Reduce the Fidelity Inconsistency in Saliency Metric

This is the project for cse586

please use conda and environment.yml to setup the environment.

Pre-trained Weights

The weights trained on MNIST is here Please download SimpleCNN_E_MNIST.pt and put it under weights. The weights for other models will be download automatically.

Get the perturbed images and output

Note that all the experiments are commented out in run.py. Please select the corresponding experiment name to run certain experiments. The following is an example experiment.

  python run.py vgg16_E_CIFAR10_InputXGradient_no_iter_test_100

You can then use the following command with corresponding argument (check generate_results.py for the avaliable args) names to produce the numbers in the paper.

  python generate_results.py MNIST_LeRF