SzczurekLab/ImmuVis_Swin

ImmuVis Swin

ImmuVis Swin is the Swin Transformer based ImmuVis masked autoencoder for multiplex imaging. It uses shared marker-agnostic encoding, marker-conditioned hyperkernels, a pan-marker encoder, and a marker-conditioned decoder.

• Source and installation: szczurek-lab/ImmuVis
• Model implementation: multiplex_model.modules.MultiplexAutoencoder
• Supported markers: 265, defined by the accompanying markers_tokenizer.yaml
• Latent width: 768
• Decoder outputs: mean logits and log variance for each requested marker

Usage

Install the ImmuVis source package (Python 3.12 or newer):

git clone https://github.com/szczurek-lab/ImmuVis.git
cd ImmuVis
uv sync --python 3.12

Run this in the installed environment:

from multiplex_model.modules import MultiplexAutoencoder

model = MultiplexAutoencoder.from_pretrained("SzczurekLab/ImmuVis_Swin")
tokenizer = model.marker_tokenizer

Use the exact tokenizer mapping shipped with these weights; do not renumber IDs. The tokenizer is shared across the ImmuVis model variants.

See the repository: szczurek-lab/ImmuVis for inference examples and tutorials.