Skip to content

[ICML 2023] Official code for "DevFormer: A Symmetric Transformer for Context-Aware Device Placement"

License

Notifications You must be signed in to change notification settings

kaist-silab/devformer

Repository files navigation

DevFormer

Streamlit Appexample workflow Code style: blackpython_sup

This repository contains the benchmark for the decoupling capacitor placement problem (DPP) and the accompanying paper "DevFormer: A Symmetric Transformer for Context-Aware Device Placement", accepted at ICML 2023. The benchmark is designed to evaluate the performance of the proposed DevFormer architecture and to facilitate future research in hardware design optimization.

The DPP and its multi-port variants are now available as environments in RL4CO! 🚀

Setup

Install dependencies

# Clone the repository and cd into it
git clone https://github.com/kaist-silab/DPP_benchmark.git && cd DPP_benchmark

# Automatically install dependencies with light the torch
pip install light-the-torch && python3 -m light_the_torch install --upgrade -r requirements.txt

The above script will automatically install PyTorch with the right GPU version for your system. Alternatively, you can use pip install -r requirements.txt

Usage

Simulator

  • Using the simulator to obtain the cost of a solution:
from src.problems.dpp.simulator import decap_sim

cost = decap_sim(probe = 23, solution = [1,5,7], keep_out = [2,3,10])

Run and evaluate models

In general, the following command is used to run the models:

python3 run.py --problem [PROBLEM] --model [MODEL] --training_dataset [DATASET]

You may also have a look at the arguments under src/options.py for more details.

  • How to evaluate pretrained DevFormer
python3 run.py --problem dpp --model devformer --resume data/dpp/pretrained/CSE_2000_epoch-50.pt --eval_only
  • How to train DevFormer
python3 run.py --problem dpp --model devformer --N_aug 4 --training_mode IL --train_dataset data/dpp/training_2000_new.pkl --guiding_action data/dpp/guiding_2000_new.pkl --EE --SE --batch_size 200

Additionally, the folder scripts/ contains scripts to reproduce the results in the paper.

Troubleshooting

  • There may be problems on multiple GPUs due to the current handling of DataParallel. You may run export CUDA_VISIBLE_DEVICES=0 to use only one GPU.
  • When running the run.py script, if data has not been download it will start downloading automatically. If you want to download the data manually, or if there are any issues with Google Drive, you may access the data at the following link and place extract the content of the .zip archive at this repository root ..

DPP Simulator GUI

The application is based on Streamlit which allows for web GUIs in Python. To run the application locally, run the following command:

streamlit run app.py

A web browser should open automatically and you can interact with the application. If it doesn't, you can manually open a browser and navigate to http://localhost:8501.

Notes on GUI development

The structure of the application is as follows:

├── app.py # landing page to `streamlit run`
└── pages/
    ├── about.py # about page in Python (as per Streamlit documentation)
    ├── assets/
    |   └── * # media such as .png images
    └── src/
       └── script.js # javascript file for modifying the GUI

Most radical modifications are not supported in Streamlit, so we hack our way and inject Javascript code to modify elements of the GUI.

Deploy the app

There are many ways to deploy the app, among which on our own server. However, Streamlit provides a free hosting service that is sufficient for our purposes. To deploy the app, simply follow the instructions there or click the "deploy" button after running the app locally!


Cite us

If you find this repository useful, please consider citing our paper:

@article{kim2023devformer,
  title={DevFormer: A Symmetric Transformer for Context-Aware Device Placement},
  author={Kim, Haeyeon and Kim, Minsu and Berto, Federico and Kim, Joungho and Park, Jinkyoo},
  year={2023},
  booktitle={International Conference on Machine Learning},
  organization={PMLR}
}