This repository contains the fastai DAWNbench result adapted to training on 1080ti. It is missing many of the optimizations that allowed the fastai team to achieve 94% accuracy in 2m 54s (no fp16, no data prefetching, etc) on an AWS p3.16xlarge instance with 8 V100 GPUs. On my box with a single 1080ti I am able to train to 94% accuracy (with TTA) in 13 minutes 30 seconds.
The second notebook adapts recent work by fastai and trains with AdamW and the 1 cycle policy cutting down the number of required epochs to 18. You can read more about this approach on the fastai blog or in the official repositiory.
You will need to have docker and nvidia-docker installed in order to run this.
Once you start the docker container, all you have to do is access https://localhost:8888 and enter jupyter as password. Open the notebook and hit run all.
- cd into cloned repo
docker build -t cifar ../run_container.sh cifar
SIDENOTE: You might need to run the commands with sudo. I prefer to do the following:
sudo groupadd docker
sudo usermod -aG docker $USER
(this effectively grants docker sudo powers so is not more secure than running docker with sudo)