Skip to content
Interpretability Methods for tf.keras models with Tensorflow 2.0
Python Makefile
Branch: master
Clone or download
RaphaelMeudec Integrated gradients (#70)
* Extract normalized grayscale from SmoothGrad

* Add Integrated Gradients method

* Add Integrated Gradients example

* Add Integrated Gradients callback

* Add Integrated Gradients to README

* Add Integrated Gradients in docs
Latest commit 58554e2 Aug 23, 2019

README.md

tf-explain

Pypi Version Build Status Documentation Status Python Versions Tensorflow Versions Code style: black

tf-explain implements interpretability methods as Tensorflow 2.0 callbacks to ease neural network's understanding.
See Introducing tf-explain, Interpretability for Tensorflow 2.0

Installation

tf-explain is available on PyPi as an alpha release. To install it:

virtualenv venv -p python3.6
pip install tf-explain

tf-explain is compatible with Tensorflow 2. It is not declared as a dependency to let you choose between CPU and GPU versions. Additionally to the previous install, run:

# For CPU version
pip install tensorflow==2.0.0-beta1
# For GPU version
pip install tensorflow-gpu==2.0.0-beta1

Available Methods

  1. Activations Visualization
  2. Occlusion Sensitivity
  3. Grad CAM (Class Activation Maps)
  4. SmoothGrad
  5. Integrated Gradients

Activations Visualization

Visualize how a given input comes out of a specific activation layer

from tf_explain.callbacks.activations_visualization import ActivationsVisualizationCallback

model = [...]

callbacks = [
    ActivationsVisualizationCallback(
        validation_data=(x_val, y_val),
        layers_name=["activation_1"],
        output_dir=output_dir,
    ),
]

model.fit(x_train, y_train, batch_size=2, epochs=2, callbacks=callbacks)

Occlusion Sensitivity

Visualize how parts of the image affects neural network's confidence by occluding parts iteratively

from tf_explain.callbacks.occlusion_sensitivity import OcclusionSensitivityCallback

model = [...]

callbacks = [
    OcclusionSensitivityCallback(
        validation_data=(x_val, y_val),
        class_index=0,
        patch_size=4,
        output_dir=output_dir,
    ),
]

model.fit(x_train, y_train, batch_size=2, epochs=2, callbacks=callbacks)

Occlusion Sensitivity for Tabby class (stripes differentiate tabby cat from other ImageNet cat classes)

Grad CAM

Visualize how parts of the image affects neural network's output by looking into the activation maps

From Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization

from tf_explain.callbacks.grad_cam import GradCAMCallback

model = [...]

callbacks = [
    GradCAMCallback(
        validation_data=(x_val, y_val),
        layer_name="activation_1",
        class_index=0,
        output_dir=output_dir,
    )
]

model.fit(x_train, y_train, batch_size=2, epochs=2, callbacks=callbacks)

SmoothGrad

Visualize stabilized gradients on the inputs towards the decision

From SmoothGrad: removing noise by adding noise

from tf_explain.callbacks.smoothgrad import SmoothGradCallback

model = [...]

callbacks = [
    SmoothGradCallback(
        validation_data=(x_val, y_val),
        class_index=0,
        num_samples=20,
        noise=1.,
        output_dir=output_dir,
    )
]

model.fit(x_train, y_train, batch_size=2, epochs=2, callbacks=callbacks)

Integrated Gradients

Visualize an average of the gradients along the construction of the input towards the decision

From Axiomatic Attribution for Deep Networks

from tf_explain.callbacks.integrated_gradients import IntegratedGradientsCallback

model = [...]

callbacks = [
    IntegratedGradientsCallback(
        validation_data=(x_val, y_val),
        class_index=0,
        n_steps=20,
        output_dir=output_dir,
    )
]

model.fit(x_train, y_train, batch_size=2, epochs=2, callbacks=callbacks)

Visualizing the results

When you use the callbacks, the output files are created in the logs directory.

You can see them in Tensorboard with the following command: tensorboard --logdir logs

Roadmap

You can’t perform that action at this time.