I am using a normalizing flow (Neural Spline Flows) to approximate a probability and after some training, the average loss is around 0.5 (so logprob = -0.5). However, when I am trying it on some new test data, I am getting some values of logprob bigger than zero, which would mean that the probability for that element is bigger than one (which doesn't make sense). Does anyone know what could cause this? Isn't the flow supposed to keep all the probabilities below 1 automatically? Thank you!
I have always seen the negative odd-logs ration when doing similar work. As it is now, I don't think you are constraining the splines appropriately Link to reading