👋 Xplique is a Neural Networks Explainability Toolbox
🦊 Xplique (pronounced \ɛks.plik\) is a Python toolkit dedicated to explainability. The goal of this library is to gather the state of the art of Explainable AI to help you understand your complex neural network models. Originally built for Tensorflow's model it also works for PyTorch models partially.
📘 Explore Xplique docs
|
Explore Xplique tutorials 🔥
Attributions
·
Concept
·
Feature Visualization
·
Metrics
The library is composed of several modules, the Attributions Methods module implements various methods (e.g Saliency, Grad-CAM, Integrated-Gradients...), with explanations, examples and links to official papers. The Feature Visualization module allows to see how neural networks build their understanding of images by finding inputs that maximize neurons, channels, layers or compositions of these elements. The Concepts module allows you to extract human concepts from a model and to test their usefulness with respect to a class. Finally, the Metrics module covers the current metrics used in explainability. Used in conjunction with the Attribution Methods module, it allows you to test the different methods or evaluate the explanations of a model.
You can find a certain number of other practical tutorials just here. This section is actively developed and more contents will be included. We will try to cover all the possible usage of the library, feel free to contact us if you have any suggestions or recommendations towards tutorials you would like to see.
Xplique requires a version of python higher than 3.7 and several libraries including Tensorflow and Numpy. Installation can be done using Pypi:
pip install xplique
Now that Xplique is installed, here are basic examples of what you can do with the available modules.
from xplique.attributions import GradCAM
# load images, labels and model
# ...
explainer = GradCAM(model)
explanations = explainer.explain(images, labels)
# or just `explainer(images, labels)`
All attributions methods share a common API described in the attributions API documentation.
In order to measure if the explanations provided by our method are faithful (it reflects well the functioning of the model) we can use a fidelity metric such as Deletion
from xplique.attributions import GradCAM
from xplique.metrics import Deletion
# load images, labels and model
# ...
explainer = GradCAM(model)
explanations = explainer(inputs, labels)
metric = Deletion(model, inputs, labels)
score_grad_cam = metric(explanations)
All attributions metrics share a common API. You can find out more about it here.
Concerning the concept-based methods, we can for example extract a concept vector from a layer of a model. In order to do this, we use two datasets, one containing inputs containing the concept: positive_samples
, the other containing other entries which do not contain the concept: negative_samples
.
from xplique.concepts import Cav
# load a model, samples that contain a concept
# (positive) and samples who don't (negative)
# ...
extractor = Cav(model, 'mixed3')
concept_vector = extractor(positive_samples,
negative_samples)
More information on CAV here and on TCAV here.
Use Craft to investigate a single class.
from xplique.concepts import CraftTf as Craft
# Cut the model in two parts: g and h
# Create a Craft concept extractor from these 2 models
craft = Craft(input_to_latent_model = g,
latent_to_logit_model = h)
# Use Craft to compute the concepts for a specific class
craft.fit(images_preprocessed, class_id=rabbit_class_id)
# Compute Sobol indices to understand which concept matters
importances = craft.estimate_importance()
# Display those concepts by showing the 10 best crops for each concept
craft.plot_concepts_crops(nb_crops=10)
More information in the CRAFT documentation.
Finally, in order to find an image that maximizes a neuron and at the same time a layer, we build two objectives that we combine together. We then call the optimizer which returns our images
from xplique.features_visualizations import Objective
from xplique.features_visualizations import optimize
# load a model...
neuron_obj = Objective.neuron(model, "logits", 200)
channel_obj = Objective.layer(model, "mixed3", 10)
obj = neuron_obj + 2.0 * channel_obj
images, obj_names = optimize(obj)
Want to know more ? Check the Feature Viz documentation
Even though the library was mainly designed to be a Tensorflow toolbox we have been working on a very practical wrapper to facilitate the integration of your PyTorch models into Xplique's framework!
import torch
from xplique.wrappers import TorchWrapper
from xplique.attributions import Saliency
from xplique.metrics import Deletion
# load images, targets and model
# ...
device = 'cuda' if torch.cuda.is_available() else 'cpu'
wrapped_model = TorchWrapper(torch_model, device)
explainer = Saliency(wrapped_model)
explanations = explainer(inputs, targets)
metric = Deletion(wrapped_model, inputs, targets)
score_saliency = metric(explanations)
Want to know more ? Check the PyTorch documentation
There are 4 modules in Xplique, Attribution methods, Attribution metrics, Concepts, and Feature visualization. In particular, the attribution methods module supports a huge diversity of tasks:Classification, Regression, Object Detection, and Semantic Segmentation. For diverse data types: Images, Time Series, and Tabular data. The methods compatible with such task are highlighted in the following table:
Attribution Method | Type of Model | Source | Images | Time Series and Tabular Data | Tutorial |
---|---|---|---|---|---|
Deconvolution | TF | Paper | C✔️ OD❌ SS❌ | C✔️ R✔️ | |
Grad-CAM | TF | Paper | C✔️ OD❌ SS❌ | ❌ | |
Grad-CAM++ | TF | Paper | C✔️ OD❌ SS❌ | ❌ | |
Gradient Input | TF, PyTorch** | Paper | C✔️ OD✔️ SS✔️ | C✔️ R✔️ | |
Guided Backprop | TF | Paper | C✔️ OD❌ SS❌ | C✔️ R✔️ | |
Integrated Gradients | TF, PyTorch** | Paper | C✔️ OD✔️ SS✔️ | C✔️ R✔️ | |
Kernel SHAP | TF, PyTorch**, Callable* | Paper | C✔️ OD✔️ SS✔️ | C✔️ R✔️ | |
Lime | TF, PyTorch**, Callable* | Paper | C✔️ OD✔️ SS✔️ | C✔️ R✔️ | |
Occlusion | TF, PyTorch**, Callable* | Paper | C✔️ OD✔️ SS✔️ | C✔️ R✔️ | |
Rise | TF, PyTorch**, Callable* | Paper | C✔️ OD✔️ SS✔️ | C✔️ R✔️ | |
Saliency | TF, PyTorch** | Paper | C✔️ OD✔️ SS✔️ | C✔️ R✔️ | |
SmoothGrad | TF, PyTorch** | Paper | C✔️ OD✔️ SS✔️ | C✔️ R✔️ | |
SquareGrad | TF, PyTorch** | Paper | C✔️ OD✔️ SS✔️ | C✔️ R✔️ | |
VarGrad | TF, PyTorch** | Paper | C✔️ OD✔️ SS✔️ | C✔️ R✔️ | |
Sobol Attribution | TF, PyTorch** | Paper | C✔️ OD✔️ SS✔️ | 🔵 | |
Hsic Attribution | TF, PyTorch** | Paper | C✔️ OD✔️ SS✔️ | 🔵 | |
FORGrad enhancement | TF, PyTorch** | Paper | C✔️ OD✔️ SS✔️ | ❌ |
TF : Tensorflow compatible
C : Classification | R : Regression | OD : Object Detection | SS : Semantic Segmentation (SS)
* : See the Callable documentation
** : See the Xplique for PyTorch documentation, and the PyTorch models: Getting started notebook.
✔️ : Supported by Xplique | ❌ : Not applicable | 🔵 : Work in Progress
Attribution Metrics | Type of Model | Property | Source |
---|---|---|---|
MuFidelity | TF, PyTorch** | Fidelity | Paper |
Deletion | TF, PyTorch** | Fidelity | Paper |
Insertion | TF, PyTorch** | Fidelity | Paper |
Average Stability | TF, PyTorch** | Stability | Paper |
MeGe | TF, PyTorch** | Representativity | Paper |
ReCo | TF, PyTorch** | Consistency | Paper |
(WIP) e-robustness |
TF : Tensorflow compatible
** : See the Xplique for PyTorch documentation, and the PyTorch models: Getting started notebook.
Concepts method | Type of Model | Source | Tutorial |
---|---|---|---|
Concept Activation Vector (CAV) | TF | Paper | |
Testing CAV (TCAV) | TF | Paper | |
CRAFT Tensorflow | TF | Paper | |
CRAFT PyTorch | PyTorch** | Paper | |
(WIP) Robust TCAV | |||
(WIP) Automatic Concept Extraction (ACE) |
TF : Tensorflow compatible
** : See the Xplique for Pytorch documentation, and the PyTorch's model: Getting started notebook
Feature Visualization (Paper) | Type of Model | Details |
---|---|---|
Neurons | TF | Optimizes for specific neurons |
Layer | TF | Optimizes for specific layers |
Channel | TF | Optimizes for specific channels |
Direction | TF | Optimizes for specific vector |
Fourrier Preconditioning | TF | Optimize in Fourier basis (see preconditioning) |
Objective combination | TF | Allows to combine objectives |
MaCo | TF | Fixed Magnitude optimisation, see Paper |
TF : Tensorflow compatible
Feel free to propose your ideas or come and contribute with us on the Xplique toolbox! We have a specific document where we describe in a simple way how to make your first pull request: just here.
This library is one approach of many to explain your model. We don't expect it to be the perfect solution, we create it to explore one point in the space of possibilities.
This library was started as a side-project by Thomas FEL who is currently a graduate student at the Artificial and Natural Intelligence Toulouse Institute under the direction of Thomas SERRE. His thesis work focuses on explainability for deep neural networks.
He then received help from some members of the DEEL team to enhance the library namely from Lucas Hervier and Antonin Poché.
If you use Xplique as part of your workflow in a scientific publication, please consider citing the 🗞️ Xplique official paper:
@article{fel2022xplique,
title={Xplique: A Deep Learning Explainability Toolbox},
author={Fel, Thomas and Hervier, Lucas and Vigouroux, David and Poche, Antonin and Plakoo, Justin and Cadene, Remi and Chalvidal, Mathieu and Colin, Julien and Boissin, Thibaut and Bethune, Louis and Picard, Agustin and Nicodeme, Claire
and Gardes, Laurent and Flandin, Gregory and Serre, Thomas},
journal={Workshop on Explainable Artificial Intelligence for Computer Vision (CVPR)},
year={2022}
}
The package is released under MIT license.