DIANNA is a Python package that brings explainable AI (XAI) to your research project. It wraps carefully selected XAI methods in a simple, uniform interface. It's built by, with and for (academic) researchers and research software engineers working on machine learning projects.
DIANNA software is addressing needs of both (X)AI researchers and mostly the various domains scientists who are using or will use AI models for their research without being experts in (X)AI. DIANNA is future-proof: one of the very few XAI library supporting the Open Neural Network Exchange (ONNX) format.
After studying the vast XAI landscape we have made choices in the parts of the XAI Taxonomy on which methods, data modalities and problems types to focus. Our choices, based on the largest usage in scientific literature, are shown graphically in the XAI taxonomy below:
![XAI_taxonomy](https://private-user-images.githubusercontent.com/3244249/296184639-9b864980-86f4-4d0e-8a83-af7d6be606f7.png?jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.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.xBwLXeRtA7WXKE7k_YcFYz5Yh9c58WgsBdOUjrNuYUE)
The key points of DIANNA:
- Provides an easy-to-use interface for non (X)AI experts
- Implements well-known XAI methods (LIME, RISE and Kernal SHAP) chosen by systematic and objective evaluation criteria
- Supports the de-facto standard format for neural network models - ONNX.
- Includes clear instructions for export/conversions from Tensorflow, Pytorch, Keras and scikit-learn to ONNX.
- Supports images, text and time series data modalities. Tabular data and even embeddings support is planned.
- Comes with simple intuitive image and text benchmarks
- Easily extendable to other XAI methods
For more information on the unique strengths of DIANNA with comparison to other tools, please see the context landscape.
DIANNA can be installed from PyPI using pip on any of the supported Python versions (see badge):
python3 -m pip install dianna
To install the most recent development version directly from the GitHub repository run:
python3 -m pip install git+https://github.com/dianna-ai/dianna.git
If you get an error related to OpenMP when importing dianna, have a look at this issue for possible workarounds.
- To install TensorFlow you can follow this tutorial.
- To install TensorFlow Addons you can follow these steps. For further reading see this issue. Note that this temporary solution works only for macOS versions >= 12.0. Note that this step may have changed already, see dianna-ai#245.
- Before installing DIANNA, comment
tensorflow
requirement insetup.cfg
file (tensorflow package for M1 is calledtensorflow-macos
).
You need:
- your trained ONNX model (convert my pytorch/tensorflow/keras/scikit-learn model to ONNX)
- 1 data item to be explained
You get:
- a relevance map overlayed over the data item
In the library's documentation, the general usage is explained in How to use DIANNA
model_path = 'your_model.onnx' # model trained on text
text = 'The movie started great but the ending is boring and unoriginal.'
Which of your model's classes do you want an explanation for?
labels = [positive_class, negative_class]
Run using the XAI method of your choice, for example LIME:
explanation = dianna.explain_text(model_path, text, 'LIME')
dianna.visualization.highlight_text(explanation[labels.index(positive_class)], text)
model_path = 'your_model.onnx' # model trained on images
image = PIL.Image.open('your_image.jpeg')
Tell us what label refers to the channels, or colors, in the image.
axis_labels = {0: 'channels'}
Which of your model's classes do you want an explanation for?
labels = [class_a, class_b]
Run using the XAI method of your choice, for example RISE:
explanation = dianna.explain_image(model_path, image, 'RISE', axis_labels=axis_labels, labels=labels)
dianna.visualization.plot_image(explanation[labels.index(class_a)], original_data=image)
model_path = 'your_model.onnx' # model trained on images
timeseries_instance = pd.read_csv('your_data_instance.csv').astype(float)
num_features = len(timeseries_instance) # The number of features to include in the explanation.
num_samples = 500 # The number of samples to generate for the LIME explainer.
Which of your model's classes do you want an explanation for?
class_names= [class_a, class_b] # String representation of the different classes of interest
labels = np.argsort(class_names) # Numerical representation of the different classes of interest for the model
Run using the XAI method of your choice, for example LIME with the following additional arguments:
explanation = dianna.explain_timeseries(model_path, timeseries_data=timeseries_instance , method='LIME',
labels=labels, class_names=class_names, num_features=num_features,
num_samples=num_samples, distance_method='cosine')
For visualization of the heatmap please refer to the tutorial
model_path = 'your_model.onnx' # model trained on tabular data
tabular_instance = pd.read_csv('your_data_instance.csv')
Run using the XAI method of your choice. Note that you need to specify the mode, either regression or classification. This case, for instance a regression task using KernelSHAP with the following additional arguments:
explanation = dianna.explain_tabular(run_model, input_tabular=data_instance, method='kernelshap',
mode ='regression', training_data = X_train,
training_data_kmeans = 5, feature_names=input_features.columns)
plot_tabular(explanation, X_test.columns, num_features=10) # display 10 most salient features
The XAI methods (explainers) are sensitive to the choice of their hyperparameters! In this work, this sensitivity to hyperparameters is researched and useful conclusions are drawn. The default hyperparameters used in DIANNA for each explainer as well as the values for our tutorial examples are given in the Tutorials README.
Explore the explanations of your trained model using the DIANNA dashboard (for now images, text and time series classification is supported). Click here for more information.
![Dianna dashboard screenshot](https://raw.githubusercontent.com/dianna-ai/dianna/main/dianna/dashboard/dashboard-screenshot.png)
DIANNA comes with simple datasets. Their main goal is to provide intuitive insight into the working of the XAI methods. They can be used as benchmarks for evaluation and comparison of existing and new XAI methods.
Dataset | Description | Examples | Generation |
---|---|---|---|
Binary MNIST ![]() |
Greyscale images of the digits "1" and "0" - a 2-class subset from the famousMNIST dataset for handwritten digit classification. | ![]() |
Binary MNIST dataset generation |
Simple Geometric (circles and triangles) ![]() |
Images of circles and triangles for 2-class geometric shape classificaiton. The shapes of varying size and orientation and the background have varying uniform gray levels. | ![]() |
Simple geometric shapes dataset generation |
Simple Scientific (LeafSnap30) ![]() |
Color images of tree leaves - a 30-class post-processed subset from the LeafSnap dataset for automatic identification of North American tree species. | ![]() |
LeafSnap30 dataset generation |
Dataset | Description | Examples | Generation |
---|---|---|---|
Stanford sentiment treebank ![]() |
Dataset for predicting the sentiment, positive or negative, of movie reviews. | This movie was actually neither that funny, nor super witty. | Sentiment treebank |
Dataset | Description | Examples | Generation |
---|---|---|---|
Coffee dataset |
Food spectographs time series dataset for a two class problem to distinguish between Robusta and Arabica coffee beans. | data source | |
Weather dataset |
The light version of the weather prediciton dataset, which contains daily observations (89 features) for 11 European locations through the years 2000 to 2010. | data source |
Dataset | Description | Examples | Generation |
---|---|---|---|
Pengiun dataset |
Palmer Archipelago (Antarctica) penguin dataset is a great intro dataset for data exploration & visualization similar to the famous Iris dataset. | ![]() |
data source |
Weather dataset |
The light version of the weather prediciton dataset, which contains daily observations (89 features) for 11 European locations through the years 2000 to 2010. | data source |
We work with ONNX! ONNX is a great unified neural network standard which can be used to boost reproducible science. Using ONNX for your model also gives you a boost in performance! In case your models are still in another popular DNN (deep neural network) format, here are some simple recipes to convert them:
- pytorch and pytorch-lightning - use the built-in
torch.onnx.export
function to convert pytorch models to onnx, or call the built-into_onnx
function on yourLightningModule
to export pytorch-lightning models to onnx. - tensorflow - use the
tf2onnx
package to convert tensorflow models to onnx. - keras - same as the conversion from tensorflow to onnx, the
tf2onnx
package also supports keras. - scikit-learn - use the
skl2onnx
package to scikit-learn models to onnx.
More converters with examples and tutorials can be found on the ONNX tutorial page.
And here are links to notebooks showing how we created our models on the benchmark datasets:
Models | Generation |
---|---|
Binary MNIST model | Binary MNIST model generation |
Simple Geometric model | Simple geometric shapes model generation |
Simple Scientific model | LeafSnap30 model generation |
Models | Generation |
---|---|
Movie reviews model | Stanford sentiment treebank model generation |
Models | Generation |
---|---|
Penguin model (classification) | Penguin model generation |
Sunshine hours prediction model (regression) | Sunshine hours prediction model generation |
We envision the birth of the ONNX Scientific models zoo soon...
DIANNA supports different data modalities and XAI methods (explainers). We have evaluated many explainers using objective criteria (see the How to find your AI explainer blog-post). The table below contains links to the relevant XAI method's papers (for some explanatory videos on the methods, please see tutorials). The DIANNA tutorials cover each supported method and data modality on a least one dataset using the default or tuned hyperparameters. Our plans to expand DIANNA with more data modalities and explainers are given in the ROADMAP.
Data \ XAI | RISE | LIME | KernelSHAP |
---|---|---|---|
Images | ✅ | ✅ | ✅ |
Text | ✅ | ✅ | |
Timeseries | ✅ | ✅ | |
Tabular | planned | ✅ | ✅ |
Embedding | planned | planned | planned |
Graphs* | work in progress | work in progress | work in progress |
LRP and PatternAttribution also feature in the top 5 of our thoroughly evaluated explainers. Contributing by adding these and more (new) post-hoc explainability methods on ONNX models is very welcome!
For detailed information on using specific DIANNA functions, please visit the documentation page hosted at Readthedocs.
If you want to contribute to the development of DIANNA, have a look at the contribution guidelines. See our developer documentation for information on developer installation, running tests, generating documentation, versioning and making a release.
If you use this package for your scientific work, please consider citing directly the software as:
Ranguelova, E., Bos, P., Liu, Y., Meijer, C., Oostrum, L., Crocioni, G., Ootes, L., Chandramouli, P., Jansen, A., Smeets, S. (2023). dianna (*[VERSION YOU USED]*). Zenodo. https://zenodo.org/record/5592606
or the JOSS paper as:
Ranguelova et al., (2022). DIANNA: Deep Insight And Neural Network Analysis. Journal of Open Source Software, 7(80), 4493, https://doi.org/10.21105/joss.04493
See also the Zenodo page or the JOSS page for exporting the software citation to BibTteX and other formats.
This package was created with Cookiecutter and the NLeSC/python-template.