Channel Activation Map

Drop a NAM model and see which channels its dilated layers actually lean on — a live channel-by-layer activation-rank heatmap, computed right in your browser.

Weights inspector

Everything runs in your browser — model files never leave your machine.

Channel × dilation activation ranks

Each cell is one channel at one dilated layer, colored by how strongly that channel activates relative to the OTHER channels in its own layer (brighter = higher rank), for a fixed broadband noise probe — not by raw loudness. Hover a cell for its exact numbers; click one to select it.

Selected channel

The dilated-conv taps and 1×1 residual routing for the channel you click in the grid above — same channel, one layer deep. Cells are colored by sign and magnitude (amber = positive, blue = negative); hover a cell for its exact value.

Classic (A1) and generalized (A2) WaveNet models render here — if a file fails (e.g. an LSTM model), the message above states exactly why.

This is the read, not the edit

The Liminam app uses this same activation map to drive channel-scaling operations like myelinate — boosting or suppressing individual channels by their learned rank. This page only shows the map. See the app →