PDFFlow is parton distribution function interpolation library written in Python and based on the TensorFlow framework. It is developed with a focus on speed and efficiency, enabling researchers to perform very expensive calculation as quick and easy as possible.
The key features of PDFFlow is the possibility to query PDF sets on GPU accelerators.
The documentation for PDFFlow can be consulted in the readthedocs page: pdfflow.readthedocs.io.
The package can be installed with pip:
python3 -m pip install pdfflow[MODE]If you prefer a manual installation just cd in the cloned folder and use:
pip install .[MODE]or if you are planning to extend or develop code just install the package in editable mode:
pip install -e .[MODE]PDFFlow assumes that the user has already installed the most optimized version
of TensorFlow for his platform. As such, by default, pip will not check it as
a requirement.
However, the user can also install it specifying a MODE option in the
pip command. The list below summarizes the valid choices for the MODE flag:
tf: installs thetensorflowpackagetf-cpu: installs thetensorflow-cpupackagetf-gpu: installs thetensorflow-gpupackagetf-amd: installs thetensorflow-rocmpackage
⚠ Note: Use the latest version of TensorFlow!
TensorFlow is updated frequently and a later version of TensorFlow will often
offer better performance in both GPUs and CPUs.
Although it can be made to work with earlier versions, PDFFlow is only
supported for TensorFlow>2.1.
PDFFlow does not do management of PDF sets, which is left to LHAPDF and so a lhapdf installation is needed.
A full lhapdf installation can be obtained by utilizing the lhapdf_management library.
python3 -m pip install lhapdf_management
lhapdf_management install NNPDF31_nnlo_as_0118Below a minimalistic example where PDFFlow is used to generate a 10 values of the PDF
for 2 members for three different flavours.
from pdfflow import mkPDFs
import tensorflow as tf
pdf = mkPDFs("NNPDF31_nnlo_as_0118", [0,2])
x = tf.random.uniform([10], dtype=tf.float64)
q2 = tf.random.uniform([10], dtype=tf.float64)*20 + 10
pid = tf.cast([-1,21,1], dtype=tf.int32)
result = pdf.xfxQ2(pid, x, q2)Note the usage of the dtype keyword inm the TensorFlow calls.
This is used to ensure that float64 is being used all across the program.
For convenience, we ship two functions, int_me and float_me which are simply
wrappers to tf.cast with the right types.
These wrappers can be used over TensorFlow types but also numpy values:
from pdfflow import mkPDFs, int_me, float_me
import tensorflow as tf
import numpy as np
pdf = mkPDFs("NNPDF31_nnlo_as_0118", [0,2])
x = float_me(np.random.rand(10))
q2 = float_me(tf.random.uniform([10])*20 + 10)
pid = int_me([-1,21,1])
result = pdf.xfxQ2(pid, x, q2)If you use the package pelase cite the following paper and zenodo references: