BNN-PYNQ

See FINN for the publications

https://www.hackster.io/adam-taylor/training-implementing-a-bnn-using-pynq-1210b9

LFC: 4 fully connected network for 28 x 28 x binary (black/white) input
CNV: 6 convolution + 2 max pooling + 3 fully connected network for 32 x 32 x 3×8-bits (RGB) input
The output activation is determined by the sign function: {+1 if x >= 0, -1 if x < 0}

See PYNQ-Z1 for the installation

sudo apt-get update
sudo apt-get upgrade

#sudo pip3 install git+https://github.com/Xilinx/BNN-PYNQ.git
sudo pip3 install git+https://github.com/Xilinx/BNN-PYNQ.git --no-build-isolation

http://192.168.2.99/

## Run bnn/CNV-QNN_Cifar10.ipynb on Jupyter Notebook

Training could be not on the FPGA board but on AWS/GCP with GPU
Install CUDA 10.2 + cuDNN 7.6.5 on Ubuntu 18.04 for faster training with Nvidia GPU
- Latest CUDA is incompatible with training scripts
Install Anaconda for easy installation
CPU Runtime: 1000 sec/epoch $\times$ 500 epoch = 140 hours
GPU Runtime: 30 sec/epoch $\times$ 500 epoch = 4 hours (Tesla T4)
Python 2.7 is necessary
Original steps are not functional with the latest libraries in 2023

https://github.com/Xilinx/BNN-PYNQ/tree/master/bnn/src/training

## Dependencies
sudo apt-get install -y git python-dev libopenblas-dev liblapack-dev gfortran
sudo apt-get install g++

## Pip
#wget https://bootstrap.pypa.io/get-pip.py && python get-pip.py --user
wget https://bootstrap.pypa.io/pip/2.7/get-pip.py && python get-pip.py --user

## PATH
export PATH="$HOME/.local/bin${PATH:+:${PATH}}"

## Theano/Lasagne
pip uninstall pyyaml
pip install pyyaml
pip install --user pathlib
pip install --user git+https://github.com/Theano/Theano.git@rel-0.9.0beta1
pip install --user https://github.com/Lasagne/Lasagne/archive/master.zip

cat << EOF > ~/.theanorc
[global]
floatX = float32
device = gpu
openmp = True
openmp_elemwise_minsize = 200000

[nvcc]
fastmath = True

[blas]
ldflags = -lopenblas

[warn]
round = False
EOF

## PyLearn2
pip install --user numpy==1.11.0
git clone https://github.com/lisa-lab/pylearn2
cd pylearn2
python setup.py develop --user
cd

## Data Download
export PYLEARN2_DATA_PATH=~/.pylearn2
mkdir -p ~/.pylearn2
cd pylearn2/pylearn2/scripts/datasets
python download_mnist.py
./download_cifar10.sh
cd

Theano shows some errors here with the latest CUDA 11.8 + cuDNN 8.9.1
The following section is a workaround solution in 2023

https://qiita.com/harmegiddo/items/6d7490214b693150cef9

## Anaconda
conda create --name pynq python=2.7
conda activate pynq
conda install tensorflow=2.1.0=gpu_py27h9cdf9a9_0
conda install pygpu

pip install cython
pip install keras==2.3.1
pip install theano==1.0.5
pip install https://github.com/Lasagne/Lasagne/archive/master.zip

cat << EOF > ~/.theanorc
[global]
floatX = float32
device = cuda
openmp = True
openmp_elemwise_minsize = 200000

[nvcc]
fastmath = True

[blas]
ldflags = -lopenblas

[warn]
round = False
EOF

git clone https://github.com/Xilinx/BNN-PYNQ.git
export XILINX_BNN_ROOT="$HOME/BNN-PYNQ/bnn/src/"

cd ./BNN-PYNQ/bnn/src/training/

The original cifar10.py is modified to avoid using pylearn2 and to use keras for downloading data:
bnn-pynq_cifar10_mod.py.gz

python cifar10_mod.py
#python cifar10_mod.py -wb 1 -ab 2
#python cifar10_mod.py -wb 2 -ab 2

Using cuDNN version 7605 on context None
Mapped name None to device cuda: Tesla T4 (0000:00:04.0)
Using TensorFlow backend.
activation_bits = 1
weight_bits = 1
batch_size = 50
alpha = 0.1
epsilon = 0.0001
W_LR_scale = Glorot
num_epochs = 500
LR_start = 0.001
LR_fin = 3e-07
LR_decay = 0.983907435305
save_path = cifar10-1w-1a.npz
train_set_size = 45000
shuffle_parts = 1
Loading CIFAR-10 dataset...
...
Training...
Epoch 1 of 500 took 31.9853491783s
  LR:                            0.001
  training loss:                 0.715714348786407
  validation loss:               0.3401033526659012
  validation error rate:         57.180000364780426%
  best epoch:                    1
  best validation error rate:    57.180000364780426%
  test loss:                     0.33863225907087324
  test error rate:               56.81000013649463%

python cifar10-gen-weights-W1A1.py
#python cifar10-gen-weights-W1A2.py
#python cifar10-gen-weights-W2A2.py

ls cnvW1A1

0-0-thres.bin     0-9-thres.bin     1-21-thres.bin    1-7-thres.bin     2-6-thres.bin     3-5-thres.bin
0-0-weights.bin   0-9-weights.bin   1-21-weights.bin  1-7-weights.bin   2-6-weights.bin   3-5-weights.bin
0-1-thres.bin     1-0-thres.bin     1-22-thres.bin    1-8-thres.bin     2-7-thres.bin     3-6-thres.bin
0-1-weights.bin   1-0-weights.bin   1-22-weights.bin  1-8-weights.bin   2-7-weights.bin   3-6-weights.bin
...
0-6-thres.bin     1-19-thres.bin    1-4-thres.bin     2-3-thres.bin     3-2-thres.bin     8-0-weights.bin
0-6-weights.bin   1-19-weights.bin  1-4-weights.bin   2-3-weights.bin   3-2-weights.bin   8-1-weights.bin
0-7-thres.bin     1-2-thres.bin     1-5-thres.bin     2-4-thres.bin     3-3-thres.bin     8-2-weights.bin
0-7-weights.bin   1-2-weights.bin   1-5-weights.bin   2-4-weights.bin   3-3-weights.bin   8-3-weights.bin
0-8-thres.bin     1-20-thres.bin    1-6-thres.bin     2-5-thres.bin     3-4-thres.bin     classes.txt
0-8-weights.bin   1-20-weights.bin  1-6-weights.bin   2-5-weights.bin   3-4-weights.bin   hw

## Copy weight files to /usr/local/lib/python3.6/dist-packages/bnn/params/original/cnvW1A1 on PYNQ
cd /usr/local/lib/python3.6/dist-packages/bnn/params/
mkdir -p original/cnvW1A1

## Run bnn/CNV-QNN_Cifar10.ipynb on Jupyter Notebook

CPU Runtime: 100 sec/epoch $\times$ 1000 epoch = 30 hours
GPU Runtime: 2 sec/epoch $\times$ 1000 epoch = 0.5 hours (Tesla T4)

The original mnist.py is modified to avoid using pylearn2 and to use keras for downloading data:
bnn-pynq_mnist_mod.py.gz

python mnist_mod.py
#python mnist_mod.py -wb 1 -ab 2

Using cuDNN version 7605 on context None
Mapped name None to device cuda: Tesla T4 (0000:00:04.0)
Using TensorFlow backend.
activation_bits = 1
weight_bits = 1
batch_size = 100
alpha = 0.1
epsilon = 0.0001
num_epochs = 1000
dropout_in = 0.2
dropout_hidden = 0.5
W_LR_scale = Glorot
LR_start = 0.003
LR_fin = 3e-07
LR_decay = 0.990831944893
save_path = mnist-1w-1a.npz
train_set_size = 50000
shuffle_parts = 1
Loading MNIST dataset...
...
Training...
Epoch 1 of 1000 took 2.09121918678s
  LR:                            0.003
  training loss:                 0.26854227716475726
  validation loss:               0.044894146784208715
  validation error rate:         6.429999990388752%
  best epoch:                    1
  best validation error rate:    6.429999990388752%
  test loss:                     0.04777069476665929
  test error rate:               6.899999972432852%

https://github.com/mohdumar644/TinyYOLO-BNN

https://marsee101.web.fc2.com/pynq.html
https://www.tkat0.dev/study-BNN-PYNQ/
https://qiita.com/harmegiddo/items/6d7490214b693150cef9
https://qiita.com/ykshr/items/6c8cff881a200a781dc3
https://discuss.pynq.io/t/new-pynq-2-7-image-got-modulenotfounderror-no-module-named-pynq-when-install-bnn-pynq/3570/5
https://github.com/Xilinx/BNN-PYNQ/tree/master/bnn/src/training
https://qiita.com/ykshr/items/08147098516a45203761
https://gist.github.com/ykshr/a68403f5b073ba8f117faeb0a14385c1
https://qiita.com/kamotsuru/items/dcd460e04385f2738cd7

Daiphys is a professional-service company for research and development of leading-edge technologies in science and engineering.
Get started accelerating your business through our deep expertise in R&D with AI, quantum computing, and space development; please get in touch with Daiphys today!

Daiphys Technologies LLC - https://www.daiphys.com/

BNN-PYNQ

Getting Started

Publications

Model Structures

Board Setup

Installation on Target Board

Sample Execution on Target Board (CNV-QNN_CIFAR10)

Sample Image

Processing Time: PL-FPGA (100 MHz) vs PS-CPU (650 MHz)

Class Rankings: CNVW1A1/CNVW1A2/CNVW2A2

Training on Host Machine (CNV-QNN_CIFAR10)

Tips

Install (Original/Failed)

Install (Successful)

Training

Weight

Run on Target Board (PYNQ-Z1)

Training on Host Machine (LFC-BNN_MNIST)

Tips

Training

Spin-Off Projects

References

Acknowledgments