Expand description
SigLIP (Sigmoid Loss for Language-Image Pre-training) Vision Transformer
Implementation based on the timm (PyTorch Image Models) SigLIP architecture. This is a standalone implementation that can be used by multimodal models like Janus.
Architecture:
- Patch embedding (Conv2d)
- Position embeddings (learned)
- Transformer blocks (self-attention + MLP with pre-norm)
- Final layer norm
- Optional attention pooling head
Structsยง
- SigLIP
Config - Configuration for SigLIP vision encoder
- SigLIP
Vision Transformer - SigLIP Vision Transformer