straightforward way to import standard FCNN from pytorch or tf?

Thanks,

Jan

[Richard Crozier]

is this a regression, pretty sure I was able to previously import non image based models using this tool?

Answer 2

[Sivylla Paraskevopoulou]

Since R2021a, both the importONNXNetwork and importTensorFlowNetwork functions can import a TensorFlow/ONNX input layer or tensor as a featureInputLayer.

For more information on how importONNXNetwork interprets the data format of the ONNX network's input tensors and converts them into MATLAB input layers, see Conversion of ONNX Input and Output Tensors into Built-In MATLAB Layers.

For more information on how importTensorFlowNetwork converts the Input TensorFlow-Keras layer into a MATLAB layer, see TensorFlow-Keras Layers Supported for Conversion into Built-In MATLAB Layers.