DINOv1-ViT-S-16-food101

DINOv1 ViT-S/16 model trained from scratch on the Food-101 dataset using a custom training repo.

Evaluation

Weighted k-NN evaluation on Food-101 validation features:

k	Top-1 accuracy	Top-5 accuracy
5	64.09%	79.93%
10	66.38%	83.47%
20	67.59%	86.17%
100	67.81%	88.72%

Run

from transformers import AutoImageProcessor, ViTModel

processor = AutoImageProcessor.from_pretrained("mnjm/DINOv1-ViT-S-16-food101")
model = ViTModel.from_pretrained("mnjm/DINOv1-ViT-S-16-food101")
inputs = processor(images=image, return_tensors="pt")
outputs = model(**inputs)
image_features = outputs.last_hidden_state[:, 0]