Boilerplate repository showing how to speedrun from idea to a reproducible experiment setup with Pytorch.
This repository contains code to train a ResNet on Cifar-10. To get started: install speedrun, fork this repository and hack away! Make sure to also install tensorboardX while you're at it.
To run the example code on CPU, do the following.
cd speedrun-springboard
mkdir experiments
python train.py experiments/FIRST-0 --inherit templates/BASE-X
This will make a new experiment directory experiments/FIRST-0
(see speedrun for more on that).
Now to train on GPU with a different learning rate, you could do the following:
python train.py experiments/FIRST-1 --inherit experiments/FIRST-0 --config.device 'cuda:0' --config.optimizer.kwargs.lr 0.01
This will make another experiment directory experiments/FIRST-1
.