Skip to content

Dataflow QNN inference accelerator examples on FPGAs

License

Notifications You must be signed in to change notification settings

hleblevec/finn-examples-fork

 
 

Repository files navigation

Dataflow Accelerator Examples

for PYNQ on Zynq and Alveo

drawing

This repository contains a variety of customized FPGA neural network accelerator examples built using the FINN compiler, which targets few-bit quantized neural networks with emphasis on generating dataflow-style architectures customized for each network.

The examples here come with pre-built bitfiles, PYNQ Python drivers and Jupyter notebooks to get started, and you can rebuild them from source. Both PYNQ on Zynq and Alveo are supported.

Need help with a problem in this repo, or got a question? Feel free to ask for help in the GitHub discussions. In the past, we also had a Gitter channel. Please be aware that this is no longer maintained by us but can still be used to search for questions previous users had.

Quickstart

We recommend PYNQ version 3.0.1, but older installations of PYNQ should also work. For PYNQ v2.6.1, please refer for set-up instructions to FINN-examples v0.0.5.

Zynq

For ZYNQ boards, all commands below must be prefixed with sudo or by first going into sudo su.

First, source the PYNQ and XRT virtual environment:

source /etc/profile.d/pynq_venv.sh
source /etc/profile.d/xrt_setup.sh

Next, ensure that your pip and setuptools installations are up-to-date on your PYNQ board:

python3 -m pip install pip==23.0 setuptools==67.1.0

Since we are going to install finn-examples without build-isolation, we need to ensure all dependencies are installed. For that, install setuptools_csm as well:

python3 -m pip install setuptools_scm==7.1.0

Install the finn-examples package using pip:

# remove previous versions with: pip3 uninstall finn-examples
pip3 install finn-examples --no-build-isolation
# to install particular git branch:
# pip3 install git+https://github.com/Xilinx/finn-examples.git@dev --no-build-isolation

Retrieve the example Jupyter notebooks using the PYNQ get-notebooks command. An example of how to run the Jupyter notebook server, assuming we are forwarding port 8888 from the target to some port on our local machine, is also shown below:

# on PYNQ boards, first cd /home/xilinx/jupyter_notebooks
pynq get-notebooks --from-package finn-examples -p . --force
jupyter-notebook --no-browser --allow-root --port=8888

Alveo

For Alveo we recommend setting up everything inside a virtualenv as described here.

First, create & source a virtual environment:

conda create -n <virtual-env> python=3.10
conda activate <virtual-env>

Next, ensure that your pip and setuptools installations are up-to-date:

python3 -m pip install --upgrade pip==23.0 setuptools==67.2.0

Finally, we can now install Pynq, FINN-examples and Jupyter (please note to source the XRT environment before):

pip3 install pynq==3.0.1
python3 -m pip install setuptools_scm==7.1.0 ipython==8.9.0
pip3 install finn-examples --no-build-isolation
# to install particular git branch:
# pip3 install git+https://github.com/Xilinx/finn-examples.git@dev --no-build-isolation
python3 -m pip install jupyter==1.0.0

Retrieve the example Jupyter notebooks using the PYNQ get-notebooks command. An example of how to run the Jupyter notebook server is also shown below:

pynq get-notebooks --from-package finn-examples -p . --force
jupyter-notebook --no-browser --port=8888

You can now navigate the provided Jupyter notebook examples, or just use the provided accelerators as part of your own Python program:

from finn_examples import models
import numpy as np

# instantiate the accelerator
accel = models.cnv_w2a2_cifar10()
# generate an empty numpy array to use as input
dummy_in = np.empty(accel.ishape_normal(), dtype=np.uint8)
# perform inference and get output
dummy_out = accel.execute(dummy_in)

Example Neural Network Accelerators

Dataset Topology Quantization Supported boards Supported build flows

CIFAR-10
CNV (VGG-11-like) several variants:
1/2-bit weights/activations
Pynq-Z1
ZCU104
Ultra96
U250
Pynq-Z1
ZCU104
Ultra96
U250


MNIST
3-layer fully-connected several variants:
1/2-bit weights/activations
Pynq-Z1
ZCU104
Ultra96
U250
Pynq-Z1
ZCU104
Ultra96
U250


ImageNet
MobileNet-v1 4-bit weights & activations
8-bit first layer weights
Alveo U250 Alveo U250


ImageNet
ResNet-50 1-bit weights 2-bit activations
4-bit residuals
8-bit first/last layer weights
Alveo U250 -


RadioML 2018
1D CNN (VGG10) 4-bit weights & activations ZCU104 ZCU104


MaskedFace-Net
BinaryCoP
Contributed by TU Munich+BMW
1-bit weights & activations Pynq-Z1 Pynq-Z1


Google Speech Commands v2
3-layer fully-connected 3-bit weights & activations Pynq-Z1 Pynq-Z1


UNSW-NB15
4-layer fully-connected 2-bit weights & activations Pynq-Z1
ZCU104
Ultra96
Pynq-Z1
ZCU104
Ultra96

Please note that the build flow for ResNet-50 for the Alveo U250 has known issues and we're currently working on resolving them. However, you can still execute the associated notebook, as we provide a pre-built FPGA bitfile generated with an older Vivado (/FINN) version targeting the xilinx_u250_xdma_201830_2 platform.
Furthermore, please note that you can target other boards (such as the Pynq-Z2 or ZCU102) by changing the build script manually, but these accelerators have not been tested.

We welcome community contributions to add more examples to this repo!

Supported Boards

Note that the larger NNs are only available on Alveo or selected Zynq boards.

finn-examples provides pre-built FPGA bitfiles for the following boards:

  • Edge: Pynq-Z1, Ultra96 and ZCU104
  • Datacenter: Alveo U250

It's possible to generate Vivado IP for the provided examples to target any modern Xilinx FPGA of sufficient size. In this case you'll have to manually integrate the generated IP into your design using Vivado IPI. You can read more about this here.

Rebuilding the bitfiles

All of the examples here are built using the FINN compiler, and can be re-built or customized. See the build/README.md for more details.

About

Dataflow QNN inference accelerator examples on FPGAs

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages

  • Jupyter Notebook 74.7%
  • Python 24.0%
  • Shell 1.3%