Mini Dino

A simplified implementation of the DINOv2 (Self-supervised Vision Transformers).

Features

• Easy-to-use DINOv2 implementation
• Image similarity computation using cosine similarity

Installation

1. Clone this repository:

git clone https://github.com/yourusername/mini_dino.git
cd mini_dino

2. Install required dependencies:

pip install torch torchvision pillow matplotlib

Usage

Run the model using:

python dino.py

Example Code

model_name = "dinov2_vits14"
dino = torch.hub.load("facebookresearch/dinov2", model_name)
model = dino_small_vit()
model = load_dino_weights(model, dino.state_dict(), N=12)

img1 = Image.open("image.png")
transforms = T.Compose([
    T.Resize((224, 224)),
    T.ToTensor()
])

img1 = transforms(img1).unsqueeze(0)
out_dino = dino(img1)

img2 = Image.open("image1.png")
img2 = transforms(img2).unsqueeze(0)

out = model(img2)
similarity = F.cosine_similarity(img1, img2)

Model Details

• Based on DINOv2 ViT-Small/14 architecture
• Input image size: 224x224
• Output: Feature embeddings for similarity comparison