Saliency Methods
If upgrading from an older version, update old imports to import saliency.tf1 as saliency. We provide wrappers to make the framework-agnostic version compatible with TF1 models. (Example TF1 notebook)
Introduction
This repository contains code for the following saliency techniques:
- Guided Integrated Gradients* (paper, poster)
- XRAI* (paper, poster)
- SmoothGrad* (paper)
- Vanilla Gradients (paper, paper)
- Guided Backpropogation (paper)
- Integrated Gradients (paper)
- Occlusion
- Grad-CAM (paper)
- Blur IG (paper)
*Developed by PAIR.
This list is by no means comprehensive. We are accepting pull requests to add new methods!
Download
# To install the core subpackage:
pip install saliency
# To install core and tf1 subpackages:
pip install saliency[tf1]
or for the development version:
git clone https://github.com/pair-code/saliency
cd saliency
Usage
The saliency library has two subpackages:
coreuses a genericcall_model_functionwhich can be used with any ML framework.tf1accepts input/output tensors directly, and sets up the necessary graph operations for each method.
Core
Each saliency mask class extends from the CoreSaliency base class. This class contains the following methods:
GetMask(x_value, call_model_function, call_model_args=None): Returns a mask of the shape of non-batchedx_valuegiven by the saliency technique.GetSmoothedMask(x_value, call_model_function, call_model_args=None, stdev_spread=.15, nsamples=25, magnitude=True): Returns a mask smoothed of the shape of non-batchedx_valuewith the SmoothGrad technique.
The visualization module contains two methods for saliency visualization:
VisualizeImageGrayscale(image_3d, percentile): Marginalizes across the absolute value of each channel to create a 2D single channel image, and clips the image at the given percentile of the distribution. This method returns a 2D tensor normalized between 0 to 1.VisualizeImageDiverging(image_3d, percentile): Marginalizes across the value of each channel to create a 2D single channel image, and clips the image at the given percentile of the distribution. This method returns a 2D tensor normalized between -1 to 1 where zero remains unchanged.
If the sign of the value given by the saliency mask is not important, then use VisualizeImageGrayscale, otherwise use VisualizeImageDiverging. See the SmoothGrad paper for more details on which visualization method to use.
call_model_function
call_model_function is how we pass inputs to a given model and receive the outputs necessary to compute saliency masks. The description of this method and expected output format is in the CoreSaliency description, as well as separately for each method.
Examples
This example iPython notebook showing these techniques is a good starting place.
Here is a condensed example of using IG+SmoothGrad with TensorFlow 2:
import saliency.core as saliency
import tensorflow as tf
...
# call_model_function construction here.
def call_model_function(x_value_batched, call_model_args, expected_keys):
tape = tf.GradientTape()
grads = np.array(tape.gradient(output_layer, images))
return {saliency.INPUT_OUTPUT_GRADIENTS: grads}
...
# Load data.
image = GetImagePNG(...)
# Compute IG+SmoothGrad.
ig_saliency = saliency.IntegratedGradients()
smoothgrad_ig = ig_saliency.GetSmoothedMask(image,
call_model_function,
call_model_args=None)
# Compute a 2D tensor for visualization.
grayscale_visualization = saliency.VisualizeImageGrayscale(
smoothgrad_ig)
TF1
Each saliency mask class extends from the TF1Saliency base class. This class contains the following methods:
__init__(graph, session, y, x): Constructor of the SaliencyMask. This can modify the graph, or sometimes create a new graph. Often this will add nodes to the graph, so this shouldn't be called continuously.yis the output tensor to compute saliency masks with respect to,xis the input tensor with the outer most dimension being batch size.GetMask(x_value, feed_dict): Returns a mask of the shape of non-batchedx_valuegiven by the saliency technique.GetSmoothedMask(x_value, feed_dict): Returns a mask smoothed of the shape of non-batchedx_valuewith the SmoothGrad technique.
The visualization module contains two visualization methods:
VisualizeImageGrayscale(image_3d, percentile): Marginalizes across the absolute value of each channel to create a 2D single channel image, and clips the image at the given percentile of the distribution. This method returns a 2D tensor normalized between 0 to 1.VisualizeImageDiverging(image_3d, percentile): Marginalizes across the value of each channel to create a 2D single channel image, and clips the image at the given percentile of the distribution. This method returns a 2D tensor normalized between -1 to 1 where zero remains unchanged.
If the sign of the value given by the saliency mask is not important, then use VisualizeImageGrayscale, otherwise use VisualizeImageDiverging. See the SmoothGrad paper for more details on which visualization method to use.
Examples
This example iPython notebook shows these techniques is a good starting place.
Another example of using GuidedBackprop with SmoothGrad from TensorFlow:
from saliency.tf1 import GuidedBackprop
from saliency.tf1 import VisualizeImageGrayscale
import tensorflow.compat.v1 as tf
...
# Tensorflow graph construction here.
y = logits[5]
x = tf.placeholder(...)
...
# Compute guided backprop.
# NOTE: This creates another graph that gets cached, try to avoid creating many
# of these.
guided_backprop_saliency = GuidedBackprop(graph, session, y, x)
...
# Load data.
image = GetImagePNG(...)
...
smoothgrad_guided_backprop =
guided_backprop_saliency.GetMask(image, feed_dict={...})
# Compute a 2D tensor for visualization.
grayscale_visualization = visualization.VisualizeImageGrayscale(
smoothgrad_guided_backprop)
Conclusion/Disclaimer
If you have any questions or suggestions for improvements to this library, please contact the owners of the PAIR-code/saliency repository.
This is not an official Google product.
213 Dec 27, 2022
789 Dec 29, 2022
93 Nov 6, 2022
13 Nov 21, 2022
5 Dec 14, 2022
49 Nov 5, 2022
75 Nov 9, 2022
5 Sep 16, 2022
1 Nov 12, 2021
865 Dec 24, 2022
463 Dec 28, 2022
2 Mar 16, 2022
113 Dec 23, 2022
224 Dec 28, 2022
25.1k Jan 2, 2023
2.8k Jan 8, 2023
19 Sep 29, 2022