Notebooks I wrote to get familiar with pytorch, and where I try to implement LRP-gamma.

• Python files in the order of dependency
  • trainable_net.py
    • TrainableNet: A class that can be subclassed instead of torch.nn.Module, which implements training and evaluating the accuracy
    • Some convenience functions like one_hot
    • input_times_gradient(net, images, target_pattern)
  • cifar10_utils.py <- trainable_net.py
    • code of LeNet taken from the pytorch cifar10 tutorial.
    • Some convenience methods like one_hot
    • A function to plot heatmaps
  • layerized_net.py <- trainable_net.py
    • A Layer and a LayerizedNet class that convert existing models to layerized versions. The layer types have to be implemented for the specific use case.
  • lrp.py <- layerized_net.py Layers and a network class that support deep taylor decomposition and LRP gamma.
• Notebooks
  • save_load_example.ipynb is a minimal demo of how to use the net. It trains, saves and loads the net.
  • lrp-gamma.ipynb <- layerized_net.py, cifar10_utils.py Implementation and a few tests with deep taylor decomposition and lrp-gamma on cifar10. Here is where lrp.py comes from
  • mnist_deeptaylor.ipynb <- lrp.py MNIST classifier and DTD usage, example how to use lrp.py
  • cifar10_adversarial_examples.ipynb generates adversarial examples. They have been introduced here
  • gradient_times_input.ipynb <- trainable_net.py, cifar10_utils.py Example with cifar10