ViT Dog Breed Classifier

Fine-tune google/vit-base-patch16-224 for dog breed classification using HuggingFace Transformers.

Setup

make install
make gpu-check

Prepare Data

Option A — CSV + image folder (recommended)

Place labels.csv and images in data/dogs/:

data/dogs/
├── labels.csv     # columns: id,breed (id = filename without extension)
├── train/
│   ├── 000bec180eb18c7604dcecc8fe0dba07.jpg
│   ├── 001513dfcb2ffafc82cccf4dbbaba97.jpg
│   └── ...
└── test/          # optional, for inference

Option B — Breed subdirectories

data/dogs/
├── labrador/
│   ├── img001.jpg
│   └── ...
├── golden_retriever/
│   └── ...
└── poodle/
    └── ...

Validate the dataset:

make data-check

Train

make train

Training config is in config.py. The best model is saved to outputs/best/.

To resume an interrupted run:

make train-resume

Evaluate

make eval

Inference

make infer IMG=path/to/dog.jpg
make infer-dir DIR=path/to/images/

Project Structure

File	Description
config.py	Model, training, and inference configuration
dataset.py	Image loading, augmentation, and preprocessing
train.py	Fine-tuning with HuggingFace Trainer
inference.py	Single-image and batch inference
Makefile	Development lifecycle commands