Search code examples
tensorflow-probability

In tensorflow probabiliy, is there a vanilla "Dense" bijector?

For example, consider a bunch of layers "a traditional keras model". One might put a prior on the weights, a prior on the data in or out. If the Dense network is a bijector, a lot of stuff is handled magically. Looks like Affine and some other layers will allow one to build this manually. There is DenseVariational but that uses a surogate posterior.

Solution

  • A traditional Dense layer is not bijective. tfb.Affine with a scale_tril parameterization (and optionally a shift to act like a bias) is the bijective equivalent. Can you say more about what you're trying to achieve?