This repository provides a means to profile (i.e., benchmark) the evaluation methodologies given by some genetic programming (GP) tools. The evolutionary mechanisms provided by the GP tools are not included when profiling——only mechanisms for calculating "fitness."
This repository was created for the EuroGP 2023 conference paper "Using FPGA Devices to Accelerate Tree-Based Genetic Programming: A Preliminary Exploration with Recent Technologies," by Crary et al., which compared the evaluation performance of an initial FPGA-based GP hardware accelerator with that of the GP software tools DEAP (version 1.3), TensorGP (Git revision 09e6d04), and Operon (Git revision 9e7ee4e).
A means for profiling for the following GP tools is given:
- DEAP - for the original paper, click here.
- TensorGP - for the original paper, click here.
- Operon - for the original paper, click here.
Source code for the FPGA accelerator is not provided at this time, although the architecture is described at length in the aforementioned paper, "Using FPGA Devices to Accelerate Tree-Based Genetic Programming: A Preliminary Exploration with Recent Technologies."
By default, the repository already contains the results
published in the relevant conference paper.
These results are contained in the experiment/results directory and
can be viewed with the experiment/tools/stats.ipynb Jupyter Notebook file.
If so desired, after successfully completing installation (as described below), you may run the entire profiling suite by executing the following within a shell program, after having navigated to the repository directory within the shell:
cd experiment
bash run.sh
After the run.sh script fully executes, to view some relevant statistics, run the Jupyter Notebook file given by the path experiment/tools/stats.ipynb.
The following has been verified via CentOS 7. It is likely that other Linux distributions are supported, but it is unlikely that Windows and MacOS operating systems are readily supported.
- Ensure that some Conda package management system (e.g., Miniconda) is installed on the relevant machine.
- Download the latest software release from GitHub, available here. Ignore the
data.tar.gzfile for now.
Upon extracting the source code, set up the relevant Conda environment and tools by executing the following within a shell program, after having navigated to the repository directory within the shell:
conda env create -f environment.yml
conda activate conference-eurogp-2023
pip install -r requirements.txt
bash install.sh
To finish installation, extract and copy the contents of the data.tar.gz file from the software release (i.e., the one folder and three .pkl files) and paste them within the experiment/results folder. These contents provide the random programs, inputs, and outputs utilized by the experiments.
NOTE: After copying the contents of the data.tar.gz file to the experiment/results folder, you may need to change file permissions for the relevant .pkl files. One way of doing so is by executing the following:
chmod 755 experiment/results/*.pkl
NOTE: If using a CPU from the Intel Skylake series (like in the conference paper), then you may need to specify this particular CPU architecture in the compilation settings for Operon before running the bash install.sh command listed above. To do so, comment out line 501 in experiment/tools/operon/custom/CMakeLists.txt and uncomment line 502.
NOTE: If using an Nvidia GPU (like in the conference paper), you may need to ensure that tensorflow can successfully utilize a GPU within the conda environment by prepending the following CUDA paths (or something similar) to the $LD_LIBRARY_PATH environment variable:
export LD_LIBRARY_PATH=$CONDA_PREFIX_1/pkgs/cudatoolkit-11.2.2-hbe64b41_10/lib:$CONDA_PREFIX_1/envs/tensorgp-test/lib:$LD_LIBRARY_PATH
If the above export command is executed, you will likely need to restart your shell to reset the $LD_LIBRARY_PATH environment variable after running any experiments.