Launching today
FeatLens
Visualize feature maps from any vision model effortlessly.
11 followers
Visualize feature maps from any vision model effortlessly.
11 followers
Model-agnostic feature-map visualization: PCA-to-RGB feature maps from any vision model and any layer.
How does this handle really deep models like a 50 layer resnet without choking on memory, and can I export the visualizations as a video to watch how features evolve across layers?
@sudenaz3r0b
Memory: it only hooks the layers you ask for and stores them detached (no grads, eval mode), so cost scales with the number of selected layers, not model depth.
Layer-evolution video: not built-in yet. video() animates frames, not layers. But visualize(..., return_data=True) gives you the per-layer RGB maps, so a "layers-as-frames" GIF is going to be implemented in future
curious how this handles models that aren't trained on imagenet, does the PCA still produce something visually meaningful or do you need to tweak the preprocessing per backbone
@harunbalkabak PCA here is unsupervised and per-image. It fits the model's own feature distribution for a single image, so it has nothing to do with ImageNet classes or labels. Any backbone with spatially coherent features (SSL, CLIP/SigLIP, MAE, V-JEPA, SAM, even a satellite/medical model via the escape hatch) produces something meaningful; the robust PCA (MAD outlier filtering) keeps a few extreme patches from washing out the colors.
the per-channel intensity scaling in the PCA-to-RGB output looks really well-tuned, colors stay distinct without blowing out even on deeper layers.