BatchMulti​ClassNonMa​xSuppressi​on - importONNXLayers/Network issue

Hi guys,

I'm currently trying to do some DeepLearning using Matlab. I saw it was possible to import ONNX model directly in Matlab for training.

So I tried to import ssd_mobilenet_v1_fpn_coco from this official model zoo tensorflow repository:

https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md

After I convert it to ONNX format (using tf2onnx) I tried to import it using "importONNXNetwork" and "importONNXLayers" (with classification as the OutputLayers). Here are the errors I get from Matlab:

The datatype of initializer 'Postprocessor/BatchMultiClassNonMaxSuppression/map/while/PadOrClipBoxList/cond/cond/pred_id/_174__cf__177:0'
('BOOL') is not supported yet.

From what I understand it seems that Matlab doesn't support this type of layers, so my question is: what can I do to solve this ? Is there a way to implement this layer or to tell matlab to delete it and replace by another ?

Thank you in advance

Answer 1

[Divya Gaddipati]

For building a custom layer, please refer to the below link:

https://www.mathworks.com/help/deeplearning/ug/define-custom-deep-learning-layers.html

And you can use the function replaceLayer to replace the layer with your custom layer

For more information on how to use replaceLayer, refer to the following link:

https://www.mathworks.com/help/deeplearning/ref/replacelayer.html

Hope this helps!

[Vincent Dufour]

Okay thank you very much for the advice I'll have a look at this and see if I can get through this.

Is there a way to import a network without certain layers (the problematic ones) and then replace them by custom layers ? Or do I need to recreate the whole network (which won't be convienent) ?

Thanks

[Veer Bhansari]

Hello,

I have followed the same procedure ( from tf2onnx to MATALB) and got the same error.

Did you get any luck with this ??

Can we import the layers which are not supported and replace them? or do we need to recreate the whole network as he mentioned ??

Thanks!!

[Vincent Dufour]

Hey Veer,

No I totally gave up on using Matlab, I thought at first it would be more efficient but they have still a long way to go.

Nevertheless if I remember well they issued a ticket to the Support department as a main matter to take care of. Maybe in further version the feature will be available.

Good luck

[Veer Bhansari]

Hi Vincent,

Yes, it definitely seems they still have a long way to go.

So Vincent, have you tried any other models from other framerwoks like Pytorch,Caffe to ONNX to MATLAB?

I am thinking of giving it a try.

Lets hope the features would be available in the further versions.

Thanks a lot!

[Vincent Dufour]

Hey,

nope I totally skipped Matlab, at first I found their feature to manage the data really cool (especially the video labelling feature) but then the problem with certain layers stopped me.

I used retinanet-examples repo from Nvidia on Github.

Cheers

[Veer Bhansari]

Hi,

Yes, the Image and Video labeller tools are handy and even I wanted to explore them. But it seems, we can't leverage them completely as of now due to these limitations.

Thanks!