TruLens is a cross-framework library for deep learning explainability. It provides a uniform abstraction over a number of different frameworks. It provides a uniform abstraction layer over TensorFlow, PyTorch, and Keras and allows input and internal explanations.
This paper is an introduction to the theoretical foundations of the library. We’ve been using TruLens at TruEra across a wide range of real-world use cases to explain deep learning models ranging from time-series RNNs to image and NLP models, and wanted to share the awesomeness with the world.
To quickly play around with the TruLens library, check out the following Colab notebooks:
- PyTorch:
- TensorFlow 2 / Keras:
- NLP with PyTorch:
- NLP with TensorFlow 2 / Keras:
These installation instructions assume that you have conda installed and added to your path.
- 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>
- 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.
- Install the trulens pip package from PyPI.
pip install trulens
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.
To get the model wrapper, use the get_model_wrapper method in trulens.nn.models. A model wrapper class exists for each backend that converts a model in the respective backend's format to the general TruLens ModelWrapper interface. The wrappers are found in the models module, and any model defined using Keras, Pytorch, or Tensorflow should be wrapped 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.ModelWrapper.
For example,
from trulens_explain.nn.models import get_model_wrapper
wrapped_model = get_model_wrapper(model_defined_via_keras)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 returns an np.Array of batched items, where each item 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.
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 specified 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 record 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 record and producing a list of sample 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 Attributions for Different Use Cases 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.
To communicate with other trulens developers, join our Slack!
To cite this repository:
curl -LH "Accept: application/x-bibtex" https://doi.org/10.5281/zenodo.4495856