multigeom.encoders.reference

Image-to-response encoders f_φ : x ↦ r̂ used throughout the paper.

Each encoder is a frozen feature-extracting backbone followed by a per-neuron
Gaussian spatial-attention readout (the same readout structure across all
three encoders, as discussed in Appendix C / D of the paper).

The canonical input is a [B, 1, 64, 64] tensor in [-1, 1] (matching the
diffusion-process input convention).  Each encoder rescales / replicates
channels internally to its backbone's expected input format.

  - InceptionEncoder: [-1, 1] → 3-channel → InceptionV3 → readout
                      (torchvision's Inception weights expect [-1, 1] input)
  - ResNet50Encoder:  [-1, 1] → 3-channel → ImageNet mean/std
                      standardisation → ResNet50 → readout
  - OnOffEncoder:     [-1, 1] → [0, 1] → 1-channel → plenoptic OnOff → readout

Trainable parameters live only in the readout. Backbone parameters are frozen
(requires_grad = False), so `forward` builds no graph for them unless the input
requires grad, in which case gradients with respect to the input image flow
(the encoder contract of `multigeom.encoders.contract`, rule R5).

Readout parameter counts (n_neurons = 50):
  - Inception (C=192): 50 * (2 + 1 + 192 + 1) =  9,800
  - ResNet50  (C=256): 50 * (2 + 1 + 256 + 1) = 13,000
  - OnOff     (C=2):   50 * (2 + 1 +   2 + 1) =    300

Functions

build_encoder(cfg)	Construct an encoder from a config namespace.
build_reference(config)	A reference encoder (fresh readout) from its description, {"reference": "inception", "n_neurons": 50, ...}.

Classes

InceptionEncoder(*args, **kwargs)	Encoder of the paper: frozen InceptionV3 backbone (truncated at layer) and a Gaussian readout.
OnOffEncoder(*args, **kwargs)	Encoder of the paper: frozen ON-OFF front-end (plenoptic) and a Gaussian readout.
ResNet50Encoder(*args, **kwargs)	Encoder of the paper: frozen ResNet-50 backbone (truncated at layer) and a Gaussian readout.