qixingzhang / SummerSchool2022-Vitis-AI

Vitis AI Lab: MNIST classifier

Vitis AI Lab: MNIST Classifier

Install Vitis AI

1. Install docker engine
   • Official site: https://docs.docker.com/engine/install/
   • For Ubuntu: https://docs.docker.com/engine/install/ubuntu/
2. Pull Vitis AI 1.4 image

   sudo docker pull xilinx/vitis-ai:1.4.916
   

3. Clone from GitHub

   $ git clone https://github.com/Xilinx/Vitis-AI.git

   $ cd Vitis-AI

   $ git checkout v1.4

4. Clone this repository to Vitis-AI folder

   $ git clone https://github.com/qixingzhang/SummerSchool2022-Vitis-AI.git

5. Launch Vitis AI

   sudo ./docker_run.sh xilinx/vitis-ai:1.4.916

Vitis AI Workflow

1. Activate TensorFlow 2.x environment

   $ conda activate vitis-ai-tensorflow2

2. Train the model if needed

   A pre-trained keras model (.h5 format) is provided, you can also train it by youself

   $ python train.py

3. Quantize the model

   $ chmod +x *.sh

   $ ./1_quantize.sh

   This script invokes the python script vitis_ai_tf2_quantize.py. It uses the python API of Vitis AI Quantizer:

   • quantizer = vitis_quantize.VitisQuantizer(model)
     • model: keras floating point model
   • quantized_model = quantizer.quantize_model(calib_dataset=dataset)
     • calib_dataset: dataset for calibration, 100 ~ 1000 training or testing images is enough.

4. Compile the model

   $ ./2_compile.sh

   This scripts uses the vai_c_tensorflow2 commmand to compile the quantized model. It has four required parameters:

   • --model quantized model with h5 format
   • --arch indicate DPU arch, it can be found in /opt/vitis_ai/compiler/arch/DPUCZDX8G for different Zynq boards.
   • --output_dir compile output path
   • --net_name your model name

   The output .xmodel file is saved in compile_output folder.

Deploy on edge board using DPU-PYNQ

1. Boot the board with PYNQ v2.7 image
   • Image download link: http://www.pynq.io/board.html
2. Install python package

   $ sudo pip3 install pynq-dpu --no-build-isolation

3. Get notebooks

   $ cd $PYNQ_JUPYTER_NOTEBOOKS

   $ pynq get-notebooks -p . pynq-dpu

4. Run the dpu_mnist_classifier notebook in pynq-dpu folder. You can upload your own .xmodel file to this folder and replace the existing dpu_mnist_classifier.xmodel