Learnable Channel Selection

Hyperspectral cameras often deliver 60–100 bands, but only a handful carry the signal you actually care about. The Channel Selector node learns which bands to keep using a two-phase training schedule: statistical warm-up followed by gradient-based refinement.

This tutorial walks through configuring and training a channel selector, then inspecting which wavelengths survived.

Run the example:

• cuvis_ai/configs/trainrun/channel_selector.yaml — trainrun config, run with restore-trainrun --mode train
• notebooks/use_cases/channel_selector_lentils.ipynb: gradient-trained variant in cuvis-ai, a Gumbel-Softmax selector learned through a frozen AdaCLIP detector on the lentils dataset.

What you'll learn

• How two-phase training combines statistical and gradient stages.
• Configuring K (number of channels to keep) and the temperature schedule.
• Reading the final channel weights to recover the selected wavelengths.

When to reach for channel selection

• You want to ship a downstream model that operates on far fewer channels than the camera delivers (smaller, faster, deployable).
• You suspect only a subset of bands carry your signal and want the model to discover them.
• You need a learnable, end-to-end alternative to hand-picking bands.