Awesome

Shapley Explanation Networks

Implementation of the paper "Shapley Explanation Networks" at ICLR 2021. Note that this repo heavily uses the experimental feature of named tensors in PyTorch. As it was really confusing to implement the ideas for the authors, we find it tremendously easier to use this feature.

Dependencies

For running only ShapNets, one would mostly only need PyTorch, NumPy, and SciPy.

Usage

For a Shapley Module:

import torch
import torch.nn as nn
from ShapNet.utils import ModuleDimensions
from ShapNet import ShapleyModule

b_size = 3
features = 4
out = 1
dims = ModuleDimensions(
    features=features,
    in_channel=1,
    out_channel=out
)

sm = ShapleyModule(
    inner_function=nn.Linear(features, out),
    dimensions=dims
)
sm(torch.randn(b_size, features), explain=True)

For a Shallow ShapNet

import torch
import torch.nn as nn
from ShapNet.utils import ModuleDimensions
from ShapNet import ShapleyModule, OverlappingShallowShapleyNetwork

batch_size = 32
class_num = 10
dim = 32

overlapping_modules = [
    ShapleyModule(
        inner_function=nn.Sequential(nn.Linear(2, class_num)),
        dimensions=ModuleDimensions(
            features=2, in_channel=1, out_channel=class_num
        ),
    ) for _ in range(dim * (dim - 1) // 2)
]
shallow_shapnet = OverlappingShallowShapleyNetwork(
    list_modules=overlapping_modules
)
inputs = torch.randn(batch_size, dim, ), )
shallow_shapnet(torch.randn(batch_size, dim, ), )
output, bias = shallow_shapnet(inputs, explain=True, )

For a Deep ShapNet

import torch
import torch.nn as nn
from ShapNet.utils import ModuleDimensions
from ShapNet import ShapleyModule, ShallowShapleyNetwork, DeepShapleyNetwork

dim = 32
dim_input_channels = 1
class_num = 10
inputs = torch.randn(32, dim, ), )


dims = ModuleDimensions(
    features=dim,
    in_channel=dim_input_channels,
    out_channel=class_num
)
deep_shapnet = DeepShapleyNetwork(
    list_shapnets=[
        ShallowShapleyNetwork(
            module_dict=nn.ModuleDict({
                "(0, 2)": ShapleyModule(
                    inner_function=nn.Linear(2, class_num),
                    dimensions=ModuleDimensions(
                        features=2, in_channel=1, out_channel=class_num
                    )
                )},
            ),
            dimensions=ModuleDimensions(dim, 1, class_num)
        ),
    ],
)
deep_shapnet(inputs)
outputs = deep_shapnet(inputs, explain=True, )

For a vision model:

import numpy as np
import torch
import torch.nn as nn

# =============================================================================
# Imports {\sc ShapNet}
# =============================================================================
from ShapNet import DeepConvShapNet, ShallowConvShapleyNetwork, ShapleyModule
from ShapNet.utils import ModuleDimensions, NAME_HEIGHT, NAME_WIDTH, \
    process_list_sizes

num_channels = 3
num_classes = 10
height = 32
width = 32
list_channels = [3, 16, 10]
pruning = [0.2, 0.]
kernel_sizes = process_list_sizes([2, (1, 3), ])
dilations = process_list_sizes([1, 2])
paddings = process_list_sizes([0, 0])
strides = process_list_sizes([1, 1])

args = {
    "list_shapnets": [
        ShallowConvShapleyNetwork(
            shapley_module=ShapleyModule(
                inner_function=nn.Sequential(
                    nn.Linear(
                        np.prod(kernel_sizes[i]) * list_channels[i],
                        list_channels[i + 1]),
                    nn.LeakyReLU()
                ),
                dimensions=ModuleDimensions(
                    features=int(np.prod(kernel_sizes[i])),
                    in_channel=list_channels[i],
                    out_channel=list_channels[i + 1])
            ),
            reference_values=None,
            kernel_size=kernel_sizes[i],
            dilation=dilations[i],
            padding=paddings[i],
            stride=strides[i]
        ) for i in range(len(list_channels) - 1)
    ],
    "reference_values": None,
    "residual": False,
    "named_output": False,
    "pruning": pruning
}

dcs = DeepConvShapNet(**args)

Citation

If this is useful, you could cite our work as

@inproceedings{
wang2021shapley,
title={Shapley Explanation Networks},
author={Rui Wang and Xiaoqian Wang and David I. Inouye},
booktitle={International Conference on Learning Representations},
year={2021},
url={https://openreview.net/forum?id=vsU0efpivw}
}