-
Notifications
You must be signed in to change notification settings - Fork 7
Run DeepDIG against a new dataset or model
You can easily run DeepDIG against any other dataset or pre-trained model. You just need to make sure to follow the same format with our developed method.
-
For each dataset, you need to develop an Autoencoder and name it
ae.py
e.g., look at this for MNIST. -
utils.py for a dataset contains four functions.
get_original_data
returns data and labels of a train or test set based on the argument.get_class_specific_data
takes an index of a class (e.g., 1 for Trouser in FashionMNIST) as well as split (test and train) and return the relevant data.save_samples
function saves samples (images) in a given directory (it usesimshow
function inutils.py
) -
For a particular (pre-trained) model, create a directory with an arbitrary name (this name will be the name of the pre-trained when running DeepDIG) e.g., 'GoogleNet' for CIFAR 10. Inside this folder, there are two files:
model.py
andtrain.py
contains training and test procedures for the dataset for the model defined inmodel.py
. Make sure the model returns scores (before softmax) and the features (before the linear model).