Is there a parametrized logistic function?


I’m designing a neural network and one part of it’s a selector: depending of one value p between 0 and 1 choose between A or B.
To do so I was planning to use a natural generalization of the sigmoid function that just multiplies the input by a scalar k, that is, sigmoid

Unfortunately I haven’t found something like that in the API and I was wondering if there is any reason for that? I was about to post an issue in github but I’ve decided to ask here first.


I think the sigmoid_k function is not available since, it seems it’s a one-liner:

def sigmoid_k(x, k):
    return (1. / (1. + torch.exp(-k * x)))

x = torch.arange(-10, 10) 
k = 2

sigmoid_k(x, k)

I haven’t tested it thoroughly, so let me know, if you encounter any issues.

Yes, I already have it implemented in my code. But I was curious about why is not already implemented. I mean, the sigmoid is in the API and is also a one-liner.

But playing a bit with your code I think I might found it out. When k grows calculating the gradient may be way too much since it the function resembles the impulse function (with gradient 0 everywhere but in one point that “has infinite” gradient).

Thanks for the answer though

I assume it’s not widely used, but if you think this method should be implemented as part of the API, you could create a feature request on GitHub. :wink:

1 Like
torch.sigmoid(k * x) ?
1 Like

hahahaha yeah, I’m dumb.

Hahaha, I’ve just written down the code without realizing the actual function. Maybe I’m just too tired.
Thanks @agadetsky for preventing more embarrassment :smiley: