README.md

metadata

license: apache-2.0
tags:
  - vit
  - image-classification
  - beans
  - transfer-learning

ViT Beans Model

This model was fine-tuned using transfer learning on the "beans" dataset from the Hugging Face Datasets Hub.
It classifies bean plant leaves into the following categories:

• LABEL_0: angular_leaf_spot
• LABEL_1: bean_rust
• LABEL_2: healthy

Model architecture

The base model is google/vit-base-patch16-224.

Training

Transfer learning was used with a ViT model pre-trained on ImageNet-21k.

Evaluation

This model was compared to a zero-shot classification using CLIP (openai/clip-vit-base-patch32).

Zero-Shot Results on Oxford Pets (as required):

• Accuracy: 0.9993189573287964
• Precision: 0.5794700118713081
• Recall: 0.10156987264053896
• Model used: openai/clip-vit-base-patch32

Example

from transformers import ViTFeatureExtractor, ViTForImageClassification
from PIL import Image
import torch

image = Image.open("example_input.png")
extractor = ViTFeatureExtractor.from_pretrained("LindiSimon/vit-beans-model")
inputs = extractor(images=image, return_tensors="pt")
model = ViTForImageClassification.from_pretrained("LindiSimon/vit-beans-model")
with torch.no_grad():
    logits = model(**inputs).logits
predicted_class = logits.argmax(-1).item()