LAST-ViT ViT-B/16 190k (Label-Supervised)

Vision Transformer backbone with LAST-ViT pooling, fine-tuned for ImageNet-1K classification (1000 classes). Converted to HuggingFace Transformers format from the LAST-ViT ViT_190k checkpoint.

Model Details

• Architecture: ViT-B/16 (Vision Transformer Base, patch size 16×16)
• Training scenario: Label-supervised (ImageNet-1K)
• Hidden size: 768
• Num layers: 12
• Num attention heads: 12
• Image size: 224×224
• Num classes: 1000 (ImageNet-1K)

Usage

from transformers import ViTForImageClassification, ViTImageProcessor
from PIL import Image

model = ViTForImageClassification.from_pretrained("BiliSakura/last-vit-vitb16-190k")
processor = ViTImageProcessor.from_pretrained("BiliSakura/last-vit-vitb16-190k")

image = Image.open("your_image.jpg").convert("RGB")
inputs = processor(images=image, return_tensors="pt")
outputs = model(**inputs)
logits = outputs.logits
predicted_class_idx = logits.argmax(-1).item()

References

• Repository: ChengShiest/LAST-ViT
• Training script: cls_pretrain
• Original weights: ViT_190k.pth