pytorch-cnn-finetune

Why we included this project

Training image classifiers in PyTorch usually means rewriting the same transfer-learning glue for every model you try. This small library collapses that into a single make_model() call: it wraps dozens of pretrained backbones (ResNet, DenseNet, VGG, Inception, Xception, SENet, NASNet and others) and automatically swaps in a classifier head sized for your own number of classes. You can also feed images at resolutions other than the original ImageNet size, which matters for VGG and AlexNet with their fully-connected layers. Teams building custom vision classifiers on top of ImageNet weights will save real time here, and the exposed preprocessing metadata (mean, std, input size) makes it easy to match training and inference. One caveat: the project is stable but lightly maintained, so treat it as a convenience layer rather than a dependency you should expect rapid updates from.