tdl77 / lrschedules

An implementation of many different types of schedules for the learning rate.

The Best Learning Rate Schedules

 

This is a basic code respository to support an overview that I wrote in my Deep (Learning) Focus newsletter regarding common learning rate schedules that are used for training neural networks. This code can be used to reproduce:

1. Triangular/Cyclical Learning Rates
2. Learning Rate Decay with Restarts
3. Learning Rate Schedules based on Profiles and Sampling Rates

Environment/Dependencies

Requires anaconda to be installed (python3) Anaconda can be installed at https://www.anaconda.com/products/individual

conda create -n lr-sched python=3.8 anaconda
conda activate lr-sched
pip install -r requirements.txt

File Arrangement

Here we summarize all files present in this repo and their purpose.

+-- figs/ : figures of different learning rate schedules saved here
+-- profiles.py : implements different growth/decay profiles
+-- sampling_rate_examples.py : generates learning rate schedules using profile + sampling rate
+-- cyclical_examples.py : generates cyclically learning rate schedules with different profiles
+-- sgdr_samples.py : generates learning rate schedules for SGD with restarts (SGDR)
+-- update_hparams.py : functions for updating LR/momentum in PyTorch optimizers