A repository for explaining feature importances and feature interactions in deep neural networks using path attribution methods.
This repository contains tools to interpret and explain machine learning models using Integrated Gradients and Expected Gradients. In addition, it contains code to explain interactions in deep networks using Integrated Hessians and Expected Hessians - methods that we introduced in our most recent paper: "Explaining Explanations: Axiomatic Feature Interactions for Deep Networks". If you use our work to explain your networks, please cite this paper.
@article{janizek2020explaining,
author = {Joseph D. Janizek and Pascal Sturmfels and Su-In Lee},
title = {Explaining Explanations: Axiomatic Feature Interactions for Deep Networks},
journal = {Journal of Machine Learning Research},
year = {2021},
volume = {22},
number = {104},
pages = {1-54},
url = {http://jmlr.org/papers/v22/20-1223.html}
}
This repository contains two important directories: the path_explain directory, which contains the packages used to interpret and explain machine learning models, and the examples directory, which contains many examples using the path_explain module to explain different models on different data types.
The easiest way to install this package is by using pip:
pip install path-explain
Alternatively, you can clone this repository to re-run and explore the examples provided.
This package was written to support TensorFlow 2.0 (in eager execution mode) with Python 3. We have no current plans to support earlier versions of TensorFlow or Python.
Although we don't yet have formal API documentation, the underlying code does a pretty good job at explaining the API. See the code for generating attributions and interactions to better understand what the arguments to these functions mean.
For a simple, quick example to get started using this repository, see the example_usage.ipynb notebook in the top-level directory of this repository. It gives an overview of the functionality provided by this repository. For more advanced examples, keep reading on.
Our repository can easily be adapted to explain attributions and interactions learned on tabular data.
# other import statements...
from path_explain import PathExplainerTF, scatter_plot, summary_plot
### Code to train a model would go here
x_train, y_train, x_test, y_test = datset()
model = ...
model.fit(x_train, y_train, ...)
###
### Generating attributions using expected gradients
explainer = PathExplainerTF(model)
attributions = explainer.attributions(inputs=x_test,
baseline=x_train,
batch_size=100,
num_samples=200,
use_expectation=True,
output_indices=0)
###
### Generating interactions using expected hessians
interactions = explainer.interactions(inputs=x_test,
baseline=x_train,
batch_size=100,
num_samples=200,
use_expectation=True,
output_indices=0)
###Once we've generated attributions and interactions, we can use the provided plotting modules to help visualize them. First we plot a summary of the top features and their attribution values:
### First we need a list of strings denoting the name of each feature
feature_names = ...
###
summary_plot(attributions=attributions,
feature_values=x_test,
feature_names=feature_names,
plot_top_k=10)
Second, we plot an interaction our model has learned between maximum achieved heart rate and gender:
scatter_plot(attributions=attributions,
feature_values=x_test,
feature_index='max. achieved heart rate',
interactions=interactions,
color_by='is male',
feature_names=feature_names,
scale_y_ind=True)
The model used to generate the above interactions is a two layer neural network trained on the UCI Heart Disease Dataset. Interactions learned by this model were featured in our paper. To learn more about this particular model and the experimental setup, see the notebook used to train and explain the model.
As discussed in our paper, we can use Integrated Hessians to get interactions in language models. We explain a transformer from the HuggingFace Transformers Repository.
from transformers import DistilBertTokenizer, TFDistilBertForSequenceClassification, \
DistilBertConfig, glue_convert_examples_to_features, \
glue_processors
# This is a custom explainer to explain huggingface models
from path_explain import EmbeddingExplainerTF, text_plot, matrix_interaction_plot, bar_interaction_plot
tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
config = DistilBertConfig.from_pretrained('distilbert-base-uncased', num_labels=num_labels)
model = TFDistilBertForSequenceClassification.from_pretrained('distilbert-base-uncased', config=config)
### Some custom code to fine-tune the model on a sentiment analysis task...
max_length = 128
data, info = tensorflow_datasets.load('glue/sst-2', with_info=True)
train_dataset = glue_convert_examples_to_features(data['train'],
tokenizer,
max_length,
'sst-2)
valid_dataset = glue_convert_examples_to_features(data['validation'],
tokenizer,
max_length,
'sst-2')
...
### we won't include the whole fine-tuning code. See the HuggingFace repository for more.
### Here we define functions that represent two pieces of the model:
### embedding and prediction
def embedding_model(batch_ids):
batch_embedding = model.distilbert.embeddings(batch_ids)
return batch_embedding
def prediction_model(batch_embedding):
# Note: this isn't exactly the right way to use the attention mask.
# It should actually indicate which words are real words. This
# makes the coding easier however, and the output is fairly similar,
# so it suffices for this tutorial.
attention_mask = tf.ones(batch_embedding.shape[:2])
attention_mask = tf.cast(attention_mask, dtype=tf.float32)
head_mask = [None] * model.distilbert.num_hidden_layers
transformer_output = model.distilbert.transformer([batch_embedding, attention_mask, head_mask], training=False)[0]
pooled_output = transformer_output[:, 0]
pooled_output = model.pre_classifier(pooled_output)
logits = model.classifier(pooled_output)
return logits
###
### We need some data to explain
for batch in valid_dataset.take(1):
batch_input = batch[0]
batch_ids = batch_input['input_ids']
batch_embedding = embedding_model(batch_ids)
baseline_ids = np.zeros((1, 128), dtype=np.int64)
baseline_embedding = embedding_model(baseline_ids)
###
### We are finally ready to explain our model
explainer = EmbeddingExplainerTF(prediction_model)
attributions = explainer.attributions(inputs=batch_embedding,
baseline=baseline_embedding,
batch_size=32,
num_samples=256,
use_expectation=False,
output_indices=1)
###
### For interactions, the hessian is rather large so we use a very small batch size
interactions = explainer.interactions(inputs=batch_embedding,
baseline=baseline_embedding,
batch_size=1,
num_samples=256,
use_expectation=False,
output_indices=1)
###We can plot the learned attributions and interactions as follows. First we plot the attributions:
### First we need to decode the tokens from the batch ids.
batch_sentences = ...
### Doing so will depend on how you tokenized your model!
text_plot(batch_sentences[0],
attributions[0],
include_legend=True)
Then we plot the interactions:
bar_interaction_plot(interactions[0],
batch_sentences[0],
top_k=5)
If you would rather plot the full matrix of attributions rather than the top interactions in a bar plot, our package also supports this. First we show the attributions:
text_plot(batch_sentences[1],
attributions[1],
include_legend=True)
And then we show the full interaction matrix. Here we've zeroed out the diagonals so you can better see the off-diagonal terms.
matrix_interaction_plot(interaction_list[1],
token_list[1])
This example - interpreting DistilBERT - was also featured in our paper. You can examine the setup more here. For more examples, see the examples directory in this repository.