Deep Learning error message

[dschatt]

hi, I know this is a long shot but I'm trying to do an instance segmentation analysis to detect shoreline docks in aerial imagery using Deep Learning in Pro.  I am going through the standard steps. as I understand them, for a deep learning analysis. 

First I  created vector polygon training samples around my feature outlines using the Label Objects for Deep Learning tool so I have those in a feature class.  I then use Export Training Data for Deep Learning, setting the "Format" to RCNN-Masks, and output a folder with all the image chips and everything looks fine.

I then try to train a model from that folder using Train Deep Learning Model and I'm setting the Model Type to MaskRCNN.  But I get an error message immediately after starting this tool: "Train Deep Learning model failed'.  No idea why, thanks for any tips..

Dan

[David_McRitchie]

Hey dschatt,

With Deep Learning there are a lot of parameters that feed into the workflow and a lot of things that can go wrong within this. To help we would need to know what documentation or videos you have used to build a bit of a better idea.

Does the error window show any further detailed error messages?

Do you have the Deep Learning Essentials package installed on the ArcGIS Pro Python environment?

Also as a note you might not need to carry out training on a Deep Learning Model. The Living Atlas has a range of pre-trained models that can be downloaded and utilised within ArcGIS Pro.

Hope that helps!

David

[dschatt]

thanks much David, I definitely have all the DL libraries, etc. downloaded and installed successfully because I've already run many Deep Learning operations successfully using the object detection tools (as opposed to the instance segmentation).  So I should have all of that unless there is something more needed for instance segmentation that I'm not aware of.   I don't think the pre-trained models will help because this is a very specialized application (detecting shoreline docks).

I am using the workflow of a YouTube video online for instance segmentation in which he basically followed the steps I described above:

Mask R-CNN Instance Segmentation and Object Detection in ArcGIS Pro 

The only difference is the initial step to define training areas: he had a pre-existing feature class to define his training areas but I created my own feature class using the Label Features for Deep Learning tool.  But the procedure after that  (using "Export Training Data for Deep Learning" and then "Train Deep Learning Model") were the same.

I've attached screen shots of my Export Training Data for Deep Learning tool input and also my Train Deep Learning Model tool input and the resulting error.  I know there is a lot here but I'm just hoping there is something very obvious in the procedure or settings that I'm missing. I don't really know how to interpret the error but maybe it's meaningful to you?

Thanks again.

[PavanYadav]

Hi @dschatt , I understand you would like to do an instance segmentation. If it’s just instance segmentation, you might want to consider using ready to use Segment Anything pre-trained model.

You are using MaskRCNN so I am curious if you have two or more classes? When using MaskRCNN in training and getting Error “local variable 'data_bunch' referenced before assignment”, can be due to the number of classes. For a test, please have them not create classvalue = 0 (anything >0 will work) and have at least two classes.

If you have just one class, maybe you can use pixel classification approach. I am not advocating one of the other, I am just sharing my thoughts. Hope this helps.

Note: We have an Arcgis Image Analyst dedicated community and here you might get response sooner.

Cheers!

Pavan Yadav | Product Engineer 
Esri | 380 New York | Redlands, 92373 | USA

https://www.linkedin.com/in/pavan-yadav-1846606/ 

[dschatt]

Thanks Pavan, I was able to resolve this with a simple fix with the help of one of your colleagues at ESRI.  I did only have one class but when naming the class during that Label Objects for Deep Learning step, I had used two words with a space between.  When I eliminated the space and had just one word as a name, everything worked fine 😊