A tool for automated mapping and optimization of 3D-CNNs for HAR onto FPGA devices
In order to clone the repository, you can use the following command:
git clone --recurse-submodules https://github.com/ptoupas/harflow3d.git
To use this tool, you will need to meet the following requirements:
- Python 3.9 or higher
- Install python packages listed in
requirements.txt
- Install fpgaconvnet-optimiser and fpconvnet-model packages as described in the relevant section below
To install the necessary Python packages, you can use pip
. Navigate to the root directory of the repository and execute the following command:
cd harflow3d
pip install -r requirements.txt
This will install all the required Python packages specified in the requirements.txt
file.
The harflow3d
tool uses the fpgaconvnet-optimiser
and fpgaconvnet-model
packages to map and optimize 3D-CNNs for HAR onto FPGA devices. These packages are included as submodules in the harflow3d
repository. To install them, you should first initialize the submodules (in case you haven't cloned the repository with the --recurse-submodules
flag):
git submodule update --init --recursive
Then, (from the root directory) navigate to the fpgaconvnet-optimiser
directory and install the package:
cd fpgaconvnet-optimiser
pip install .
Finally, (from the root directory) navigate to the fpgaconvnet-model
directory and install the package:
cd fpgaconvnet-model
pip install .
Once you have installed the necessary dependencies or set up Docker, you can use the harflow3d
tool to map and optimize 3D-CNNs for HAR onto FPGA devices.
To use the harflow3d
tool, you first need to download the necessary ONNX model files. We have provided a script named get_onnx_models.sh
to help you with this. The script will download the following models:
- x3d_m
- c3d
- slowonly
- r2plus1d_18
- r2plus1d_34
The models will be stored in the correct folder of the fpgaconvnet-optimiser
submodule, which is located in the fpgaconvnet-optimiser/examples/models
directory.
To run the script, simply navigate to the root directory of the harflow3d
repository and execute the following command:
./get_onnx_models.sh
The harflow3d
tool uses wandb to store the results of the optimization process. For the best experience and visualization of the results, we recommend that you use wandb. It will provide you with a dashboard with very detailed information about the optimization process and the interpretation of the results.
To use wandb, you need to create an account and initialize it. After you have created an account, you can initialize wandb by running the following command:
wandb login or wandb init
This will promt you to enter your wandb API key. You can find your API key by clicking on your profile picture in the top right corner of the wandb dashboard and then clicking on the "Settings" button. You can find more information about wandb initialization here. Note that this step only needs to be done once.
The fpgaconvnet-optimiser
package provides a configuration file that allows you to specify the mapping and optimization parameters. The configuration file is located in the fpgaconvnet-optimiser/examples/
directory. The default configuration file is named latency_optimiser_example.toml
. You can edit this file to change the mapping and optimization parameters.
To run the optimiser you should navigate to the fpconvnet-optimiser directory and run the following command:
cd fpgaconvnet-optimiser
./run_optimiser.sh
The run_optimiser.sh
script will run the optimiser with the default configuration file (the one described in step 2). The script accepts the following arguments:
-p
. This argument allows you to specify the target FPGA platform to use. The default value iszcu104
. You can define more than one device and the optimiser will generate one design for each FPGA platform defined. You can do that as follows-p "zcu104 zcu102"
. The supported FPGA platforms are the following:vus440
vc707
vc709
zc706
zcu102
zcu104
-m
. This argument allows you to specify the model to use. The default value isc3d
. You can define more than one model and the optimiser will generate one design for each model defined. You can do that as follows-m "x3d_m c3d"
. The supported models are the following:x3d_m
c3d
slowonly
r2plus1d_18
r2plus1d_34
-n
. This argument allows you to specify the number of times each pair of FPGA-Model will be executed. The default value is1
. For each pair of FPGA-Model, the optimiser will generate a design and run itn
times.
The results of the optimization process will be stored in a wandb project named after the name of the tool (harflow3d) followed by the name of the model with a postfix "latency". For example, if you run the optimiser for c3d model, the results will be stored in a wandb project named harflow3d-c3d-latency
. The wandb project can be accessed by clicking here. You can find your run by using the filters and tools provided by wandb. For example, you can filter the runs by model name, platform name, etc. You can also use the wandb tools to compare the results of different runs.
After the final configuration of the 3D CNN model is produced by the optimizer, you can request the bitstream and host code necessary to execute the model on a specific FPGA device. To obtain these files, you will need to provide us with the configuration file which can be found under the fpgaconvnet-optimizer/outputs/{model_name}/{platform_name}/config.json
or the link of the specific run from the wandb
project page.
Upon request we will generate the bitstream and host code for you using our closed-source backend tool that translates the given configuration of the layers into hardware IPs. We integrate the layers IPs into our proposed Vivado design and provide the final bitstream. Please note that this service is provided on a best-effort basis, and the timeframe for generating the bitstream and host code may vary depending on the complexity of the model and the current workload.
If you are interested in obtaining the bitstream and host code for your model, please contact us at [[email protected] or [email protected]].
Alternatively, you can use Docker to run this tool. Docker allows you to run the tool in a self-contained environment without worrying about dependencies or system configuration.
To use this tool with Docker, follow these steps:
- Install Docker on your system (if you haven't already).
- Clone the repository and navigate to the root folder
harflow3d
. - Download the ONNX model files (see step 1 above).
- Build the Docker image:
docker build -t harflow3d -f docker/Dockerfile .
This will build a Docker image named harflow3d
using the Dockerfile
in the docker
directory.
- Run the Docker image:
To run the Docker image, you can use the following command:
docker run -it --rm harflow3d /bin/bash
This will start a container and give you access to a shell inside the container. From there, you can run the harflow3d
tool as usual (see step 4 above). Remember that you will need to initialize wandb inside the container with your API key as explained above in step 2.
If you find this project useful in your research, please consider cite:
@article{toupas2023harflow3d,
title={HARFLOW3D: A Latency-Oriented 3D-CNN Accelerator Toolflow for HAR on FPGA Devices},
author={Toupas, Petros and Montgomerie-Corcoran, Alexander and Bouganis, Christos-Savvas and Tzovaras, Dimitrios},
doi={10.48550/arXiv.2303.17218},
journal={arXiv preprint arXiv:2303.17218},
year={2023}
}