PyGlove is a library for building Python programs that are easily evolvable, whether it be manually or automatically. For example, to find an optimal neural architecture for an image classifier, the architecture needs to be evolved by humans or by machines. With adding a single line of code, PyGlove upgrades a regular Python class into a symbolic class, whose objects can be manipulated safely at run time as if their source code were modified. On top of that, meta-programs can thrive, which not only add new aspects to such classes with ease (e.g. serialization, visualization), but also invites algorithms to participate in the modifications of the program automatically, enabling applications such as program search.
PyGlove was created for scaling automated machine learning (AutoML) in research and application. Its method was published in 2020 and has also been used widely in Google since 2019: through PyGlove, a better algorithm has been found on top of DQN by evolving reinforcement learning algorithms; multi-shot and one-shot neural architecture search were unified under the same programming interface; the co-design of neural architectures and hardware accelerators was possible for maximizing the model performance while meeting the hardware constraints. PyGlove is also being used in key products under Alphabet, such as Waymo's self-driving cars, Youtube's recommendation systems, Google Cloud's AI solutions, and etc.
Behind the scene, PyGlove introduces symbolic object-oriented programming, a new programming paradigm that adds symbolic programmability to objects, making it a useful tool for programs that need mutable representations, such as neural modeling, or a machine learning program as a whole. Moreover, PyGlove enhances the programmability for daily programming tasks. It is carefully implemented to be a general-purpose programming toolkit, and has no dependencies beyond the Python interpreter. PyGlove is developed by Daiyi Peng and his colleagues in Google Research Brain team.
pip install pyglove
import pyglove as pg
@pg.symbolize
class Hello:
def __init__(self, subject):
self._greeting = f'Hello, {subject}!'
def greet(self):
print(self._greeting)
hello = Hello('World')
hello.greet()Hello, World!
hello.rebind(subject='PyGlove')
hello.greet()Hello, PyGlove!
hello.rebind(subject=pg.oneof(['World', 'PyGlove']))
for h in pg.iter(hello):
h.greet()Hello, World!
Hello, PyGlove!
@inproceedings{peng2020pyglove,
title={PyGlove: Symbolic programming for automated machine learning},
author={Peng, Daiyi and Dong, Xuanyi and Real, Esteban and Tan, Mingxing and Lu, Yifeng and Bender, Gabriel and Liu, Hanxiao and Kraft, Adam and Liang, Chen and Le, Quoc},
booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
volume={33},
pages={96--108},
year={2020}
}
Disclaimer: this is not an officially supported Google product.