PhySO

Web: https://physo.readthedocs.io/en/latest/
GitHub: https://github.com/WassimTenachi/PhySO

ceuadmin/PhySO/1.1.11

module load ceuadmin/python/3.12.10
python -m venv PhySO
source PhySO/bin/activate
pip index versions physo
pip install physo
# to run .pynb
pip install ipykernel
# noninteractively:
# jupyter nbconvert --to notebook --execute --inplace --ExecutePreprocessor.timeout=7200 physo.ipynb
pip install nbconvert
# to run .py, e.g., paste into a Python session
pip install psutil

showing that

physo (1.1.11)
Available versions: 1.1.11, 1.1.10, 1.1.9, 1.1.8, 1.0

For the full test shown earlier, revise code to access fitted free constants via prog.free_consts.class_values, replacing the obsolete prog.free_const_values attribute. The Pareto-front output can also be made robust to expressions that raise ZeroDivisionError during SymPy simplification, falling back to unsimplified expression output so that processing can continue. In more details,

# Changes:
# - PhySO 1.1.11 uses free_consts.class_values instead of free_const_values.
# - Handle ZeroDivisionError when simplifying some Pareto-front expressions.

free_consts = expression.free_consts.class_values.detach().cpu().numpy().flatten()
for j in range(len(free_consts)):
    print("%s = %f" % (expression.library.free_const_names[j], free_consts[j]))

pareto_front_complexities, pareto_front_expressions, pareto_front_r, pareto_front_rmse = logs.get_pareto_front()

for i, prog in enumerate(pareto_front_expressions):

    # Showing expression
    try:
        print(prog.get_infix_pretty(do_simplify=True))
    except ZeroDivisionError:
        print(prog.get_infix_pretty(do_simplify=False))

    # Showing free constants
    free_consts = prog.free_consts.class_values.detach().cpu().numpy().flatten()
    for j in range(len(free_consts)):
        print("%s = %f" % (prog.library.free_const_names[j], free_consts[j]))

    # Showing RMSE
    print("RMSE = {:e}".format(pareto_front_rmse[i]))
    print("-------------")

ceuadmin/PhySO/1.0-dev0

It is recommended that the following commands are necessary,

module load ceuadmin/PhySO
. /usr/local/Cluster-Apps/miniconda3/4.5.1/etc/profile.d/conda.sh
conda activate ${PhySO}

Line containing conda.sh could also be in ~/.bashrc, and conda deactivate will unload the module.

However, it appears simpler with

module load ceuadmin/PhySO
source activate $PhySO

with source deactivate. In both cases, PhySO is already defined by the ceuadmin/PhySO module.

Installation

The download is as usual.

git clone https://github.com/WassimTenachi/PhySO
cd PhySO

It takes more than usual resources, which is necessary to get around with SLURM.

cat << 'EOL' > ~/PhySO.sb
#!/usr/bin/bash

#SBATCH --account CARDIO-SL0-CPU
#SBATCH --partition cardio
#SBATCH --qos=cardio
#SBATCH --mem=28800
#SBATCH --time=12:00:00
#SBATCH --job-name=PhySO
#SBATCH --output=PhySO.o
#SBATCH --error=PhySO.e

export CEUADMIN=/usr/local/Cluster-Apps/ceuadmin

. /etc/profile.d/modules.sh
module load miniconda3/4.5.1 texlive
. /usr/local/Cluster-Apps/miniconda3/4.5.1/etc/profile.d/conda.sh
conda activate $CEUADMIN/PhySO
conda install -p $CEUADMIN/PhySO --file requirements.txt
conda install -p $CEUADMIN/PhySO --file requirements_display1.txt
pip install -r requirements_display2.txt
pip install -e .
EOL

function create()
{
 # This step is doable from a login node
 conda create -y -p $CEUADMIN/PhySO python=3.8
 # The following statements are all automatic recommendations
 ## conda update -n base -c defaults conda
 echo ". /usr/local/Cluster-Apps/miniconda3/4.5.1/etc/profile.d/conda.sh" >> ~/.bashrc
 ## All users
 sudo ln -s /usr/local/Cluster-Apps/miniconda3/4.5.1/etc/profile.d/conda.sh /etc/profile.d/conda.sh
 echo "conda activate" >> ~/.bashrc
}

where we have made it available from /usr/local/Cluster-Apps/ceuadmin with all the steps to be generated into ~/PhySO.sb, called with sbatch.

Note that requirements.txt includes pytorch for conda install but torch with pip install -r requirements.txt. It also appears that dot2tex, pdflatex (needs to be pip install pdflatex) and pdf2image are necessary. Some notes are available from requirements_display2.txt.

:star: :star: :star: It appears that there is conflict between texlive/2015 from CSD3 and texlive-core so we execute conda uninstall -p $PhySO texlive-core to be in line with the former.

Tests of package loading and units follow suit from the documentation.

python -c 'import physo;print("OK")'
python -m unittest discover -p "*UnitTest.py"

Note that at the Getting started section Python session needs to be started followed by a statement import numpy as np to proceed.

By-products

Besides torch, it is noticeable that jupyterlab and scikit-learn are also made available as dependencies. In fact, quite some packages are installed to lib/python3.8/site-packages/, leading to an overall size of ~7GB.

A full list is visible through conda list.

A full test

Demos are available from demo/ in forms of both .ipynb and .py.

This is extracted from README.md,

import numpy as np
import physo

z = np.random.uniform(-10, 10, 50)
v = np.random.uniform(-10, 10, 50)
X = np.stack((z, v), axis=0)
y = 1.234*9.807*z + 1.234*v**2

# set 1. Symbolic regression with default hyperparameters.
expression, logs = physo.SR(X, y,
                            X_units = [ [1, 0, 0] , [1, -1, 0] ],
                            y_units = [2, -2, 1],
                            fixed_consts       = [ 1.      ],
                            fixed_consts_units = [ [0,0,0] ],
                            free_consts_units  = [ [0, 0, 1] , [1, -2, 0] ],
)
# set 2. Hyperparameters configurations.
expression, logs = physo.SR(X, y,
                            X_units = [ [1, 0, 0] , [1, -1, 0] ],
                            y_units = [2, -2, 1],
                            fixed_consts       = [ 1.      ],
                            fixed_consts_units = [ [0,0,0] ],
                            free_consts_units  = [ [0, 0, 1] , [1, -2, 0] ],
                            run_config = physo.config.config1.config1
)
# set 3. Selectable symbolic operations.
expression, logs = physo.SR(X, y,
                            X_names = [ "z"       , "v"        ],
                            X_units = [ [1, 0, 0] , [1, -1, 0] ],
                            y_name  = "E",
                            y_units = [2, -2, 1],
                            fixed_consts       = [ 1.      ],
                            fixed_consts_units = [ [0,0,0] ],
                            free_consts_names = [ "m"       , "g"        ],
                            free_consts_units = [ [0, 0, 1] , [1, -2, 0] ],
                            op_names = ["mul", "add", "sub", "div", "inv", "n2", "sqrt", "neg", "exp", "log", "sin", "cos"]
)

print(expression.get_infix_pretty(do_simplify=True))
print(expression.get_infix_latex(do_simplify=True))
print(expression.free_const_values.cpu().detach().numpy())

pareto_front_complexities, pareto_front_expressions, pareto_front_r, pareto_front_rmse = logs.get_pareto_front()
for i, prog in enumerate(pareto_front_expressions):
    # Showing expression
    print(prog.get_infix_pretty(do_simplify=True))
    # Showing free constant
    free_consts = prog.free_const_values.detach().cpu().numpy()
    for j in range (len(free_consts)):
        print("%s = %f"%(prog.library.free_const_names[j], free_consts[j]))
    # Showing RMSE
    print("RMSE = {:e}".format(pareto_front_rmse[i]))
    print("-------------")

Three sets of parameters are provided which no longer requires to be run separately. These are also wrapped as a physo.ipynb including outputs.

fonts

This section is according to https://alexanderlabwhoi.github.io/post/2021-03-missingfont/.

rm ~/.cache/matplotlib -rf
conda install -p $PhySO -c conda-forge mscorefonts

and use them,

import matplotlib
matplotlib.rcParams['font.family'] = "sans-serif"
matplotlib.rcParams['font.sans-serif'] = "Comic Sans MS"

This will reveal location of the default specification,

import matplotlib
print(matplotlib.matplotlib_fname())

which in this case, /usr/local/Cluster-Apps/ceuadmin/PhySO/lib/python3.8/site-packages/matplotlib/mpl-data/matplotlibrc, where we make changes such as,

font.family:  sans-serif
font.sans-serif: Arial
pdf.fonttype: 42