ray air cifar10 demo for ray meetup

Setup: This demo uses Anyscale workspace. g4dn.12xlarge for both head node and worker node. Overall 8 GPUs. Each trial is using 4GPUs. Demo using Ray Tune to distribute two trials.

Note: Before running the demo, please make sure that you run prepare_data.py on both the head node and worker node. To ssh onto worker node, do the following (replace with the right worker node ip address. You can find it in Ray Dashboard)

ssh -i $HOME/ray_bootstrap_key.pem [email protected]
ubuntu@ip-172-31-13-42:~$ docker exec -it ray_container bash

The training/tuning takes about 11min. best_checkpoint folder stores the best checkpoint for demoing the following operations, so that we don't have to wait!

To get an AIR checkpoint out of it, simply do

from ray.air.checkpoint import Checkpoint
best_checkpoint = Checkpoint.from_directory("./best_checkpoint")