maceythm/vit-90-animals

vit-90-animals

Model description

This model is a fine-tuned Vision Transformer version of google/vit-base-patch16-224 on the animal image dataset from kaggle - trained to classify images into 90 different animal species. It achieves high accuracy on unseen data and was trained using supervised learning. The model can be used for general-purpose image classification in the animal domain and serves as a comparison baseline for zero-shot classification models such as CLIP.

The model achieves the following results on the evaluation set:

• —Loss: 0.0840
• —Accuracy: 0.9796

Intended uses & limitations

Intended uses

• —Animal image classification (educational, demo, prototyping)
• —Benchmarking against zero-shot classification models
• —Use in Gradio interfaces or image analysis tools

Limitations

• —The model is limited to the 90 animal classes it was trained on
• —It may not generalize well to image domains outside of its training distribution
• —Performance can degrade with poor image quality or occlusions

Training and evaluation data

The model was trained on a dataset containing 5,400 animal images categorized into 90 distinct classes. The dataset was obtained from Kaggle and according to the creator originally sourced from Google Images. The training/validation/test split was 80/10/10, and the label distribution is relatively balanced across classes.

Evaluation was conducted on the test split and compared to results from a zero-shot model (openai/clip-vit-large-patch14) using the same label set.

Training procedure

• —Base model: google/vit-base-patch16-224
• —Fine-tuning method: Supervised training using the Hugging Face Trainer class
• —Data augmentation: Applied during training (e.g., RandomHorizontalFlip, ColorJitter)
• —Training time: ~5 epochs with and without augmentation
• —Optimizer: AdamW (default settings)
• —Evaluation metrics: Accuracy, precision, and recall
• —Best performance (no augmentation): 98.3% test accuracy

Training hyperparameters

The following hyperparameters were used during training:

• —learning_rate: 0.0003
• —trainbatchsize: 16
• —evalbatchsize: 8
• —seed: 42
• —optimizer: Use adamwtorch with betas=(0.9,0.999) and epsilon=1e-08 and optimizerargs=No additional optimizer arguments
• —lrschedulertype: linear
• —num_epochs: 5

Training results

Framework versions

• —Transformers 4.50.0
• —Pytorch 2.6.0+cu124
• —Datasets 3.4.1
• —Tokenizers 0.21.1