C++ bindings for Python Data Pipeline API

This directory contains C++ bindings for the Python data pipeline API.

(See DiRAC CSD3 for specific notes on installing on that machine.)

Requirements

• Python 3.7 or greater with a working python3-config

• Create and activate a virtual environment for the required packages from the top-level repository directory:

  python3 -m venv .venv
  source .venv/bin/activate
  

• CMake will fetch and install all other requirements.

Building the C++ wrapper library

cmake . -Bbuild
cmake --build build

Run the C++ wrapper tests

The test binary can be found in the build folder after the build has completed:

./build/bin/datapipeline-tests

Usage

Where possible the usage of the API in C++ mirrors that of Python however there are a few exceptions.

Reading and Writing Distributions

Note that the procedure to read distributions is different to writing them. Due to the differences between C++ and Python, the retrieval method converts the Python scipy.stats object to a simple class Distribution from which parameters can be extracted. The reason for this wrapper class is that some variables exist within the kwargs member of the scipy.stats object, whereas others exist within args, the class unpacks these to be a single set in a way that varies depending on the type of distribution received.

To retrieve a parameter that is a single value (i.e. not an array) use the getParameter(<param_name>) method:

DataPipeline* dat_pipe = new DataPipeline("/path/to/config.yaml", github_uri, version);

// Retrieve the parameter 'k' from the Gamma distribution 'example-distribution'
const double k = dat_pipe->read_distribution("parameter", "example-distribution").getParameter("k");

for parameters which are an array, use getArrayParameter(<param_name>) which returns a std::vector<double>:

// Retrieve the parameter 'p' from the Multinomial distribution 'example-distribution'
const std::vector<double> dat_pipe->read_distribution("parameter", "example-distribution").getArrayParameter("p");

Writing distributions is easy, the following functions return pybind11::objects which are then handed to the API:

Gamma(double k, double theta);

Normal(double mu, double sigma);

Poisson(double lambda);

Multinomial(double n, std::vector<double> p);

Uniform(double a, double b);

Beta(double alpha, double beta);

Binomial(int n, double p);

for example, to write a Poisson distribution named "my_dist":

dat_pipe->write_distribution("distribution", "my_dist", Poisson(5));

Reading and Writing Parameters

Parameters are read and written in a manner identical to the Python API:

dat_pipe->read_estimate("parameter", "example-estimate");
dat_pipe->write_estimate("output-parameter", "example-estimate", 1.0);

Reading and Writing Tables

Tables are read into a Table object which is a set of STL vectors representing columns. The columns can then be retrieved:

Table table = pDataPipeline_->read_table("object", "example-table");

std::vector<long> col_a = table.get_column<long>("a");

Tables are assembled using STL vectors also:

Table table;
const std::vector<std::string> _alpha = {"A", "B", "C", "D", "E", "F"};
const std::vector<double> _numero = {0.5, 2.2, 3.4, 4.6, 5.2, 6.1};
const std::vector<int> _id = {0,1,2,3,4,5};
table.add_column("ALPHA", _alpha);
table.add_column("NUMERO", _numero);
table.add_column("ID", _id);
pDataPipeline_->write_table("output-table", "example-table", table);

Machine-specific notes

DiRAC CSD3

Follow the above instructions, but be sure to run

module load python/3.7

first. The default python does not have python3-config, which is required for the compilation.

Developer notes

• We have not been successful in running with Conda. The problem appears to be related to the provided Python being compiled with a different compiler than is used to build the C++ bindings.