第二部分 Build Machine Learning Models for DPU

Build Machine Learning Models for DPU

This folder helps users recompile their own DPU models so they can be deployed
on the board. The recompilation is needed if users want to retarget
a different DPU configuration.

We provide a compile.sh script that helps users compile their own deployable
models from Vitis AI model zoo. For advanced users, a jupyter notebook
train_mnist_model.ipynb is provided to show how to train a ML
model, quantize it, and compile it using Vitis AI tools.

Prerequisites

1. Repository

Make sure you have cloned this repository onto your host machine:

git clone --recursive --shallow-submodules https://github.com/Xilinx/DPU-PYNQ.git

2. Docker

If you have not installed docker on your host machine, please refer to the
Vitis AI getting started page
to install docker.

3. (optional) dpu.hwh

There are 3 cases for different boards:

1. For ZCU104, our DPU configuration is consistent with the ZCU104 design of
   the Vitis AI release.
2. For ZCU111, our DPU configuration is consistent with the ZCU102 design of
   the Vitis AI release.
3. For Ultra96 or other boards, the Vitis AI release does not contain
   necessary files to recompile the DPU models directly.

For case 1 and 2, users can leverage existing files on the released docker
image. No action is required.

For case 3 (Ultra96 or any other board), the DPU configuration file
(Ultra96.dcf or <Board>.dcf) does not exist on the released docker image.
We need to prepare the dpu.hwh so we can compile it into a DPU configuration
file.

One way to get the dpu.hwh is to
download it for Ultra96;
remember to rename it to dpu.hwh if necessary.

Alternatively, if you have rebuilt the DPU hardware design by yourself,
you should see 3 overlay files (dpu.hwh, dpu.bit, and dpu.xclbin)
inside folder DPU-PYNQ/Boards/<Board>. You can take the dpu.hwh
there as well.

Note: if you have changed the DPU configurations in your hardware design,
you must prepare your new dpu.hwh.

4. (optional) Docker

If you want to run the jupyter notebook train_mnist_model.ipynb on the host
machine, you will need to make sure you have an AVX2 and FMA compatible
machine. To check that:

grep avx2 /proc/cpuinfo
grep fma /proc/cpuinfo

Both commands should return a list of available supported instruction sets.
If one of the 2 commands returns nothing, your machine will have difficulties
importing tensorflow package.

Also, on the docker image, you need to do the following before running any Jupyter
notebook.

conda activate vitis-ai-tensorflow
yes | pip install matplotlib keras==2.2.5

Build DPU Models from Vitis AI Model Zoo

On your host machine, as mentioned in the previous section, if you are
building models for Ultra96, you need to put the corresponding dpu.hwh file
in folder DPU-PYNQ/host.

We can run the following commands now.

cd DPU-PYNQ/host
mkdir -p docker
cp -rf ../vitis-ai-git/docker_run.sh .
cp -rf ../vitis-ai-git/docker/PROMPT.txt docker
chmod u+x docker_run.sh
./docker_run.sh xilinx/vitis-ai-cpu:latest

The docker_run.sh will download a Vitis AI docker image after users accept
the license agreements. It may take a long time to download since the image
is about 8GB. After the download is complete, the docker_run.sh script
will help you log onto that docker image.

The Vitis AI docker image gives users access to the Vitis AI utilities
and compilation tools. If you have run docker_run.sh before, it will simply
launch the docker image without downloading again.

Once you are in the docker environment, you can run the compile.sh script.

./compile.sh <Board> <model_name>

Here Board can be Ultra96, ZCU104, and ZCU111. For model_name,
users can check the model information page as shown below.

[外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-PGpMfCrA-1593136974001)(images/model_info.png)]

The compile.sh eases the compilation of existing DPU models. Users can also
adjust the script to compile their own models. The compile.sh does the
following things.

(1) Adding Ultra96 support

We prepare the Ultra96.dcf using the dpu.hwh file.

(2) Downloading a model from model zoo

We download a deployable model as a zip file from
Vitis AI Model Zoo.
The contents of the extracted zip file (e.g., cf_resnet50_imagenet_224_224_7.7G)
will contain multiple versions of the model:

• floating point frozen graph (under float)
• quantized evaluation model (under quantized)
• quantized deployment model (under quantized)

In our case we only need the following files inside the quantized directory:
(1) deploy.caffemodel and (2) deploy.prototxt.

(3) Compiling the model

We will compile the model into a *.elf file.
If everything is successful, you should see a screen as shown below.

[外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-oQP25k4e-1593136974002)(images/vai_c_output_caffe.png)]

A new model file (e.g. dpu_resnet50_0.elf) should appear in your working
directory; this is the model file that can be deployed on the board.

After you are done with the docker environment, type in:

exit

to exit.

Train Your Own DPU Models from Scratch

Instead of using the deployable models from the Vitis AI model zoo, advanced
users may even train their own machine learning models. We will show
one example in train_mnist_model.ipynb.

Once you are in the docker environment, if you have not done the following,
make sure you do it before running the notebook:

conda activate vitis-ai-tensorflow
yes | pip install matplotlib keras==2.2.5

Then launch jupyter notebook and run the train_mnist_model.ipynb
step-by-step.

jupyter notebook --ip=0.0.0.0 --port=8080

For more information such as training your own model, please refer to the
Vitis AI user guide
and Vitis AI Tutorials.

References

• Vitis AI Github
• Vitis AI User Guide
• Vitis AI Model Zoo

Copyright © 2020 Xilinx, Inc