multigeom.diffusion.config

Configuration of the conditional diffusion model of an encoder.

The diffusion model that G(x) needs is conditioned on the responses of one encoder, so
one model is trained per encoder. Everything in its configuration is shared (UNet,
scheduler, training recipe) except the width of the first layer of the conditioning
network, which is the number of neurons of the encoder: `DiffusionConfig.for_encoder`
sets it.

Functions

encoder_n_neurons(encoder[, image_shape])

The number of neurons (output width) of an encoder: its n_neurons if it has one, else the width of its output.

Classes

DiffusionConfig()

Namespace of the diffusion-model configuration (multigeom/configs/diffusion.yaml) and its per-encoder version.