Well-trained deep neural networks (DNNs) treat all test samples equally during prediction. Adaptive DNN inference with early exiting leverages the observation that some test examples can be easier to predict than others. This repository presents LEENet, a brand-new approach to early exit neural DNN models. Instead of traditional early exit networks where gate thesholds are either hardcoded or learned via inference budget thresholds, LEENet learns to optimize a cost/accuracy tradeoff balance based on a single alpha hyperparameter set by the user. This allows the model to choose an appropriate risk level for certain images which the model is less confident at earlier exit indices. Initial experiments are being conducted via DenseNet121, ResNet50, and VGG11 networks for computer vision tasks (ImageNetTE, CIFAR-10, CIFAR-100). Our initial results demonstrate that by solely tuning alpha, LEENet networks can vastly outperform traditional DNNs and also outperforms other early exit techniques in certain circumstances.

Any python version above 3.8 will suffice for this project. If this is a CUDA-enabled PC, install your latest version of CUDA software and the appropriate PyTorch software. All other packages are in the requirements.txt file linked in the repository.

All appropriate model architectures are either hardcoded in this repository, or are downloaded via public channels. Pretrained transfer learning models can be found here. Any model not in this list is already trained on the proper dataset and does not require an initial .pth file.

This is currently a multi-step process. Please begin by training the classifiers via main.ipynb, then train the appropriate alpha configurations as desired via alpha_tuning.ipynb.