cnn_finetune

Why we included this project

If you've done any transfer learning with Keras, you know the pain of hunting down a pretrained ImageNet checkpoint and then guessing at the right preprocessing. This repo fixes that by packaging working fine-tuning examples for VGG, Inception, ResNet, and DenseNet as standalone scripts, each with its pretrained weights linked. Run one on Cifar10 and you'll see the whole loop end to end: load the model, swap the classifier head, retrain. It also covers the preprocessing details that trip everyone up, like converting RGB to BGR and subtracting the ImageNet mean pixel. For anyone learning transfer learning or adapting a reference implementation to their own data, it's a clear, concrete starting point.