How to do do multi-label classification

[antbertrand]

Hi all,

I'm building a system to analyse the behaviors of a mouse in its cage. I am looking at 6 different behaviors, that are not always exclusive. For example a mouse can be eating and rearing at the same time (rearing meaning it is on its back legs).

I've trained a I3D network on the task. But I can't seem to find a way for the different classes to be considered independent. As you can see on my attached picture, the probabilities sum to 1 when I would like them to be independent for each class. Typically this is done by using a sigmoid function and not softmax on the predictions. Would you have any idea where and how I could change this ?

Thanks a lot for your help,

[dreamerlin]

Sorry for the late reply. You can refer to this

[Grutschus]

Hi,
I'd be interested in the same problem for my research. Unfortunately, your reference is no longer available. Is there a newer reference?

Thanks a lot for your help!

[rlleshi]

Check for that tsn config file in the 0.X branch.

[Grutschus]

Thanks, really appreciate your reply
For everyone else, here is a permalink to the file: tsn_r101_1x1x5_50e_mmit_rgb.py