ReAct

A (summer) project to embed adaptive-computation mechanisms in attention based models as an inductive bias to help towards OOD extrapolation.

Writeup/Website: Notion Website

WandB link: here

Code: You're here 😉

Vast.ai

To reproduce the training runs: Find the precise commit id you want to replicate from the WandB or simply use the defaults.

Navigate to the docker "temporary slot" and enter the Docker image creds:

DockerHub location: neel04/react_image:latest
version tag: latest

Paste this as the onstart.sh script. Make sure to fill in the appropriate VAST and Wandb.ai key.

For commit, you can use the default or use the commit id from the WandB runs. Here are some I'd recommend:

bAdd: ac3b5a4bf328e01ea1f37ccdbe0cd1053f0abe53
reverse_string : 07af7653514a976d3e0355d544f32c1693a563c6
prefix_sum : e006c2e859bfab3106c110f76a065fbe3a89fb45

export VAST_API_KEY=...
export WANDB_API_KEY=...

export TASK="main"  # refers to the branches of repo. "main" is `bAdd`, then `reverse_string` and `prefix_sum`
export COMMIT_ID="default"  # "default" will not checkout any commit

cd /workspace/
rm -rf /fsx/

# VAST config to shutdown instance after the job is done
wget https://raw.githubusercontent.com/vast-ai/vast-python/master/vast.py -O vast; chmod +x vast;
echo $VAST_API_KEY > ~/.vast_api_key

# Clone ReAct repository and move it to /fsx/awesome/DPT/
git clone -b $TASK https://github.com/neel04/ReAct.git /fsx/awesome/DPT/
cd /fsx/awesome/DPT

# If commit id is not default, checkout that commit
if [ "$COMMIT_ID" != "default" ]; then
  git checkout $COMMIT_ID
fi

# Create a directory for the outputs
mkdir -p /fsx/awesome/DPT/outputs

# Change directory to /fsx/awesome/DPT/
cd /fsx/awesome/DPT

# Set CUDA memory allocation configuration to max_split_size_mb:512
export PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:512

# Run React training
chmod +x ./dpt_exec.sh
sh ./dpt_exec.sh

# Stop the instance
cd /workspace; ./vast stop instance ${VAST_CONTAINERLABEL:2}

Rent a 3090 instance (24Gb+ VRAM) and relax as it caches the docker file and runs the training.
Enjoy

Credits

Credits go to Avi Schwarzchild's and Arpit Bansal's (et al.) repository on which this code is built on. Check out their amazing work here!

Huge thanks to Algovera.ai for sponsoring this project 🚀!

Important commands

Docker container command:

docker run -it --rm -v /workspaces/ReAct:/fsx/awesome/DPT -w /fsx/awesome/DPT neel04/react_image:latest
# ... git clone and stuff
cd /fsx/awesome/DPT; sh ./dpt_exec.sh

Runs the training script by executing DeepThinking.ipynb, which in turn modifies some files and configs, finally executing the main trigger program.

(This rigmarole was due to this codebase originally working with SLURM on an HPC and then I never cleaned it up... But I suppose that's a story for another time.)

Made with 🩸, 💧 & 😢

Email: neelgupta04@outlook.com

Github: neel04 (links to the code for this project)

Discord: awesome_ruler_007 - or you can usually find me on Yannic's server or "Learn AI Together”