wbott/ml-classification-butterflies — reverse-engineered prompt

Build me a clean butterfly image classifier project that can identify about 75 species from photos. I want it to feel like a real, reusable machine learning pipeline, not just one notebook. It should be able to pull the butterfly dataset from Kaggle, preprocess the images, train a few model options, compare results, and make it obvious that the transfer learning approach is the best one to use here. A simple dense baseline and a custom CNN are fine too, but the main model should be the VGG16 based classifier.

Please include an easy way to run training, evaluate accuracy and loss, save the trained model, and predict the species for a new image. I would also like a notebook or simple report that shows training curves and a clear summary of model performance. Keep the code organized, configurable, and tested so it is easy to trust and rerun. If anything is missing or outdated, look up current docs online and make it work end to end.