DeepLIFT Lamarr

This is a reimplementation of the DeepLIFT algorithm in PyTorch which aims to attribute contributions to the difference of an output to a reference output to the input values. The original DeepLIFT paper can be found here and its improved version to approximate Shapley values can be found here.

Usage

The main functionality lies in the deeplift.DeepLiftClass class. This can be instatiated with a SequentialLoadedModel model which in turn is instanciated with a torchscript model. DeepLiftClass also takes a reference value or reference dataset (as a torch.Tensor). Attributions can be made with the attribute method. This method takes a torch.Tensor as input and returns the contributions of the input values to the difference of the output to the reference output. The method also returns the delta values, which are the diference between (output_of_the_net + contributions) and the reference output. The following example shows how to use the DeepLiftClass class:

import torch
import torch.nn as nn
import src.deeplift.deeplift as deeplift
import src.deeplift.parsing as parsing

my_model = nn.Sequential(
    nn.Linear(10, 20),
    nn.ReLU(),
    nn.Linear(20, 5)
)

explainer = deeplift.DeepLiftClass(model=parsing.SequentialLoadedModel(torch.jit.script(my_model)),
                                   reference_value=torch.zeros(1, 10))
explanations, _ = explainer.attribute(torch.ones(1, 10))
print(explanations)

A corresponding example can be found in deeplift/example.py.

Current restrictions

Models

Models for which contributions are calculated have to be given in the torchscript format. Models have to be of type nn.Sequential and activation functions must be given as their own layer instead of them being part of other layers. The output of models needs to be one dimensional. Explanations can also be calculated with respect to an error function. The only error function currently supported is MSE. Only certain layers are supported. These are:

• Linear
• Conv1d
• Conv2d
• Conv3d
• MaxPool1d
• MaxPool2d
• MaxPool3d
• AvgPool1d
• AvgPool2d
• AvgPool3d
• Flatten
• ReLU
• Sigmoid
• Dropout

Input

Input values always need to contain a batch dimension. The batch dimension is always the first dimension. The input values need to be of type torch.Tensor.