🐶 vit-base-oxford-iiit-pets

A fine-tuned Vision Transformer (ViT-Base) model on the Oxford-IIIT Pet Dataset for fine-grained image classification of 37 cat and dog breeds.

---

📈 Evaluation Results

• Final Validation Loss: 0.2496
• Final Accuracy: 93.10%

Training Loss	Epoch	Step	Validation Loss	Accuracy
1.7389	1.0	185	0.4195	0.8999
0.3005	2.0	370	0.2997	0.9161
0.2178	3.0	555	0.2647	0.9256
0.1928	4.0	740	0.2533	0.9323
0.1734	5.0	925	0.2496	0.9310

---

🧠 Model Description

• Architecture: Vision Transformer (ViT-Base)
• Patch Size: 16x16
• Input Resolution: 224x224
• Classes: 37 (cat and dog breeds)
• Base Model: google/vit-base-patch16-224
• Trained on: Oxford-IIIT Pet Dataset

---

📦 Intended Use

• Pet breed classification (cats & dogs)
• Transfer learning on similar classification tasks
• Educational and benchmarking use for ViT models

---

⚠️ Limitations

• Only trained on pet images — may not generalize to wild animals
• Not optimized for real-time or mobile deployment
• Sensitive to poor image quality or unusual aspect ratios

---

🧪 Dataset

• Name: Oxford-IIIT Pet Dataset
• Images: 7,349 total
• Classes: 37 pet breeds
• Split: Random train-validation (80/20)

---

🛠️ Training Details

Hyperparameters

• learning_rate: 3e-4
• train_batch_size: 32
• eval_batch_size: 16
• num_train_epochs: 5
• optimizer: AdamW (betas=(0.9, 0.999), epsilon=1e-08)
• lr_scheduler_type: linear
• seed: 42

Framework Versions

• Transformers: 4.51.3
• PyTorch: 2.6.0+cu124
• Datasets: 3.6.0
• Tokenizers: 0.21.1

---

🚀 How to Use

from transformers import AutoFeatureExtractor, AutoModelForImageClassification
from PIL import Image
import torch

# Load model and feature extractor
model = AutoModelForImageClassification.from_pretrained("rakib730/vit-base-oxford-iiit-pets")
processor = AutoFeatureExtractor.from_pretrained("rakib730/vit-base-oxford-iiit-pets")

# Load an image
image = Image.open("your_image.jpg")
inputs = processor(images=image, return_tensors="pt")

# Predict
with torch.no_grad():
    outputs = model(**inputs)
    logits = outputs.logits
    predicted_label = logits.argmax(-1).item()

print("Predicted class:", model.config.id2label[predicted_label])