cannot instantiate SegmentationKS for training

[kashefy]

Hey @chrschy, I was trying out demo for training SegmentationKS and ran into an error. The constructor checks if the model file exists and produces an error if it cannot find it. This happens here.

In order to train SegmentationKS I would still need to create an instance and it will not get past this line. Or am I missing something?

If this is indeed the case, I was thinking we could

1. pass a flag to the constructor to indicate that we're instantiating the object in "train mode" and skip this check accordingly.
2. Turn the train method into a static method or a factory method that does not need the object.

Option 1 seems simpler. What do you think?

[hagenw]

Hi @kashefy,

what exactly did you run to produce your error?

When I go to the segmentation folder of TWOEARS/examples and run demo_train_segmentation everything runs fine for me.

[kashefy]

Hi @hagenw,
I'm working off of demo_train_segmentation as well. I get the error when changing one of the KS's parameters.

A change of parameters involves re-training the KS. Changing the parameters changes the object's reqHashs attribute. The hash value determines the model filename. If you run the train script without changing the parameters you are basically re-training an existing model. In this case the constructor finds the existing model and no errors occur. I changed the nChannels parameter from 16 to 32. I think this leads to a different hash value for which a model does not exist yet.

I wanted to change nChannels to better align with the features parameters required for the IdentityKS downstream.

[hagenw]

Yes, I can recreate this. Change nChannels after running the demo (in my case from 32 to 16) results in:

>> demo_train_segmentation
Downloading file https://dev.qu.tu-berlin.de/projects/twoears-database/repository/revisions/master/raw//learned_models/SegmentationKS/DemoKS/DemoKS_models_66cc1d312cb06e04624ffd91283691a0d6d3ec028e79306c718aee17dd2b922b.mat
Error using SegmentationKS (line 168)
No trained models are available for this KS. Please ensure to run KS training first.

Error in demo_train_segmentation (line 33)
segKS = SegmentationKS(ksName, ...