Benchmark and baselines for offline multi-objective optimization.
Our proposed offline collected data can be accessed and downloaded via Google Drive.
In order to run our Regex, RFP, and ZINC tasks, following LaMBO (Paper, Code), you may first download FoldX Emulator.
FoldX is available under a free academic license. After creating an account you will be emailed a link to download the FoldX executable and supporting assets. Copy the contents of the downloaded archive to ~/foldx. You may also need to rename the FoldX executable (e.g. mv -v ~/foldx/foldx_20221231 ~/foldx/foldx).
After installing FoldX, generate an instance proxy_rfp_problem.pkl of RFP task by running
cd off_moo_bench/problem/lambo/
python scripts/black_box_opt.py optimizer=mf_genetic optimizer/algorithm=nsga2 task=proxy_rfp tokenizer=proteinMake sure that the lines of saving instance of proxy_rfp_problem.pkl exist in line 203 of off_moo_bench/problem/lambo/lambo/optimizers/pymoo.py such that
if round_idx == self.num_rounds:
import pickle
with open('proxy_rfp_problem.pkl', 'wb+') as f:
pickle.dump(problem, f)After successfully installing FoldX, run
bash install.shfor a quick installation.
To test with our offline MO-NAS problems, installing EvoXBench (paper, code) is needed. Before running pip install evoxbench, you should first download their database via Google Drive or Baidu Netdisk, and data via Google Drive or Baidu Netdisk, and save it to the path off_moo_bench/problem/mo_nas/database and off_moo_bench/problem/mo_nas/data.
Following How to download from Google Drive, we propose a more stable method for downloading in a commandline interface. Please first go to OAuth 2.0 Playground to obtain your Google Drive download APIs, then download it with
curl -H "Authorization: Bearer <Your Google Drive APIs>" https://drive.google.com/file/d/11bQ1paHEWHDnnTPtxs2OyVY_Re-38DiO/view?usp=sharing -o database.zip curl -H "Authorization: Bearer <Your Google Drive APIs>" https://drive.google.com/file/d/1r0iSCq1gLFs5xnmp1MDiqcqxNcY5q6Hp/view?usp=sharing -o data.zipThen configure the EvoXBench as:
from evoxbench.database.init import config
config("Path to database", "Path to data")
# For instance:
# With this structure:
# /home/Downloads/
# └─ database/
# | | __init__.py
# | | db.sqlite3
# | | ...
# |
# └─ data/
# └─ darts/
# └─ mnv3/
# └─ ...
# Then, execute:
# config("PATH_TO_M2BO/off_moo_bench/problem/mo_nas/database", "PATH_TO_M2BO/off_moo_bench/problem/mo_nas/data")
### MuJoCo
1. We use [MuJoCo](https://github.com/google-deepmind/mujoco) with version of ``2.1.0``, which can be downloaded on [GitHub](https://github.com/google-deepmind/mujoco/releases/download/2.1.0/mujoco210-linux-x86_64.tar.gz), then run
```shell
mkdir ~/.mujoco
tar -zxvf mujoco210_linux_x86_64.tar.gz -C ~/.mujocoto put MuJoCo under ~/.mujoco and set environment variables.
Make sure that you have installed needed dependancy.
sudo apt install libosmesa6-dev libgl1-mesa-glx libglfw3Install required packages cython, mujoco_py, gym and set environmental variables.
pip install cython==3.0.0a10 mujoco-py==2.1.2.14 gym==0.14.0
conda env config vars set LD_LIBRARY_PATH=$LD_LIBRARY_PATH:~/.mujoco/mujoco210/bin:/usr/lib/nvidiaAfter the above steps, you can check MuJoCo installation by
import mujoco_py
import os
mj_path = mujoco_py.utils.discover_mujoco()
xml_path = os.path.join(mj_path, 'model', 'humanoid.xml')
model = mujoco_py.load_model_from_path(xml_path)
sim = mujoco_py.MjSim(model)print(sim.data.qpos)
sim.step() print(sim.data.qpos)
### Notice during installation
1. Due to conflict versions among different packages, we design special orders to install packages correctly. Note that conducting ``conda env create -f environment.yml`` may raise some errors during installing Pip dependancy. Do not stop then and keep going on with rest scripts we proposed in ``install.sh``.
2. Due to packages conflicts, we use ``fix_contents.sh`` to solve such conflict bugs by running
```shell
bash fix_contents.sh ${YOUR_PATH_TO_CONDA}/envs/off-moo/lib/python3.8/site-packages/sklearn/cross_decomposition/pls_.py "pinv2" "pinv"tests/ folder to check for successful installation.libstdc++.so.6: version `GLIBCXX_3.4.29' not found issue, since the reasons can be various, we recommend that you refer to StackOverflow: Where can I find GLIBCXX_3.4.29? or GitHub: libstdc++.so.6: version `GLIBCXX_3.4.29' not found for further advice to solve this issue.To reproduce the performance of baseline algorithms reported in our work, you may then run run_multi_head.sh/run.sh/run_single_obj.sh/run_mobo_once.sh, or run the following series of commands in a bash terminal. Also, please ensure that the conda environment off-moo is activated in the bash session.
for env_name in "zdt3" "re21"; do
for seed in 2024; do
for train_data_mode in "onlybest_1"; do
for train_mode in "grad_norm" "pcgrad"; do
for solver in "nsga2"; do
for output_size in 256; do
python config_evoxbench.py
CUDA_VISIBLE_DEVICES=0 \
python scripts/multi_head_nn.py \
--env-name ${env_name} \
--seed ${seed} \
--normalize-x \
--normalize-y \
--num-solutions ${output_size} \
--filter-type best \
--train-data-mode ${train_data_mode} \
--train-mode ${train_mode} \
--mo-solver ${solver} \
--reweight-mode none \
--df-name $seed-test-hv.csv
done
done
done
done
done
doneGradNorm, COMs, etc), data pruning is needed, which corresponds to onlybest_1 in --train-data-mode.model and trainer under algorithm folder, then choose one MO-solver as what is done in scripts/multi_obj_nn.py.If you have any questions, feel free to contact Rongxi Tan or raise an issue.