Brevitas Fixed Point Training for Classification Models

This repo contains training scripts to train and evaluate popular deep learning classification models: Lenet, Alexnet and VGG using Brevitas.

Requirements

• Pytorch >= 1.1.0
• Brevitas (https://github.com/Xilinx/brevitas)

Install

After installing Pytorch, install Brevitas:

         git clone https://github.com/Xilinx/brevitas
         cd brevitas
         pip install .

Clone this repo

         git clone https://github.com/MinahilRaza/Brevitas_Fixed_Point.git

Results

This repo includes pretrained models with the following accuracies

Name	Dataset	Weight quantization	Activation quantization	Brevitas Top1
Lenet	MNIST	8 bit	8 bit	98.99%
Lenet	MNIST	8 bit for conv only	8 bit for conv only	99.08%
Alexnet	CIFAR10	8 bit	8 bit	85.42%
VGG	CIFAR10	8 bit	8 bit	86.22%

Train

In order to launch training, run the following commands

MNIST for Lenet

From within the training_scripts folder:

python run.py --network Lenet --dataset MNIST

AlexNet for CIFAR10

From within the training_scripts folder:

python run.py --network AlexNet --dataset CIFAR10

VGG for CIFAR10

From within the training_scripts folder:

python run.py --network VGG --dataset CIFAR10

Resuming Training and Evaluation

In order ro resume training from saved checkpoint use --resume flag. Set its value to true for resuming training. For evaluation on validation set, set --evaluate flag to true.

README format

This README format was inspired by brevitas_cnv_lfc (https://github.com/ussamazahid96/brevitas_cnv_lfc)