Clarification on make_maf function

[Jice-Zeng]

Hi Simon,
I looked into the codes and have some questions about the design of the surjective flow.
There is a function:
def make_maf(
n_dimension: int,
n_layers: Optional[int] = 5,
n_layer_dimensions: Optional[Iterable[int]] = None,
hidden_sizes: Iterable[int] = (64, 64),
activation: Callable = jax.nn.tanh,
)

You provided an example:
neural_network = make_maf(10, n_layer_dimensions=(10, 10, 5, 5, 5))
Does this mean we have a total of 5 layers, with the 1st and 2nd layers being dimensionality-preserving (keeping the dimension at 10), and the 3rd, 4th, and 5th layers being dimensionality-reducing (reducing the dimension by a factor of 2, from 10 to 5)?

If my initial dimension is 200 and I want the reduction factor to be 0.75, but only for the middle layer to be dimensionality-reducing, should the network be:
neural_network = make_maf(200, n_layer_dimensions=(200, 200, 200x0.75, 200, 200))?
I appreciate your assistance!

Best regard,

Jice

[dirmeier]

Hi Jice,

It should be

 neural_network = make_maf(200, n_layer_dimensions=(200, 200, int(200*0.75), int(200*0.75), int(200*0.75)))

The layer_dimensions represent the dimensionalities in the latent space. Once you reduce from 200 to 200*.75, the following dimensions need to be at least as big as that. Thanks for point this out! I need to make this clearer and the docu and maybe catch this n_layer_dimensions=(200, 200, 200x0.75, 200, 200) as error.

Best, Simon