_test_distributed.py

import copy
import io
import socket
import subprocess
from concurrent.futures import ThreadPoolExecutor
from pathlib import Path
from typing import Any, Dict, Generator, List

import numpy as np
import pytest
from sklearn.datasets import make_blobs, make_regression
from sklearn.metrics import accuracy_score

TESTS_DIR = Path(__file__).absolute().parent


@pytest.fixture(scope='module')
def executable(pytestconfig) -> str:
    """Returns the path to the lightgbm executable."""
    return pytestconfig.getoption('execfile')


def _find_random_open_port() -> int:
    """Find a random open port on localhost."""
    with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
        s.bind(('', 0))
        port = s.getsockname()[1]
    return port


def _generate_n_ports(n: int) -> Generator[int, None, None]:
    return (_find_random_open_port() for _ in range(n))


def _write_dict(d: Dict, file: io.TextIOWrapper) -> None:
    for k, v in d.items():
        file.write(f'{k} = {v}\n')


def create_data(task: str, n_samples: int = 1_000) -> np.ndarray:
    """Create the appropriate data for the task.

    The data is returned as a numpy array with the label as the first column.
    """
    if task == 'binary-classification':
        centers = [[-4, -4], [4, 4]]
        X, y = make_blobs(n_samples, centers=centers, random_state=42)
    elif task == 'regression':
        X, y = make_regression(n_samples, n_features=4, n_informative=2, random_state=42)
    dataset = np.hstack([y.reshape(-1, 1), X])
    return dataset


class DistributedMockup:
    """Simulate distributed training."""

    default_train_config = {
        'task': 'train',
        'pre_partition': True,
        'machine_list_file': TESTS_DIR / 'mlist.txt',
        'tree_learner': 'data',
        'force_row_wise': True,
        'verbose': 0,
        'num_boost_round': 20,
        'num_leaves': 15,
        'num_threads': 2,
    }

    default_predict_config = {
        'task': 'predict',
        'data': TESTS_DIR / 'train.txt',
        'input_model': TESTS_DIR / 'model0.txt',
        'output_result': TESTS_DIR / 'predictions.txt',
    }

    def __init__(self, executable: str):
        self.executable = executable

    def worker_train(self, i: int) -> subprocess.CompletedProcess:
        """Start the training process on the `i`-th worker."""
        config_path = TESTS_DIR / f'train{i}.conf'
        cmd = [self.executable, f'config={config_path}']
        return subprocess.run(cmd)

    def _set_ports(self) -> None:
        """Randomly assign a port for training to each worker and save all ports to mlist.txt."""
        ports = set(_generate_n_ports(self.n_workers))
        i = 0
        max_tries = 100
        while i < max_tries and len(ports) < self.n_workers:
            n_ports_left = self.n_workers - len(ports)
            candidates = _generate_n_ports(n_ports_left)
            ports.update(candidates)
            i += 1
        if i == max_tries:
            raise RuntimeError('Unable to find non-colliding ports.')
        self.listen_ports = list(ports)
        with open(TESTS_DIR / 'mlist.txt', 'wt') as f:
            for port in self.listen_ports:
                f.write(f'127.0.0.1 {port}\n')

    def _write_data(self, partitions: List[np.ndarray]) -> None:
        """Write all training data as train.txt and each training partition as train{i}.txt."""
        all_data = np.vstack(partitions)
        np.savetxt(str(TESTS_DIR / 'train.txt'), all_data, delimiter=',')
        for i, partition in enumerate(partitions):
            np.savetxt(str(TESTS_DIR / f'train{i}.txt'), partition, delimiter=',')

    def fit(self, partitions: List[np.ndarray], train_config: Dict = {}) -> None:
        """Run the distributed training process on a single machine.

        For each worker i:
            1. The i-th partition is saved as train{i}.txt.
            2. A random port is assigned for training.
            3. A configuration file train{i}.conf is created.
            4. The lightgbm binary is called with config=train{i}.conf in another thread.
            5. The trained model is saved as model{i}.txt. Each model file only differs in data and local_listen_port.
        The whole training set is saved as train.txt.
        """
        self.train_config = copy.deepcopy(self.default_train_config)
        self.train_config.update(train_config)
        self.n_workers = self.train_config['num_machines']
        self._set_ports()
        self._write_data(partitions)
        self.label_ = np.hstack([partition[:, 0] for partition in partitions])
        futures = []
        with ThreadPoolExecutor(max_workers=self.n_workers) as executor:
            for i in range(self.n_workers):
                self.write_train_config(i)
                train_future = executor.submit(self.worker_train, i)
                futures.append(train_future)
            results = [f.result() for f in futures]
        for result in results:
            if result.returncode != 0:
                raise RuntimeError('Error in training')

    def predict(self, predict_config: Dict[str, Any] = {}) -> np.ndarray:
        """Compute the predictions using the model created in the fit step.

        predict_config is used to predict the training set train.txt
        The predictions are saved as predictions.txt and are then loaded to return them as a numpy array.
        """
        self.predict_config = copy.deepcopy(self.default_predict_config)
        self.predict_config.update(predict_config)
        config_path = TESTS_DIR / 'predict.conf'
        with open(config_path, 'wt') as file:
            _write_dict(self.predict_config, file)
        cmd = [self.executable, f'config={config_path}']
        result = subprocess.run(cmd)
        if result.returncode != 0:
            raise RuntimeError('Error in prediction')
        y_pred = np.loadtxt(str(TESTS_DIR / 'predictions.txt'))
        return y_pred

    def write_train_config(self, i: int) -> None:
        """Create a file train{i}.conf with the required configuration to train.

        Each worker gets a different port and piece of the data, the rest are the
        model parameters contained in `self.config`.
        """
        with open(TESTS_DIR / f'train{i}.conf', 'wt') as file:
            output_model = TESTS_DIR / f'model{i}.txt'
            data = TESTS_DIR / f'train{i}.txt'
            file.write(f'output_model = {output_model}\n')
            file.write(f'local_listen_port = {self.listen_ports[i]}\n')
            file.write(f'data = {data}\n')
            _write_dict(self.train_config, file)


def test_classifier(executable):
    """Test the classification task."""
    num_machines = 2
    data = create_data(task='binary-classification')
    partitions = np.array_split(data, num_machines)
    train_params = {
        'objective': 'binary',
        'num_machines': num_machines,
    }
    clf = DistributedMockup(executable)
    clf.fit(partitions, train_params)
    y_probas = clf.predict()
    y_pred = y_probas > 0.5
    assert accuracy_score(clf.label_, y_pred) == 1.


def test_regressor(executable):
    """Test the regression task."""
    num_machines = 2
    data = create_data(task='regression')
    partitions = np.array_split(data, num_machines)
    train_params = {
        'objective': 'regression',
        'num_machines': num_machines,
    }
    reg = DistributedMockup(executable)
    reg.fit(partitions, train_params)
    y_pred = reg.predict()
    np.testing.assert_allclose(y_pred, reg.label_, rtol=0.2, atol=50.)