Introducing darkspark, a gui for your neural network. It traces your pytorch code and brings up a visual representation for you to interact with. We have a hosted gallery of popular model architectures pre-traced and ready to explore. Here’s stable-diffusion-v1.5

We’re also capturing activations during the trace. Check out the simple color and line detectors in the first layer of convnext_zepto from the timm library

It’s like openai’s Microscope tool but for your own models. See if you can identify these later features from timm’s contrastively-trained vit_base_patch16_siglip_224

We can also explore the attention patterns. The siglip model from above has a head that appears to move information around hamburgers

We’ve got all the models from the timm library as well as many others from huggingface’s transformers and diffusers libraries already hosted.

The github repo isn’t yet ready for public release but you can explore all the pre-traced models at Works best with a large screen.

Or start with a few neurons from DINO-v2

Very cool—I wrote a package (TorchLens) that does visuals for DNNs but without the interactivity. Lemme know if you ever want to compare notes!

yes torchlens was an inspiration for me!

sd1.5 having 7.7b params seems very wrong

why?