Generate Image Embeddings with CXR-DINO-V1

1. Installation

pip install torch torchvision pillow
pip install lightly-train

2. Embed a Folder of Images (CLI)

lightly-train embed \
  data="images/" \
  checkpoint="/cxr-dino-model/model.pt" \
  out="embeddings.pth" \
  format="torch"

• data → folder containing images
• checkpoint → your cxr-dino model
• out → output file with embeddings

3. Minimal Python Example

import lightly_train

lightly_train.embed(
    data="images/",
    checkpoint="cxr-dino-model/model.pt",
    out="embeddings.pth",
    format="torch"
)

print("Saved embeddings to embeddings.pth")

4. Load Embeddings

import torch

data = torch.load("embeddings.pth")
embeddings = data["embeddings"]
paths = data["image_paths"]

print(embeddings.shape)
print(paths[:5])