This repository contains the code for reproducing the results reported in the following paper:
Orhan AE (2018) A simple cache model for image recognition. NeurIPS 2018 [arxiv:1805.08709].
The code was written and tested in Tensorflow (v1.4.0) and Keras (v2.0.9) on a GPU cluster. Other configurations may or may not work. Please let me know if you have any trouble running the code. A brief description of the directories follows:
adversarial: contains code for running both black-box and white-box attacks against the baseline and cache models. The code here requires the Foolbox toolbox.depth-expts: contains code for testing the effect of layer depth in cache models, as reported in Fig. 2 in the paper.imagenet: contains code for running ImageNet experiments. I had to divide the training and validation data into several chunks to deal with memory issues (see the pre-processing files in the folder). This may or may not be suitable for your needs.saved_models: contains saved ResNet and DenseNet models trained on the CIFAR-10 and CIFAR-100 datasets.
Files with a _grid in their name can be used to run hyper-parameter searches. Files with a _evaluate in their name can be used to evaluate the models. Files with a _justmem in their name are associated with the cache-only (CacheOnly) models.