Multihead Model
Explore the embeddings of the multihead model.
Default Model
Explore the embeddings of the default model.
Embeddings Viewer
Programming with Julius
This website hosts UMAP embeddings visualized in 3D space, representing activations from a trained neural network. The network predicts the in-game location (XYZ) based on Counter-Strike 2 screenshots. The activations from the penultimate layer are reduced to 3D to create the maps.
The multihead model also predicts which map the screenshot is from, resulting in a dramatically different embedding space. The default model only predicts the in-game location.
The visualization loads 10.000 images and you may experience lag while the images are loading. For the multihead model in particular, it is necessary to look at each cluster at least once before the images are cached and the visualization becomes smooth.
This website is not functional on mobile devices. For the best experience, use a desktop or laptop computer.
Controls:
- Click to enter control mode
- Press Esc to exit control mode
- WASD + Mouse to move
- Hold Shift to go slower