PyTorch Explain: Logic Explained Networks in Python.
Project description
PyTorch, Explain! is an extension library for PyTorch to develop explainable deep learning models called Logic Explained Networks (LENs).
It consists of various methods for explainability from a variety of published papers, including the APIs required to get first-order logic explanations from deep neural networks.
Quick start
You can install torch_explain along with all its dependencies from PyPI:
pip install torch-explain
Example
For this simple experiment, let’s solve the XOR problem (augmented with 100 dummy features):
import torch
import torch_explain as te
from torch.nn.functional import one_hot
x0 = torch.zeros((4, 100))
x_train = torch.tensor([
[0, 0],
[0, 1],
[1, 0],
[1, 1],
], dtype=torch.float)
x_train = torch.cat([x_train, x0], dim=1)
y_train = torch.tensor([0, 1, 1, 0], dtype=torch.long)
y_train_1h = one_hot(y_train).to(torch.float)
We can instantiate a simple feed-forward neural network with 3 layers using the EntropyLayer as the first one:
layers = [
te.nn.EntropyLinear(x_train.shape[1], 10, n_classes=y_train_1h.shape[1]),
torch.nn.LeakyReLU(),
torch.nn.Linear(10, 4),
torch.nn.LeakyReLU(),
torch.nn.Linear(4, 1),
]
model = torch.nn.Sequential(*layers)
We can now train the network by optimizing the cross entropy loss and the entropy_logic_loss loss function incorporating the human prior towards simple explanations:
optimizer = torch.optim.AdamW(model.parameters(), lr=0.001)
loss_form = torch.nn.BCEWithLogitsLoss()
model.train()
for epoch in range(2001):
optimizer.zero_grad()
y_pred = model(x_train).squeeze(-1)
loss = loss_form(y_pred, y_train_1h) + 0.0001 * te.nn.functional.entropy_logic_loss(model)
loss.backward()
optimizer.step()
Once trained we can extract first-order logic formulas describing how the network composed the input features to obtain the predictions:
from torch_explain.logic.nn import entropy
from torch.nn.functional import one_hot
y1h = one_hot(y_train)
explanation, _ = entropy.explain_class(model, x_train, y1h, x_train, y1h, target_class=1)
Explanations will be logic formulas in disjunctive normal form. In this case, the explanation will be y=1 IFF (f1 AND ~f2) OR (f2 AND ~f1) corresponding to y=1 IFF f1 XOR f2.
The quality of the logic explanation can quantitatively assessed in terms of classification accuracy and rule complexity as follows:
from torch_explain.logic.metrics import test_explanation, complexity
accuracy, preds = test_explanation(explanation, x_train, y1h, target_class=1)
explanation_complexity = complexity(explanation)
In this case the accuracy is 100% and the complexity is 4.
Experiments
Training
To train the model(s) in the paper, run the scripts and notebooks inside the folder experiments.
Results
Results on test set and logic formulas will be saved in the folder experiments/results.
Data
The original datasets can be downloaded from the links provided in the supplementary material of the paper.
Theory
Theoretical foundations can be found in the following papers.
Logic Explained Networks:
@article{ciravegna2021logic, title={Logic explained networks}, author={Ciravegna, Gabriele and Barbiero, Pietro and Giannini, Francesco and Gori, Marco and Li{\'o}, Pietro and Maggini, Marco and Melacci, Stefano}, journal={arXiv preprint arXiv:2108.05149}, year={2021} }
Entropy-based LENs:
@article{barbiero2021entropy, title={Entropy-based Logic Explanations of Neural Networks}, author={Barbiero, Pietro and Ciravegna, Gabriele and Giannini, Francesco and Li{\'o}, Pietro and Gori, Marco and Melacci, Stefano}, journal={arXiv preprint arXiv:2106.06804}, year={2021} }
Psi network (“learning of constraints”):
@inproceedings{ciravegna2020constraint, title={A Constraint-Based Approach to Learning and Explanation.}, author={Ciravegna, Gabriele and Giannini, Francesco and Melacci, Stefano and Maggini, Marco and Gori, Marco}, booktitle={AAAI}, pages={3658--3665}, year={2020} }
Learning with constraints:
@inproceedings{marra2019lyrics, title={LYRICS: A General Interface Layer to Integrate Logic Inference and Deep Learning}, author={Marra, Giuseppe and Giannini, Francesco and Diligenti, Michelangelo and Gori, Marco}, booktitle={Joint European Conference on Machine Learning and Knowledge Discovery in Databases}, pages={283--298}, year={2019}, organization={Springer} }
Constraints theory in machine learning:
@book{gori2017machine, title={Machine Learning: A constraint-based approach}, author={Gori, Marco}, year={2017}, publisher={Morgan Kaufmann} }
Licence
Copyright 2020 Pietro Barbiero, Francesco Giannini, Gabriele Ciravegna, and Dobrik Georgiev.
Licensed under the Apache License, Version 2.0 (the “License”); you may not use this file except in compliance with the License. You may obtain a copy of the License at: http://www.apache.org/licenses/LICENSE-2.0.
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an “AS IS” BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and limitations under the License.
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Hashes for torch_explain-1.0.0-py3-none-any.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | 762ca47cdc02dda4ce40087b6ef68b709cc8b82b8667d1dce0544b8ff593bfe4 |
|
MD5 | a4146ae9b59a872af70e972bf62eec93 |
|
BLAKE2b-256 | 2e6707c3b63299de9a236ca1da9a86f29dcd976b0bc655b7b202ff5d560c2791 |