voxelkit embed-preview¶
Renders your embedding matrix as a PNG heatmap and saves it to disk. Each row is one sample, each column is one dimension. Dead dimensions (no variance) show up as flat mid-grey stripes — they're immediately obvious in the image.
Usage¶
Arguments¶
| Argument | Description |
|---|---|
FILE |
Path to a .npy 2D (N_samples, D_dims) float array |
Flags¶
| Flag | Default | Description |
|---|---|---|
--output PATH |
— | Required. Output .png file path |
--max-samples INT |
256 |
Maximum number of sample rows to render. Large matrices are randomly subsampled |
Examples¶
# basic — renders up to 256 samples
voxelkit embed-preview features.npy --output heatmap.png
# render more rows for a denser view
voxelkit embed-preview features.npy --max-samples 1024 --output heatmap_full.png
What the image shows¶
The heatmap is per-column-normalised: each dimension (column) is independently scaled to 0–255, so you can see internal structure in dimensions regardless of their absolute scale.
- Bright/dark variation in a column → that dimension carries signal
- Flat uniform colour in a column → dead dimension (no variance)
- Isolated extremely bright or dark rows → potential outlier samples
Use this alongside voxelkit embed-report — the report gives you the numbers, the heatmap gives you the picture.