How to configure a network with a large number of output neurons?

Question

Python

I will make a reservation right away that I do not use libraries. Own network. All banal classification tests work.
Essence.
I'm trying to create a network where there will be 50+ neurons on the output layer in the future.
As an activation function - Sigmoidal.
Regardless of the number of layers and neurons in the layer, the error stops at 1.3333...
In the end result, in the test run, all output neurons are zero.
Actually the question is... How to configure a network with 50+ neurons on the output layer, provided that as a result several neurons can take a value close to one, and not one of them, as in the case of a simple classifier?
UPD: I ran it on small layers: 5->10->5 - everything works.

Answer 1

Do a similar test using one of the common frameworks (keras, pytorch).
If everything works fine there, then there is a bug in your self-written library.