Skip to content
master
Go to file
Code

Latest commit

 

Git stats

Files

Permalink
Failed to load latest commit information.
Type
Name
Latest commit message
Commit time
L4
 
 
 
 
 
 
 
 
 
 

README.md

L4 Stepsize Adaptation Scheme

By Michal Rolínek, Georg Martius.

Autonomous Learning Group, Max Planck Institute for Intelligent Systems.

Table of Contents

  1. Introduction
  2. Requirements
  3. Installation
  4. Usage
  5. Notes

Introduction

This repository contains TensorFlow implementation code for the paper "L4: Practical loss-based stepsize adaptation for deep learning". This work proposes an explicit rule for stepsize adaptation on top of existing optimizers such as Adam or momentum SGD.

Disclaimer: This code is a PROTOTYPE and most likely contains bugs. It should work with most Tensorflow models but most likely it doesn't comply with TensorFlow production code standards. Use at your own risk.

Requirements

TensorFlow >= 1.4.

For a PyTorch implementation of L4, see l4-pytorch (also linked here as a submodule).

Installation

Either use one of the following python pip commands,

python -m pip install git+https://github.com/martius-lab/l4-optimizer
python3 -m pip install git+https://github.com/martius-lab/l4-optimizer

or simply drop the L4/L4.py file to your project directory

Usage

Exaclty as you would expect from a TensorFlow optimizer. Empirically, good values for the 'fraction' parameter are 0.1 < fraction < 0.3, where 0.15 is set default and should work well enough in most cases. Decreasing 'fraction' is typically a good idea in case of a small batch size or more generally for very little signal in the gradients. Too high values of 'fraction' behave similarly to too high learning rates (i.e. divergence or very early plateauing).

import L4

...
opt = L4.L4Adam(fraction=0.20)
opt.minimize(loss)
...

or

import L4

...
opt = L4.L4Mom()  # default value fraction=0.15 is used
grads_and_vars = opt.compute_gradients(loss)
...
# Gradient manipulation
...

opt.apply_gradients(grads_and_vars) # (!) Passing the loss is no longer needed (!)
...

Notes

Contribute: If you spot a bug or some incompatibility, contribute via a pull request! Thank you!

About

Code for paper "L4: Practical loss-based stepsize adaptation for deep learning"

Resources

License

Releases

No releases published

Packages

No packages published