Keras

Keras is “a high-level neural networks API, written in Python and capable of running on top of TensorFlow, CNTK, or Theano”.

Keras can be used for many Machine Learning tasks, and it has support for both popular and experimental neural network architectures.

explain_prediction

Currently ELI5 supports eli5.explain_prediction() for Keras image classifiers. eli5.explain_prediction() explains image classifications through Grad-CAM.

The returned eli5.base.Explanation instance contains some important objects:

  • image represents the image input into the model. A Pillow image with mode “RGBA”.

  • targets represents the explanation values for each target class (currently only 1 target is supported). A list of eli5.base.TargetExplanation objects with the following attributes set:

    • heatmap a grayscale “localization map” (rank 2 (2D) numpy array, with float values in the interval [0, 1]). The numbers indicate how important the region in the image is for the target class (even if the target class was not the predicted class). Higher numbers mean that the region tends to increase the predicted value for a class. Lower numbers mean that the region has smaller effect on the predicted class score.
    • target the integer ID of the class (same as the argument to targets if one was passed, or the predicted class ID if no argument was passed).
    • score the output of the network for the predicted class.

Important arguments to eli5.explain_prediction() for Model and Sequential:

  • doc is an image as a tensor that can be inputted to the model.

    • The tensor must be an instance of numpy.ndarray.
    • Usually the tensor has the format (batch, dims, …, channels) (channels last format, dims=(height, width), batch=1, one image), i.e. BHWC.
    • Check model.input_shape to confirm the required dimensions of the input tensor.
  • target_names are the names of the output classes.

    • Currently not implemented.
  • targets are the output classes to focus on. Possible values include:

    • A list of integers (class ID’s). Only the first prediction from the list is currently taken. The list must be length one.
    • None for automatically taking the top prediction of the model.
  • layer is the layer in the model from which the heatmap will be generated. Possible values are:

    • An instance of Layer, a name (str), or an index (int)
    • None for automatically getting a suitable layer, if possible.

All other arguments are ignored.

Note

Top-level eli5.explain_prediction() calls are dispatched to eli5.keras.explain_prediction_keras() for keras.models.Model and keras.models.Sequential.

show_prediction

ELI5 supports eli5.show_prediction() to conveniently invoke explain_prediction with format_as_image, and display the explanation in an IPython cell.

Grad-CAM

ELI5 contains eli5.keras.gradcam.gradcam() and eli5.keras.gradcam.gradcam_backend().

These functions can be used to obtain finer details of a Grad-CAM explanation.