This will not address the case when y=1., where we will still see infinity. How about we clamp this using

clamped = torch.clamp(y.reciprocal() + (1 - y).reciprocal(), max=torch.finfo(y.dtype).max)
return -clamped.log()

@neerajprad It looks good ! I've tried a testing, it seems when std from Normal distribution is large, there is still possibliity to get 'inf, e.g.

d = TransformedDistribution(Normal(torch.tensor(0.0), torch.tensor(100.0)), [StableSigmoidTransform()])
x = d.sample()
print(x)
d.log_prob(x)
>>> tensor(1.)
>>> tensor(-inf)

Yes, that would just fix the jacobian term, but in TransformedDistribution.log_prob we still evaluate base_dist.log_prob(y) and in this case y = inf. I think what you really want is something like a ClippedSigmoidTransform which differs from SigmoidTransform in that it clips the output of _call to lie in (0, 1):

    def _call(self, x):
        finfo = torch.finfo(x.dtype)
        clamped = torch.clamp(torch.sigmoid(x), min=finfo.tiny, max=1. - finfo.eps)
        return clamped

Then you shouldn't see any infs in your example above.

@alicanb, @fritzo - Do you think this is something we should do by default (tensorflow doesn't seem to do this either), or have a separate transform class ClippedSigmoidTransform?

@neerajprad @alicanb @fritzo Thanks a lot for the proposal. I think this is rather important for some applications where we evaluate log_prob of a sample drawn from a transformed distribution that is conditional on the input of the NN. Without such fixes, it is easily happen to get NaN/Inf. An example use case is tanh transformed policy network in reinforcement learning context, e.g. SAC algorithm.

Or this can be provided with an optionally flag in SigmoidTransform, without creating a new class.

I think that’s a good idea - we can have an optional keyword arg, clip=False.

Sorry, I'm late to the party. @neerajprad your suggestion sounds great, do you think users would want to change clip limits (@zuoxingdong, have you ever had the need?) ? If that's the case, instead of a bool arg, we can keep the clamping symmetric (min=lim, max=1-lim) and have lim as the argument. If not, then bool arg sound great

Nice! I think this is the best way to handle the numerical stability issue.

This change looks great, but just to confirm, this still doesn't handle the TransformedDistribution case you mentioned?

This is so nice!!!!!!!!!!!!!!!!!!!! A great way to address the overflow/underflow while computing y.