Fail to compose a single predictive model

[lachioma]

Hi
Thank you very much for providing a Matlab implementation of the SLEAP inference!
I am trying to run the demo code on a centroid+topdown models that I trained on my data, but I get the following error:

>> modelNet = sleap.model.topdown.compose(centroidNet, topdownNet, inputScale, bboxSize, minThreshold, centroidConfig, topdownConfig);
Error using  == 
Matrix dimensions must agree.

Error in sleap.model.topdown.compose (line 36)
sameNames = tdNames==centroidNames;

This error clearly occurs because the centroid and topdown models have a different number of layers.

Thank you for any help!
-alessandro

[talmo]

Hi @lachioma,

Thanks for reporting this! This seems like a bug to me. Specifically, this part of the compose function attempts to deal with having layer names that are the same between the centroid and centered instance models:

sleap-matlab/+sleap/+model/+topdown/compose.m

Lines 31 to 39 in cf0f813

	% Since array layer names for both centroid and topdown models have
	% commonalities, create unique names for topdown model inorder to combine
	% all into a single model
	centroidNames = arrayfun(@(x) string(x.Name), net.Layers);
	tdNames = arrayfun(@(x) string(x.Name), topdownNet.Layers);
	sameNames = tdNames==centroidNames;
	newtdNames = tdNames;
	newtdNames(sameNames) = matlab.lang.makeUniqueStrings(centroidNames(sameNames), tdNames(sameNames));
	larrayTd = topdownNet.Layers;

This is necessary MATLAB can't have layers with the same name within the same DAGNetwork (while this is handled via scoping in TensorFlow). The names are the same because we use the same general architecture (UNet) for the backbone, though they might be configured differently, resulting in different variants which have different numbers of layers.

I think the fix here will be to change up the logic in that snippet above to use ismember or string contains or intersect to find the repeated layer names and only rename those.

Maybe an even easier solution that mimics what TensorFlow does is to add a scoping prefix to all layers from each model type.

We'll work on a fix and report back here when we have a solution.