Skip to content
This repository has been archived by the owner on Jan 31, 2022. It is now read-only.
/ gtn Public archive

Automatic differentiation with weighted finite-state transducers.

License

Notifications You must be signed in to change notification settings

facebookresearch/gtn

Repository files navigation

logo

Deprecated

This repository has been deprecated.

Future development will be in the GTN org repository.

GTN: Automatic Differentiation with WFSTs

Quickstart | Installation | Documentation

facebookresearch Documentation Status

What is GTN?

GTN is a framework for automatic differentiation with weighted finite-state transducers. The framework is written in C++ and has bindings to Python.

The goal of GTN is to make adding and experimenting with structure in learning algorithms much simpler. This structure is encoded as weighted automata, either acceptors (WFSAs) or transducers (WFSTs). With gtn you can dynamically construct complex graphs from operations on simpler graphs. Automatic differentiation gives gradients with respect to any input or intermediate graph with a single call to gtn.backward.

Also checkout the repository gtn_applications which consists of GTN applications to Handwriting Recognition (HWR), Automatic Speech Recognition (ASR) etc.

Quickstart

First install the python bindings.

The following is a minimal example of building two WFSAs with gtn, constructing a simple function on the graphs, and computing gradients. Open In Colab

import gtn

# Make some graphs:
g1 = gtn.Graph()
g1.add_node(True)  # Add a start node
g1.add_node()  # Add an internal node
g1.add_node(False, True)  # Add an accepting node

# Add arcs with (src node, dst node, label):
g1.add_arc(0, 1, 1)
g1.add_arc(0, 1, 2)
g1.add_arc(1, 2, 1)
g1.add_arc(1, 2, 0)

g2 = gtn.Graph()
g2.add_node(True, True)
g2.add_arc(0, 0, 1)
g2.add_arc(0, 0, 0)

# Compute a function of the graphs:
intersection = gtn.intersect(g1, g2)
score = gtn.forward_score(intersection)

# Visualize the intersected graph:
gtn.draw(intersection, "intersection.pdf")

# Backprop:
gtn.backward(score)

# Print gradients of arc weights 
print(g1.grad().weights_to_list()) # [1.0, 0.0, 1.0, 0.0]

Installation

Requirements

  • A C++ compiler with good C++14 support (e.g. g++ >= 5)
  • cmake >= 3.5.1, and make

Python

Install the Python bindings with

pip install gtn

Building C++ from source

First, clone the project:

git clone [email protected]:facebookresearch/gtn.git && cd gtn

Create a build directory and run CMake and make:

mkdir -p build && cd build
cmake ..
make -j $(nproc)

Run tests with:

make test

Run make install to install.

Python bindings from source

Setting up your environment:

conda create -n gtn_env
conda activate gtn_env

Required dependencies:

cd bindings/python
conda install setuptools

Use one of the following commands for installation:

python setup.py install

or, to install in editable mode (for dev):

python setup.py develop

Python binding tests can be run with make test, or with

python -m unittest discover bindings/python/test

Run a simple example:

python bindings/python/examples/simple_graph.py

License

GTN is licensed under a MIT license. See LICENSE.

About

Automatic differentiation with weighted finite-state transducers.

Resources

License

Code of conduct

Security policy

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published