Welcome to our public repository, implementing the extraction algorithm from our ICML 2018 paper, Extracting Automata from Recurrent Neural Networks Using Queries and Counterexamples.
To use the main notebook here without installing anything, you can go straight to google colaboratory: https://drive.google.com/file/d/1tkJK1rJVEg9e-QcWOxErDb3cQq9UR-yR/view?usp=sharing
Open the dfa_from_rnn
notebook for a full demonstration and run through of how to use it yourself (we have provided all of the Tomita grammars, but you can also define, train, and extract your own languages!). For the impatient, the dfa_from_rnn_no_documentation
notebook is exactly like dfa_from_rnn
, only without all the explanation blocks. And if you want to train and keep track of several RNNs, you can use dfa_from_rnn_notebook_for_several_rnns
, which is like dfa_from_rnn_no_documentation
only it keeps all of your RNNs in neat little wrappers with their target languages and then keeps all of those in a list.
Everything here is implemented in Python 3. To use these notebooks, you will also need to install:
- DyNet (for working with our LSTM and GRU networks, which are implemented in DyNet)
- Graphviz (for drawing the extracted DFAs).
- NumPy and SciPy (for Scikit-Learn)
- Scikit-Learn (for the SVM classifier)
- Matplotlib (for plots of our networks' loss during training)
- Jupyter (for the python notebooks themselves)
If you are on a mac using Homebrew, then NumPy, SciPy, Scikit-Learn, Matplotlib, Graphviz and Jupyter should all hopefully work with brew install numpy
, brew install scipy
, etc.
If you don't have Homebrew, or wherever brew install
doesn't work, try pip install
instead.
For Graphviz you may first need to download and install the package yourself Graphviz, after which you can run pip install graphviz
. If you're lucky, brew install graphviz
might take care of all of this for you by itself. On colab, we got Graphviz using pip install graphviz
and then apt-get install graphviz
.
DyNet is installed by pip install dynet
from the command line (for the basic CPU version. For the GPU version, check their site).
You can also apply the code directly to your own networks without most of these packages. The main extraction function is in Extraction.py
and called extract
. You can run it on any network that implements the API described in our dfa_from_rnn
notebook, which is viewable in-browser in git even if you don't have Jupyter, and reiterated here for completeness.
classify_word(word)
returns a True or False classification for a word over the input alphabetget_first_RState()
returns a tuple (v,c) where v is a continuous vector representation of the network's initial state (an RState), and c is a boolean signifying whether it is an accepting stateget_next_RState(state,char)
given an RState, returns the next RState the network goes to on input characterchar
, in the same format asget_first_RState
(i.e., a tuple (v,c) of vector + boolean)
To run only the extraction code you will only need the NumPy, SciPy, Scikit-Learn, and Graphviz packages. If you want, you can also skip the Graphviz package, at the cost of the ability to visualise your DFAs. Remove the graphviz import from DFA.py
and set the body of the draw_nicely
function of the DFA
class to pass
. You only need the DFA
, Extraction
, Lstar
, Helper_Functions
, Observation_Table
, Quantisations
, Teacher
, and WhiteboxRNNCounterexampleGenerator
modules for extraction.
You can cite this work using:
@InProceedings{weiss-goldberg-yahav,
title = {Extracting Automata from Recurrent Neural Networks Using Queries and Counterexamples},
author = {Gail Weiss and Yoav Goldberg and Eran Yahav},
booktitle = {Proceedings of the 35th International Conference on Machine Learning},
year = {2018},
editor = {Jennifer Dy and Andreas Krause},
volume = {80},
series = {Proceedings of Machine Learning Research},
address = {Stockholmsm\"{a}ssan, Stockholm, Sweden},
month = {10--15 Jul},
publisher = {PMLR}
}