I already have a well-labeled image set of potato leaves in JPG and PNG format, covering Blight, Early blight, and Leaf spot. I now need a production-ready hybrid model that fuses EfficientNet’s feature extraction strengths with Vision Transformer (ViT) attention mechanisms to detect these diseases accurately. The work I’d like you to handle includes the full pipeline: clean and augment the data, design and train the EfficientNet + ViT architecture, fine-tune it for my three disease classes, then benchmark the results with clear metrics (accuracy, precision, recall, confusion matrix). Deliverables • A fully trained hybrid model (saved weights plus exportable ONNX or TensorFlow format) • A reproducible training notebook or script (Python; TensorFlow/Keras or PyTorch) • An inference script or minimal REST API that takes a new leaf image and returns the predicted class with confidence • Brief documentation so I can retrain or update the model later If you have prior experience blending CNN backbones with transformer heads, particularly on agricultural datasets, I’m keen to see examples.