Tiny MLPerf™ v0.1 hls4ml Xilinx PYNQ-Z2 Open Submission

By hls4ml team

• Contacts:
  • Ben Hawks, email: [email protected], GitHub: @ben-hawks
  • Nhan Tran, email: [email protected], GitHub: @nhanvtran
  • Javier Duarte, email: [email protected], GitHub: @jmduarte
  • Giuseppe DiGuglielmo, email: [email protected], GitHub: @GiuseppeDiGuglielmo
• Team members:
  • Nicolò Ghielmetti, CERN
  • Jules Muhizi, Fermilab/Harvard
  • Shvetank Prakash, Columbia/Harvard
  • Rushil Roy, UCSD

Hardware

• Board is a TUL PYNQ-Z2 based on Xilinx Zynq SoC (See https://www.tul.com.tw/productspynq-z2.html for more information).

Code structure

The code is structured as follows

hls4ml
├── code
│   ├── ad
│   │   └── AD03
│   │       ├── inference
│   │       │   ├── hls
│   │       │   ├── sdk
│   │       │   ├── sys
│   │       │   └── utils
│   │       └── training
│   │           ├── convert.py
│   │           ├── model
│   │           │   └── ad03
│   │           │       └── model_ToyCar.h5
│   │           ├── train.py
│   └── ic
│       └── RN06
│           ├── inference
│           │   ├── hls
│           │   ├── sdk
│           │   ├── sys
│           │   └── utils
│           └── training
│               ├── convert.py
│               ├── resnet_v1_eembc_RN06
│               │   └── model_best.h5
│               └── train.py
├── results
│   └── pynqz2
│       ├── ad
│       │   ├── accuracy
│       │   └── performance
│       └── ic
│           ├── accuracy
│           └── performance
└── systems

• For both the anomaly detection model (AD03) and the image classification model (RN06), there are training and inference subdirectories.
• Under training, there are scripts to train the model with QKeras (train.py) and convert it to a Xilinx HLS/Vivado/SDK project using hls4ml (convert.py).
• The configruation is controlled by yml files.
• For convenience, the pretrained models in .h5 format are provided in the repository as indicated.
• Under inference, the Xilinx HLS, Vivado, and SDK projects will be automatically created after successfully running convert.py in the hls, sys, and sdk folders respectively.

Setup

• Install miniconda from here: https://docs.conda.io/en/latest/miniconda.html
• Create the environment:

conda-env create -f environment.yml

• Activate the environment:

conda activate tiny-mlperf-env

• Install Vivado 2019.1 from https://www.xilinx.com/support/download/index.html/content/xilinx/en/downloadNav/vivado-design-tools/archive.html
• Download PYNQ-Z2 board files (from https://dpoauwgwqsy2x.cloudfront.net/Download/pynq-z2.zip) and install appropriately by extracting and copying files to:

<path_to_Vivado>/Vivado/2019.1/data/boards/board_files

• Setup Vivado 2019.1:

source <path_to_Vivado>/Vivado/2019.1/settings64.sh

• Ensure PYNQ-Z2 board is connected (and powered) by USB and visible.

Training with QKeras

In this step, you will download the dataset and perform a quantization-aware training with QKeras.

AD03 model

• Change directory

cd code/ad/AD03/training/

• Download dataset for AD03:

./get_dataset.sh

• Train AD03, pretrained model is provided as model/ad03/model_ToyCar.h5:

python train.py -c AD03.yml

n.b. if you don't have a GPU, you can comment out the import setGPU (true also for later python scripts)

RN06 model

• Change directory

cd code/ic/RN06/training/

• Train RN06, pretrained model is provided as resnet_v1_eembc_RN06/model_best.h5:

python train.py -c RN06_pynqz2.yml

Conversion with hls4ml

In this step, you will ingest the quantization-aware training performed in the previous step and convert it to firmware using hls4ml. The hls4ml configuration, pynqz2.yml has details such as the implementation architecture.

AD03 model

• Change directory

cd code/ad/AD03/training/

• Get test data:

python generate_test_data.py -c AD03.yml

• Convert AD03:

python convert.py -c pynqz2.yml

RN06 model

• Change directory

cd code/ic/RN06/training/

• Get test data:

source get_test_data.sh

• Convert RN06:

python convert.py -c RN06_pynqz2.yml

Program FPGA and run software

• Change directory

cd code/ic/<model_name>/inference/sdk/

• Open Xilinx SDK GUI

make gui

• Program the FPGA with the bit file in SDK
  • 
• Run test harness software in SDK
  • 
• Download EEMBC runner GUI and AD/IC benchmark datasets (See https://github.com/eembc/ulpmark-ml)
• Open EEMBC runner GUI and and perform measurements, follow the instructions on the eembc README
  • 

Boot from Flash

The PYNQ--Z2 supports Quad SPI Flash. Please follow these instructions to program and boot from the Flash memory.