A unified framework of perturbation and gradient-based attribution methods for Deep Neural Networks interpretability
Clone or download
marcoancona Merge pull request #29 from harkous/master
update setup.py to enable installing via pip
Latest commit a4d6dcd Dec 15, 2018

README.md

DeepExplain: attribution methods for Deep Learning Build Status

DeepExplain provides a unified framework for state-of-the-art gradient and perturbation-based attribution methods. It can be used by researchers and practitioners for better undertanding the behavior of existing models, as well for benchmarking other attribution methods.

It supports Tensorflow as well as Keras with Tensorflow backend. Support for PyTorch is planned.

Implements the following methods:

Gradient-based attribution methods

Methods marked with (*) are implemented as modified chain-rule, as better explained in Towards better understanding of gradient-based attribution methods for Deep Neural Networks, Ancona et al, ICLR 2018. As such, the result might be slightly different from the original implementation.

Pertubration-based attribution methods

What are attributions?

Consider a network and a specific input to this network (eg. an image, if the network is trained for image classification). The input is multi-dimensional, made of several features. In the case of images, each pixel can be considered a feature. The goal of an attribution method is to determine a real value R(x_i) for each input feature, with respect to a target neuron of interest (for example, the activation of the neuron corresponsing to the correct class).

When the attributions of all input features are arranged together to have the same shape of the input sample we talk about attribution maps (as in the picture below), where red and blue colors indicate respectively features that contribute positively to the activation of the target output and features having a suppressing effect on it. Attribution methods comparison on InceptionV3

This can help to better understand the network behavior, which features mostly contribute to the output and possible reasons for missclassification.

DeepExplain Quickstart

Installation

pip install -e git+https://github.com/marcoancona/DeepExplain.git#egg=deepexplain

Notice that DeepExplain assumes you already have installed Tensorflow > 1.0 and (optionally) Keras > 2.0.

Usage

Working examples for Tensorflow and Keras can be found in the example folder of the repository. DeepExplain consists of a single method: explain(method_name, target_tensor, input_tensor, samples, ...args).

Parameter name Type Description
method_name string, required Name of the method to run (see Which method to use?).
target_tensor Tensor, required Tensorflow Tensor object representing the output of the model for which attributions are seeked (see Which tensor to target?).
input_tensor Tensor, required Symbolic input to the network.
samples numpy array, required Batch of input samples to be fed to input_tensor and for which attributions are seeked. Notice that the first dimension must always be the batch size.
...args various, optional Method-specific parameters (see below).

The method explain must be called within a DeepExplain context:

# Pseudo-code
from deepexplain.tensorflow import DeepExplain

# Option 1. Create and train your model within a DeepExplain context

with DeepExplain(session=...) as de:  # < enter DeepExplain context
    model = init_model()  # < construct the model
    model.fit()           # < train the model

    attributions = de.explain(...)  # < compute attributions

# Option 2. First create and train your model, then apply DeepExplain.
# IMPORTANT: in order to work correctly, the graph to analyze
# must always be (re)constructed within the context!

model = init_model()  # < construct the model
model.fit()           # < train the model

with DeepExplain(session=...) as de:  # < enter DeepExplain context
    new_model = init_model()  # < assumes init_model() returns a *new* model with the weights of `model`
    attributions = de.explain(...)  # < compute attributions

When initializing the context, make sure to pass the session parameter:

# With Tensorlow
import tensorflow as tf
# ...build model
sess = tf.Session()
# ... use session to train your model if necessary
with DeepExplain(session=sess) as de:
    ...

# With Keras
import keras
from keras import backend as K

model = Sequential()  # functional API is also supported
# ... build model and train

with DeepExplain(session=K.get_session()) as de:
    ...

See concrete examples here.

Which method to use?

DeepExplain supports several methods. The main partition is between gradient-based methods and perturbation-based methods. The former are faster, given that they estimate attributions with a few forward and backward iterations through the network. The latter perturb the input and measure the change in output with respect to the original input. This requires to sequentially test each feature (or group of features) and therefore takes more time, but tends to produce smoother results.

Some methods allow tunable parameters. See the table below.

Method method_name Optional parameters Notes
Saliency saliency [Gradient] Only positive attributions.
Gradient * Input grad*input [Gradient] Fast. May be affected by noisy gradients and saturation of the nonlinerities.
Integrated Gradients intgrad steps, baseline [Gradient] Similar to Gradient * Input, but performs steps iterations (default: 100) though the network, varying the input from baseline (default: zero) to the actual provided sample. When provided, baseline must be a numpy array with the size of the input (but no batch dimension since the same baseline will be used for all inputs in the batch).
epsilon-LRP elrp epsilon [Gradient]Computes Layer-wise Relevance Propagation. Only recommanded with ReLU or Tanh nonlinearities. Value for epsilon must be greater than zero (default: .0001).
DeepLIFT (Rescale) deeplift baseline [Gradient] In most cases a faster approximation of Integrated Gradients. Do not apply to networks with multiplicative units (ie. LSTM or GRU). When provided, baseline must be a numpy array with the size of the input, without the batch dimension (default: zero).
Occlusion occlusion window_shape, step [Perturbation] Computes rolling window view of the input array and replace each window with zero values, measuring the effect of the perturbation on the target output. The optional parameters window_shape and step behave like in skimage. By default, each feature is tested independently (window_shape=1 and step=1), however this might be extremely slow for large inputs (such as ImageNet images). When the input presents some local coherence (eg. images), you might prefer larger values for window_shape. In this case the attributions of the features in each window will be summed up. Notice that the result might vary significantly for different window sizes.

Which neuron to target?

In general, any tensor that represents the activation of any hidden or output neuron can be user as target_tensor. If your network performs a classification task (ie. one output neuron for each possible class) you might want to target the neuron corresponding to the correct class for a given sample, such that the attribution map might help you undertand the reasons for this neuron to (not) activate. However you can also target the activation of another class, for example a class that is often missclassified, to have insight about features that activate this class.

Important: Tensors in Tensorflow and Keras usually include the activations of all neurons of a layer. If you pass such a tensor to explain you will get the average attribution map for all neurons the Tensor refers to. If you want to target a specific neuron yuo need either to slice the component you are interested in or multiply it for a binary mask that only select the target neuron.

# Example on MNIST (classification, with 10 output classes)
# ... model is created and trained
logits = Tensor(shape=(1, 10)) # output layer, 2-dimensional Tensor, where first dimension is the batch size
ys = [[0, 1, 0, 0, 0, 0, 0, 0, 0, 0]]  # numpy array of shape (1, 10) with one-hot encoding of labels
target_tensor = logits * ys # < masked target tensor: only the second component of `logits` will be used to compute attributions

Softmax: if the network last activation is a Softmax, it is recommanded to target the activations before this normalization.

NLP / Embedding lookups

The most common cause of ValueError("None values not supported.") is run() being called with a tensor_input and target_tensor that are disconnected in the backpropagation. This is common when an embedding lookup layer is used, since the lookup operation does not propagate the gradient. To generate attributions for NLP models, the input of DeepExplain should be the result of the embedding lookup instead of the original model input. Then, attributions for each word are found by summing up along the appropriate dimension of the resulting attribution matrix.

Tensorflow pseudocode:

input_x = graph.get_operation_by_name("input_x").outputs[0]
# Get a reference to the embedding tensor
embedding = graph.get_operation_by_name("embedding").outputs[0]
pre_softmax = graph.get_operation_by_name("output/scores").outputs[0]

# Evaluate the embedding tensor on the model input (in other words, perform the lookup)
embedding_out = sess.run(embedding, {input_x: x_test})
# Run DeepExplain with the embedding as input
attributions = de.explain('elrp', pre_softmax * y_test_logits, embedding, embedding_out)

Multiple inputs

Models with multiple inputs are supported for gradient-based methods. Instead, the Occlusion method will raise an exception if called on a model with multiple inputs (how perturbation should be generated for multiple inputs is actually not well defined).

For a minimal (toy) example see the example folder.

Contributing

DeepExplain is still in active development. If you experience problems, feel free to open an issue. Contributions to extend the functinalities of this framework and/or to add support for other methods are welcome.

License

MIT