/bnn_pynq

Primary LanguagePython

BNN-PYNQ Brevitas experiments

This repo contains training scripts and pretrained models to recreate the LFC and CNV models used in the BNN-PYNQ repo using Brevitas. These pretrained models and training scripts are courtesy of Alessandro Pappalardo and Ussama Zahid.

Experiments

Name Input quantization Weight quantization Activation quantization Dataset Top1 accuracy
TFC_1W1A 1 bit 1 bit 1 bit MNIST 93.17%
TFC_1W2A 2 bit 1 bit 2 bit MNIST 94.79%
TFC_2W2A 2 bit 2 bit 2 bit MNIST 96.60%
SFC_1W1A 1 bit 1 bit 1 bit MNIST 97.81%
SFC_1W2A 2 bit 1 bit 2 bit MNIST 98.31%
SFC_2W2A 2 bit 2 bit 2 bit MNIST 98.66%
LFC_1W1A 1 bit 1 bit 1 bit MNIST 98.88%
LFC_1W2A 2 bit 1 bit 2 bit MNIST 98.99%
CNV_1W1A 8 bit 1 bit 1 bit CIFAR10 84.22%
CNV_1W2A 8 bit 1 bit 2 bit CIFAR10 87.80%
CNV_2W2A 8 bit 2 bit 2 bit CIFAR10 89.03%

Train

A few notes on training:

  • An experiments folder at /path/to/experiments must exist before launching the training.
  • Set training to 1000 epochs for 1W1A networks, 500 otherwise.
  • Enabling the JIT with the env flag BREVITAS_JIT=1 significantly speeds up training.

To start training a model from scratch, e.g. LFC_1W1A, run:

BREVITAS_JIT=1 brevitas_bnn_pynq_train --network LFC_1W1A --experiments /path/to/experiments

Evaluate

To evaluate a pretrained model, e.g. LFC_1W1A, run:

BREVITAS_JIT=1 brevitas_bnn_pynq_train --evaluate --network LFC_1W1A --pretrained

To evaluate your own checkpoint, of e.g. LFC_1W1A, run:

BREVITAS_JIT=1 brevitas_bnn_pynq_train --evaluate --network LFC_1W1A --resume /path/to/checkpoint.tar

Determined AI

First log in with port 8080:

ssh -L 8080:localhost:8080 <USERNAME>@prp-gpu-1.t2.ucsd.edu

Install a conda environment with determined (first time):

conda create python=3.6 -n determined
conda activate determined
pip install determined

Check out this repo (first time):

git clone https://github.com/jmduarte/bnn_pynq

Activate conda environment and enter repo (each time):

conda activate determined
cd bnn_pynq

Log in with your username and password

det user login $USER

Run a single experiment (with constant parameters):

det -m 'localhost' experiment create const.yaml .