node-calls-python

node-calls-python - call Python from Node.js directly in-process without spawning processes

Suitable for running your ML or deep learning models from Node directly

Motivation

Current solutions spawn a new process whenever you want to run Python code in Node.js and communicate via IPC using sockets, stdin/stdout, etc. But creating new processes every time you want to run Python code could be a major overhead and can lead to significant performance penalties. If the execution time of your Python code is less than creating a new process, you will see significant performance problems because your Node.js code will keep creating new processes instead of executing your Python code. Suppose you have a few NumPy calls in Python: do you want to create a new process for that? I guess your answer is no. In this case, running the Python code in-process is a much better solution because using the embedded Python interpreter is much faster than creating new processes and does not require any IPC to pass the data around. The data can stay in memory and requires only some conversions between Python and Node types (using the N-API and Python C API).

Installation

npm install node-calls-python

Installation FAQ

Sometimes you have to install prerequisites to make it work.

Linux: install node, npm, node-gyp, python3, python3-dev, g++ and make

Install Node

sudo apt install curl
curl -sL https://deb.nodesource.com/setup_13.x | sudo -E bash -
sudo apt install nodejs

Install Python

sudo apt install python3
sudo apt install python3-dev

Install Node-gyp

sudo apt install make
sudo apt install g++
sudo npm install -g node-gyp

Windows: install NodeJS and Python

Install Node-gyp if missing

npm install --global --production windows-build-tools
npm install -g node-gyp

Mac: install XCode from AppStore, Node.js and Python

npm install node-calls-python

If you see installation problems on Mac with ARM (E.g. using M1 Pro), try to specify 'arch' and/or 'target_arch' parameters for npm

npm install --arch=arm64 --target_arch=arm64 node-calls-python

Examples

Calling a simple python function

Let's say you have the following python code in test.py

import numpy as np

def multiple(a, b):
    return np.multiply(a, b).tolist()

Then to call this function directly you can do this in Node

const nodecallspython = require("node-calls-python");

const py = nodecallspython.interpreter;

py.import("path/to/test.py").then(async function(pymodule) {
    const result = await py.call(pymodule, "multiple", [1, 2, 3, 4], [2, 3, 4, 5]);
    console.log(result);
});

Or to call this function by using the synchronous version

const nodecallspython = require("node-calls-python");

const py = nodecallspython.interpreter;

py.import("path/to/test.py").then(async function(pymodule) {
    const result = py.callSync(pymodule, "multiple", [1, 2, 3, 4], [2, 3, 4, 5]);
    console.log(result);
});

Creating python objects

Let's say you have the following python code in test.py

import numpy as np

class Calculator:
    vector = []

    def __init__(self, vector):
        self.vector = vector

    def multiply(self, scalar, vector):
        return np.add(np.multiply(scalar, self.vector), vector).tolist()

Then to instance the class directly in Node

const nodecallspython = require("node-calls-python");

const py = nodecallspython.interpreter;

py.import("path/to/test.py").then(async function(pymodule) {
    const pyobj = await py.create(pymodule, "Calculator", [1.4, 5.5, 1.2, 4.4]);
    const result = await py.call(pyobj, "multiply", 2, [10.4, 50.5, 10.2, 40.4]);
});

Or to instance the class synchronously and directly in Node

const nodecallspython = require("node-calls-python");

const py = nodecallspython.interpreter;

py.import("path/to/test.py").then(async function(pymodule) {
    const pyobj = py.createSync(pymodule, "Calculator", [1.4, 5.5, 1.2, 4.4]);
    const result = await py.callSync(pyobj, "multiply", 2, [10.4, 50.5, 10.2, 40.4]); // you can use async version (call) as well
});

Running python code

const nodecallspython = require("node-calls-python");

const py = nodecallspython.interpreter;

py.import("path/to/test.py").then(async function(pymodule) {
    await py.exec(pymodule, "run_my_code(1, 2, 3)"); // exec will run any python code but the return value is not propagated
    const result = await py.eval(pymodule, "run_my_code(1, 2, 3)"); // result will hold the output of run_my_code
    console.log(result);
});

Running python code synchronously

const nodecallspython = require("node-calls-python");

const py = nodecallspython.interpreter;

const pymodule = py.importSync("path/to/test.py");
await py.execSync(pymodule, "run_my_code(1, 2, 3)"); // exec will run any python code but the return value is not propagated
const result = py.evalSync(pymodule, "run_my_code(1, 2, 3)"); // result will hold the output of run_my_code
console.log(result);

Reimporting a python module

You have to set allowReimport paramter to true when calling import/importSync.

const nodecallspython = require("node-calls-python");

const py = nodecallspython.interpreter;

let pymodule = py.importSync("path/to/test.py");
pymodule = py.importSync("path/to/test.py", true);

Passing kwargs

Javascript has no similar concept to kwargs of Python. Therefore a little hack is needed here. If you pass an object with __kwargs property set to true as a parameter to call/callSync/create/createSync the object will be mapped to kwargs.

const nodecallspython = require("node-calls-python");

const py = nodecallspython.interpreter;

let pymodule = py.importSync("path/to/test.py");
py.callSync(pymodule, "your_function", arg1, arg2, {"name1": value1, "name2": value2, "__kwargs": true })

def your_function(arg1, arg2, **kwargs):
    print(kwargs)

Doing some ML with Python and Node

Let's say you have the following python code in logreg.py

from sklearn.datasets import load_iris, load_digits
from sklearn.linear_model import LogisticRegression

class LogReg:
    logreg = None

    def __init__(self, dataset):
        if (dataset == "iris"):
            X, y = load_iris(return_X_y=True)
        else:
            X, y = load_digits(return_X_y=True)

        self.logreg = LogisticRegression(random_state=42, solver='lbfgs', multi_class='multinomial')
        self.logreg.fit(X, y)

    def predict(self, X):
        return self.logreg.predict_proba(X).tolist()

Then you can do this in Node

const nodecallspython = require("node-calls-python");

const py = nodecallspython.interpreter;

py.import("logreg.py")).then(async function(pymodule) { // import the python module
    const logreg = await py.create(pymodule, "LogReg", "iris"); // create the instance of the classifier

    const predict = await py.call(logreg, "predict", [[1.4, 5.5, 1.2, 4.4]]); // call predict
    console.log(predict);
});

Using Python venv

You have to add the proper import path so that python could use your installed packages from your venv.

If you have created a venv by python -m venv your-venv your installed python packages can be found under your-venv/Lib/site-packages. So you have to use addImportPath before importing any module to pick up the python packages from your venv.

const nodecallspython = require("node-calls-python");

const py = nodecallspython.interpreter;

py.addImportPath(your-venv/Lib/site-packages)

Working Around Linking Errors on Linux

If you get an error like this while trying to call Python code ImportError: /usr/local/lib/python3.7/dist-packages/cpython-37m-arm-linux-gnueabihf.so: undefined symbol: PyExc_RuntimeError

You can fix it by passing the name of your libpython shared library to fixlink

const nodecallspython = require("node-calls-python");

const py = nodecallspython.interpreter;
py.fixlink('libpython3.7m.so');

See more examples here

Supported data mapping

From Node to Python

  - undefined to None
  - null to None
  - boolean to boolean
  - number to double or long (as appropriate)
  - int32 to long
  - uint32 to long
  - int64 to long
  - string to unicode (string)
  - array to list
  - object to dictionary

From Python to Node

  - None to undefined
  - boolean to boolean
  - double to number
  - long to int64
  - unicode (string) to string
  - list to array
  - tuple to array
  - set to array
  - dictionary to object
  - numpy.array to array (this has limited support, will convert everything to number or string)