109 2nd-year EDA project DEMO

Overview

This repository contains the PyTorch implementation of 109 2nd-year EDA project DEMO.

Dependencies

Please refer to requirements.txt

User can use the command shows below to install dependencies:

pip3.6 install -r requirements.txt

Usage

For test results for alexnet_brevitas:

python3.6 -O main.py \
   	--batch-size 32 \
   	--workers 16 \
   	--seed 11111 \
   	--sess brevitas \
   	--resume

For training alexnet_brevitas and export finn-onnx model:

python3.6 -O main.py \
	--batch-size 32 \
   	--workers 16 \
   	--seed 11111 \
	--lr 0.1 \
	--epochs 100 \
   	--sess brevitas \
   	--train \
	--export_finn

Benchmark on Google Speech Commands (12-cmd)

The following table shows the train accuracy and test accuracy in a 100-epoch training session.

Model	Training	Testing	Checkpoint
alexnet_brevitas	76.263%	73.794%	ckpt.t7.brevitas_11111.pth

To use pre-trained checkpoint, please download checkpoint .pth files into checkpoint folder.

Acknowledgement

The project is adapted from the COT repository by henry8527.

The models are adapted from the BinaryNet.pytorch repository by itayhubara, M5 refer from Speech Command Recognition with torchaudio

The dataset is adapted from the Speech Command Recognition repository by douglas125