os-hxfan/BayesNN_FPGA

FPGA-based hardware acceleration for dropout-based Bayesian Neural Networks.

FPGA-based Accelerator for Bayesian Neural Networks

FPGA-based hardware acceleration for dropout-based Bayesian Neural Networks (BayesNNs). We support both multi-exit Monte Carlo Dropout (MCD) and multi-exit Masksembles on FPGA.

This repo contains the artifacts of our DAC'23 paper and TCAD'23 submission. Pls cite us if you found it helpful. The software is based on Pytorch, and the hardware implementation is based on HLS4ML and QKeras. Pls cite our paper and give this repo ⭐ if you feel the code and paper are helpful!

Structure

.
├── README.md  
├── Software_Artifact                     # Evaluating accuracy and ECE for MCME-based networks
    ├── README.md    
    ├── requirements.txt
    ├── script_tables          # Scripts to generate the tables in the paper
    ├── software               # Code to build and train multi-exit BNNs
        ├── main.py
        ├── datasets           # Code containing dataset files and loaders
           ├── data            # Code containing the train/test data
        ├── models             # Code containing model files and loaders
           ├── model_weights   # Code containing model weight files
        ├── train              # Code containing training files
           ├── loss            # Code containing loss loaders
        ├── snapshots          # Checkpoints of the model during training
           ├── figures         # Figures of the loss vs epoch
        ├── runs_db            # Database containing hyperparameters and performance of each experiment
        ├── logs               # Logs containing outputs used in paper
├── Hardware_Artifact          # Evaluating hardware performance of proposed FPGA-based accelerator
  ├── README.md
  ├── autobayes                # All the scripts to generate data in the paper
      ├── models               # Keras models used for the experiment and testing
      ├── timing               # Timing reports generated here
      ├── diff_dropouts        # Synthesis reports for BNN with different number of dropout layers generated here
  ├── converter                # All code related to converting NN to BNN
      ├── keras                # Code for converting keras NN model to BNN model
      ├── pytorch              # Code for converting torch NN model to BNN model
  ├── bayes_hw                 # FPGA-based hardware accelerator for MCME, based on qkears and hls4ml 
      ├── models               # Keras model definition
      ├── scripts              # Scripts to generate the tables and figures in the paper
  └── requirements.txt   

How to Run

1. Setup environment by following README.md under both ./Software_Artifact and ./Hardware_Artifact. Virtual environment management system such as Conda is highly recommended.
2. Run the scripts under both folder to generate the results shown in the paper.

Citation

Our paper is online now (link)! If you found it helpful, pls cite us using:

@inproceedings{fan2023when,
  title={When Monte-Carlo Dropout Meets Multi-Exit: Optimizing Bayesian Neural Networks on FPGA},
  author={Fan, Hongxiang and Chen, Hao and Castelli, Liam and Que, Zhiqiang and Li, He and Long, Kenneth and Luk, Wayne},
  booktitle={Proceedings of the 60th ACM/IEEE Design Automation Conference},
  year={2023}
}