oxford-pets-3class-vit

oxford-pets-3class-vit

This model is a fine-tuned version of google/vit-base-patch16-224-in21k for a simplified pet image classification task.

It was trained on a custom 3-class subset of the Oxford-IIIT Pet dataset with the following classes:

• Egyptian Mau
• leonberger
• samoyed

Model description

This is a transfer learning model created for an educational computer vision project.

The goal of the project was to compare:

• a fine-tuned ViT model
• a zero-shot CLIP model
• a closed-source OpenAI vision model

The model is designed to classify images into one of the three selected pet classes.

Intended uses & limitations

Intended use

This model is intended for:

• educational use
• demonstration of transfer learning
• comparison against CLIP and OpenAI vision models
• classification of images belonging to the selected 3 classes

Limitations

This model has important limitations:

• it was trained on a very small dataset
• it only supports 3 classes
• it is not suitable for real-world production use
• predictions on unrelated animals or unseen categories may be unreliable

Training and evaluation data

Dataset source

• Hugging Face dataset loader: load_dataset("pcuenq/oxford-pets")

Dataset used in this project

A custom subset was created from the Oxford-IIIT Pet dataset.

Selected classes:

• Egyptian Mau
• leonberger
• samoyed

Dataset size

• Total images: 90
• Train: 60 images total (20 per class)
• Validation: 15 images total (5 per class)
• Test: 15 images total (5 per class)

Preprocessing

Training transforms

• Random resized crop
• Random horizontal flip
• Conversion to tensor
• Normalization with ViT image processor values

Validation / test transforms

• Resize
• Center crop
• Conversion to tensor
• Normalization with ViT image processor values

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-5
• train_batch_size: 8
• eval_batch_size: 8
• num_epochs: 5
• optimizer: AdamW (Trainer default)
• best model selected using validation accuracy

Training results

Final evaluation

• Validation accuracy: 1.0
• Test accuracy: 1.0

Because the task was simplified to only 3 classes and the dataset is small, the model performs very well on this limited setup.

Example prediction behavior

Example: leonberger image

• ViT model: leonberger
• CLIP: leonberger
• OpenAI: leonberger

Example: Egyptian Mau image

• ViT model: Egyptian Mau
• CLIP: Egyptian Mau
• OpenAI: Egyptian Mau

Related resources

• App: https://huggingface.co/spaces/vasanthi8134/pet-classification-comparison
• Model repo: https://huggingface.co/vasanthi8134/oxford-pets-3class-vit