Skip to main content

Library containing attribution and interpretation methods for deep nets.

Project description

Welcome to TruLens!

Library containing attribution and interpretation methods for deep nets.


These installation instructions assume that you have conda installed and added to your path.

  1. Create a virtual environment (or modify an existing one).
conda create -n "<my_name>" python=3.7  # Skip if using existing environment.
conda activate <my_name>
  1. Install dependencies.
conda install tensorflow-gpu=1  # Or whatever backend you're using.
conda install keras             # Or whatever backend you're using.
conda install matplotlib        # For visualizations.
  1. Install the trulens package
pip install trulens


Quick Usage

To quickly play around with the TruLens library, check out the following CoLab notebooks:

  • PyTorch: Open In Colab
  • Tensorflow 2 / Keras: Open In Colab


The trulens library supports several common machine learning libraries, including Keras, Pytorch, and TensorFlow.

In order to set the backend to the backend of your choice, use the TRULENS_BACKEND flag, e.g., to use the Keras backend, the following code could be used before TruLens imports:

import os
os.environ['TRULENS_BACKEND'] = 'keras'


Model Wrappers

In order to support a wide variety of backends with different interfaces for their respective models, TruLens uses its own ModelWrapper class which provides a general model interface to simplify the implementation of the API functions. A model wrapper class exists for backend's model that converts a model in the respective backend's format to the general TruLens ModelWrapper interface. The wrappers are found in the trulens.nn.models module, and any model defined using Keras, Pytorch, or Tensorflow should be wrapped with the appropriate wrapper before being used with the other API functions that require a model -- all other TruLens functionalities expect models to be an instance of trulens.nn.models.Model.

For example,

wrapped_model = KerasModel(model_defined_via_keras)

Attribution Methods

Attribution methods, in the most general sense, allow us to quantify the contribution of particular variables in a model towards a particular behavior of the model. In many cases, for example, this may simply measure the effect each input variable has on the output of the network.

Attribution methods extend the AttributionMethod class, and many concrete instances are found in the trulens.nn.attribution module.

Once an attribution method has been instantiated, its main function is its attributions method, which takes an np.Array of batched instances, where each instance matches the shape of the input to the model the attribution method was instantiated with.

See the method comparison demo for further information on the different types of attribution methods, their uses, and their relationships with one another.

Slices, Quantities, and Distributions

In order to obtain a high degree of flexibility in the types of attributions that can be produced, we implement Internal Influence, which is parameterized by a slice, quantity of interest, and distribution of interest, explained below.

The slice essentially defines a layer to use for internal attributions. The slice for the InternalInfluence method can be specified by an instance of the Slice class in the trulens.nn.slices module. A Slice object specifies two layers: (1) the layer of the variables that we are calculating attribution for (e.g., the input layer), and (2) the layer whose output defines our quantity of interest (e.g., the output layer, see below for more on quantities of interest).

The quantity of interest (QoI) essentially defines the model behavior we would like to explain using attributions. The QoI is a function of the model's output at some layer. For example, it may select the confidence score for a particular class. In its most general form, the QoI can be pecified by an implementation of the QoI class in the trulens.nn.quantities module. Several common default implementations are provided in this module as well.

The distribution of interest (DoI) essentially specifies for which points surrounding each instance the calculated attribution should be valid. The distribution can be specified via an implementation of the DoI class in the trulens.nn.distributions module, which is a function taking an input instance and producing a list of input points to aggregate attribution over. A few common default distributions implementing the DoI class can be found in the trulens.nn.distributions module.

See the parameterization demo for further explanations of the purpose of these parameters and examples of their usage.


In order to interpret the attributions produced by an AttributionMethod, a few useful visualizers are provided in the trulens.visualizations module. While the interface of each visualizer varies slightly, in general, the visualizers are a function taking an np.Array representing the attributions returned from an AttributionMethod and producing an image that can be used to interpret the attributions.

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Built Distribution

trulens-0.0.2-py3-none-any.whl (63.1 kB view hashes)

Uploaded py3

Supported by

AWS AWS Cloud computing Datadog Datadog Monitoring Facebook / Instagram Facebook / Instagram PSF Sponsor Fastly Fastly CDN Google Google Object Storage and Download Analytics Huawei Huawei PSF Sponsor Microsoft Microsoft PSF Sponsor NVIDIA NVIDIA PSF Sponsor Pingdom Pingdom Monitoring Salesforce Salesforce PSF Sponsor Sentry Sentry Error logging StatusPage StatusPage Status page