karl-schulz/attribution-experiments

Re-Implementing recent DNN Attribution methods

Attribution Experiments

Implementation and tests of several attribution methods:

• Saliency maps: https://arxiv.org/abs/1312.6034
• Integrated Gradients: https://arxiv.org/abs/1703.01365
• SmoothGrad: https://arxiv.org/abs/1706.03825
• Grad-CAM: https://arxiv.org/abs/1610.02391
• Occlusion: https://arxiv.org/abs/1311.2901
• Information Dropout: https://arxiv.org/abs/1611.01353 (repurpose as attribution method)

... and maybe some others soon.

Example

import matplotlib.pyplot as plt
from attribution.methods.grad_based import SmoothGrad
from attribution.utils.viz import Heatmap

# the model to be inspected
model = ...  # type: torch.nn.Module

# the image to analyse
img = ...  # type: torch.Tensor

# the target label to derive from, ground truth or any other
target = ...  # type: torch.Tensor

# get heatmap for this input image
method = SmoothGrad(model, std=0.2)
hmap = method.get_map(img=img, target=target)

# visualize the result
fig = plt.figure() # the figure to paint to, may be subfigure
visualizer = Heatmap()
visualizer.show(hmap=hmap, img=img, fig=fig)

Good Example

"sturgeon"

Bad Example

"basketball"

Information Dropout

On cluttered MNIST