andrew-pathology encoders
Two prostate-histopathology encoders that produce the pre-extracted features used by the downstream MIL tasks in the andrew-prostate-pathology repo:
| model | role | arch | derived from | files |
|---|---|---|---|---|
| andrew-tile | tile / patch encoder | ViT-H/14 (DINOv2-trained) | paige-ai/Virchow | andrew-tile/ |
| andrew-slide | slide encoder | PRISM (CoCa-style) | paige-ai/Prism | andrew-slide/ |
Pipeline: each tile → andrew-tile → concat([CLS, mean(patch tokens)]) = 2560-dim
embedding → andrew-slide aggregates the tile embeddings into one slide-level vector.
Repository layout
andrew-tile/
model.safetensors # ViT-H/14 weights (timm format)
config.json # timm model config
andrew-slide/
andrew-slide.model.pth # fine-tuned PRISM state_dict
Usage
andrew-tile (patch encoder)
import timm, torch
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file
cfg = hf_hub_download("EstelleXIA/andrew", "andrew-tile/config.json")
wts = hf_hub_download("EstelleXIA/andrew", "andrew-tile/model.safetensors")
model = timm.create_model("vit_huge_patch14_224", pretrained=False,
img_size=224, patch_size=14, embed_dim=1280, depth=32,
num_heads=16, mlp_ratio=5.3375, global_pool="",
num_classes=0, reg_tokens=0, init_values=1e-5)
model.load_state_dict(load_file(wts))
model.eval()
# x: [B, 3, 224, 224], normalized with mean=std=0.5
with torch.inference_mode():
tokens = model(x) # [B, 257, 1280]
emb = torch.cat([tokens[:, 0], tokens[:, 1:].mean(dim=1)], dim=-1) # [B, 2560]
andrew-slide (slide encoder)
import torch
from huggingface_hub import hf_hub_download
from transformers import AutoModel
model = AutoModel.from_pretrained("paige-ai/Prism", trust_remote_code=True)
ckpt = hf_hub_download("EstelleXIA/andrew", "andrew-slide/andrew-slide.model.pth")
state = torch.load(ckpt, map_location="cpu")
state = state.get("model_state_dict", state)
model.load_state_dict(state, strict=False)
model.eval()
License
Released for non-commercial academic research only, as derivatives of paige-ai/Virchow and paige-ai/Prism. You must comply with the original Virchow and PRISM licenses. No commercial use; no redistribution of the weights.
Citation
If you use these weights, please cite the andrew-prostate-pathology repository and the original Virchow / PRISM / DINOv2 works.