LAST-ViT DINO ViT-B/16 (Self-Supervised)

Vision Transformer backbone with LAST-ViT pooling, pre-trained with DINO (self-supervised). Converted to HuggingFace Transformers format from the LAST-ViT dino0080 checkpoint.

Use this model as a feature extractor for downstream vision tasks (classification heads, detection, segmentation).

Model Details

• Architecture: ViT-B/16 (Vision Transformer Base, patch size 16×16)
• Training scenario: Self-supervised (DINO)
• Hidden size: 768
• Num layers: 12
• Num attention heads: 12
• Image size: 224×224

Usage

from transformers import ViTModel, ViTImageProcessor
from PIL import Image

model = ViTModel.from_pretrained("BiliSakura/last-vit-dino-vitb16")
processor = ViTImageProcessor.from_pretrained("BiliSakura/last-vit-dino-vitb16")

image = Image.open("your_image.jpg").convert("RGB")
inputs = processor(images=image, return_tensors="pt")
outputs = model(**inputs)

# Pooled output (CLS token)
pooled = outputs.last_hidden_state[:, 0, :]  # (batch, 768)
# Or all patch tokens
patch_tokens = outputs.last_hidden_state[:, 1:, :]  # (batch, 196, 768)

References

• Repository: ChengShiest/LAST-ViT
• DINO training: facebookresearch/dino
• Original weights: dino0080.pth