Build Custom CNN From Dataset

I already have a complete image-based dataset and now I want a robust Convolutional Neural Network built around it. I need your help choosing or designing the right architecture, training it from scratch or with transfer learning, and then packaging the finished model so I can reuse it easily for inference. You can work in TensorFlow/Keras, PyTorch or any modern deep-learning stack you prefer—as long as the code is clean, well-commented and fully reproducible. At the end I expect: • Source code for preprocessing, training and evaluation • Trained model weights plus a simple predict/inference script • A concise report of the hyper-parameters you settled on and the metrics you achieved on my hold-out data If something in my dataset format needs adjustment, let me know and we can refine it together. I’m happy to provide quick feedback throughout, and I’d like the whole pipeline ready to run on a standard GPU environment.