I want a model that outputs the pixel coordinates of the tip of my forefinger, and whether it's touching something or not. Those would be 3 output neurons: 2 for the X-Y coordinates and 1, with a sigmoid activation, wich predicts the probability whether it's touching or not.

What do I need to change in the squeezenet model in order to do this?

(PS: the trained model needs to be the fastest possible (in latency), that's why I wanted to use SqueezeNet)

  • $\begingroup$ Can you clarify why you why need 3 output neurons? Also, why would one have the sigmoid? $\endgroup$
    – nbro
    May 6, 2020 at 1:29
  • $\begingroup$ In general, any architecture can be combined with different "heads". Yes, you could take an existing model and use a different output layer. Instead of predicting a vector with 3 elements, it would make sense to predict 1) 2 elements using a regression loss like MSE 2) another scalar with sigmoid using a binary cross entropy loss. That would be a multi-task learning setup. $\endgroup$ May 6, 2020 at 7:54


You must log in to answer this question.