test_dask.py

# coding: utf-8
"""Tests for lightgbm.dask module"""

import inspect
import pickle
import random
import socket
from itertools import groupby
from os import getenv
from platform import machine
from sys import platform

import pytest

import lightgbm as lgb

if not platform.startswith('linux'):
    pytest.skip('lightgbm.dask is currently supported in Linux environments', allow_module_level=True)
if not lgb.compat.DASK_INSTALLED:
    pytest.skip('Dask is not installed', allow_module_level=True)

import cloudpickle
import dask.array as da
import dask.dataframe as dd
import joblib
import numpy as np
import pandas as pd
import sklearn.utils.estimator_checks as sklearn_checks
from dask.array.utils import assert_eq
from dask.distributed import Client, LocalCluster, default_client, wait
from distributed.utils_test import client, cluster_fixture, gen_cluster, loop
from pkg_resources import parse_version
from scipy.sparse import csr_matrix
from scipy.stats import spearmanr
from sklearn import __version__ as sk_version
from sklearn.datasets import make_blobs, make_regression

from .utils import make_ranking

sk_version = parse_version(sk_version)

# time, in seconds, to wait for the Dask client to close. Used to avoid teardown errors
# see https://distributed.dask.org/en/latest/api.html#distributed.Client.close
CLIENT_CLOSE_TIMEOUT = 120

tasks = ['binary-classification', 'multiclass-classification', 'regression', 'ranking']
distributed_training_algorithms = ['data', 'voting']
data_output = ['array', 'scipy_csr_matrix', 'dataframe', 'dataframe-with-categorical']
boosting_types = ['gbdt', 'dart', 'goss', 'rf']
group_sizes = [5, 5, 5, 10, 10, 10, 20, 20, 20, 50, 50]
task_to_dask_factory = {
    'regression': lgb.DaskLGBMRegressor,
    'binary-classification': lgb.DaskLGBMClassifier,
    'multiclass-classification': lgb.DaskLGBMClassifier,
    'ranking': lgb.DaskLGBMRanker
}
task_to_local_factory = {
    'regression': lgb.LGBMRegressor,
    'binary-classification': lgb.LGBMClassifier,
    'multiclass-classification': lgb.LGBMClassifier,
    'ranking': lgb.LGBMRanker
}

pytestmark = [
    pytest.mark.skipif(getenv('TASK', '') == 'mpi', reason='Fails to run with MPI interface'),
    pytest.mark.skipif(getenv('TASK', '') == 'gpu', reason='Fails to run with GPU interface'),
    pytest.mark.skipif(machine() != 'x86_64', reason='Fails to run with non-x86_64 architecture')
]


@pytest.fixture()
def listen_port():
    listen_port.port += 10
    return listen_port.port


listen_port.port = 13000


def _create_ranking_data(n_samples=100, output='array', chunk_size=50, **kwargs):
    X, y, g = make_ranking(n_samples=n_samples, random_state=42, **kwargs)
    rnd = np.random.RandomState(42)
    w = rnd.rand(X.shape[0]) * 0.01
    g_rle = np.array([len(list(grp)) for _, grp in groupby(g)])

    if output.startswith('dataframe'):
        # add target, weight, and group to DataFrame so that partitions abide by group boundaries.
        X_df = pd.DataFrame(X, columns=[f'feature_{i}' for i in range(X.shape[1])])
        if output == 'dataframe-with-categorical':
            for i in range(5):
                col_name = "cat_col" + str(i)
                cat_values = rnd.choice(['a', 'b'], X.shape[0])
                cat_series = pd.Series(
                    cat_values,
                    dtype='category'
                )
                X_df[col_name] = cat_series
        X = X_df.copy()
        X_df = X_df.assign(y=y, g=g, w=w)

        # set_index ensures partitions are based on group id.
        # See https://stackoverflow.com/questions/49532824/dask-dataframe-split-partitions-based-on-a-column-or-function.
        X_df.set_index('g', inplace=True)
        dX = dd.from_pandas(X_df, chunksize=chunk_size)

        # separate target, weight from features.
        dy = dX['y']
        dw = dX['w']
        dX = dX.drop(columns=['y', 'w'])
        dg = dX.index.to_series()

        # encode group identifiers into run-length encoding, the format LightGBMRanker is expecting
        # so that within each partition, sum(g) = n_samples.
        dg = dg.map_partitions(lambda p: p.groupby('g', sort=False).apply(lambda z: z.shape[0]))
    elif output == 'array':
        # ranking arrays: one chunk per group. Each chunk must include all columns.
        p = X.shape[1]
        dX, dy, dw, dg = [], [], [], []
        for g_idx, rhs in enumerate(np.cumsum(g_rle)):
            lhs = rhs - g_rle[g_idx]
            dX.append(da.from_array(X[lhs:rhs, :], chunks=(rhs - lhs, p)))
            dy.append(da.from_array(y[lhs:rhs]))
            dw.append(da.from_array(w[lhs:rhs]))
            dg.append(da.from_array(np.array([g_rle[g_idx]])))

        dX = da.concatenate(dX, axis=0)
        dy = da.concatenate(dy, axis=0)
        dw = da.concatenate(dw, axis=0)
        dg = da.concatenate(dg, axis=0)
    else:
        raise ValueError('Ranking data creation only supported for Dask arrays and dataframes')

    return X, y, w, g_rle, dX, dy, dw, dg


def _create_data(objective, n_samples=1_000, output='array', chunk_size=500, **kwargs):
    if objective.endswith('classification'):
        if objective == 'binary-classification':
            centers = [[-4, -4], [4, 4]]
        elif objective == 'multiclass-classification':
            centers = [[-4, -4], [4, 4], [-4, 4]]
        else:
            raise ValueError(f"Unknown classification task '{objective}'")
        X, y = make_blobs(n_samples=n_samples, centers=centers, random_state=42)
    elif objective == 'regression':
        X, y = make_regression(n_samples=n_samples, n_features=4, n_informative=2, random_state=42)
    elif objective == 'ranking':
        return _create_ranking_data(
            n_samples=n_samples,
            output=output,
            chunk_size=chunk_size,
            **kwargs
        )
    else:
        raise ValueError("Unknown objective '%s'" % objective)
    rnd = np.random.RandomState(42)
    weights = rnd.random(X.shape[0]) * 0.01

    if output == 'array':
        dX = da.from_array(X, (chunk_size, X.shape[1]))
        dy = da.from_array(y, chunk_size)
        dw = da.from_array(weights, chunk_size)
    elif output.startswith('dataframe'):
        X_df = pd.DataFrame(X, columns=['feature_%d' % i for i in range(X.shape[1])])
        if output == 'dataframe-with-categorical':
            num_cat_cols = 2
            for i in range(num_cat_cols):
                col_name = "cat_col" + str(i)
                cat_values = rnd.choice(['a', 'b'], X.shape[0])
                cat_series = pd.Series(
                    cat_values,
                    dtype='category'
                )
                X_df[col_name] = cat_series
                X = np.hstack((X, cat_series.cat.codes.values.reshape(-1, 1)))

            # make one categorical feature relevant to the target
            cat_col_is_a = X_df['cat_col0'] == 'a'
            if objective == 'regression':
                y = np.where(cat_col_is_a, y, 2 * y)
            elif objective == 'binary-classification':
                y = np.where(cat_col_is_a, y, 1 - y)
            elif objective == 'multiclass-classification':
                n_classes = 3
                y = np.where(cat_col_is_a, y, (1 + y) % n_classes)
        y_df = pd.Series(y, name='target')
        dX = dd.from_pandas(X_df, chunksize=chunk_size)
        dy = dd.from_pandas(y_df, chunksize=chunk_size)
        dw = dd.from_array(weights, chunksize=chunk_size)
    elif output == 'scipy_csr_matrix':
        dX = da.from_array(X, chunks=(chunk_size, X.shape[1])).map_blocks(csr_matrix)
        dy = da.from_array(y, chunks=chunk_size)
        dw = da.from_array(weights, chunk_size)
    else:
        raise ValueError("Unknown output type '%s'" % output)

    return X, y, weights, None, dX, dy, dw, None


def _r2_score(dy_true, dy_pred):
    numerator = ((dy_true - dy_pred) ** 2).sum(axis=0, dtype=np.float64)
    denominator = ((dy_true - dy_true.mean(axis=0)) ** 2).sum(axis=0, dtype=np.float64)
    return (1 - numerator / denominator).compute()


def _accuracy_score(dy_true, dy_pred):
    return da.average(dy_true == dy_pred).compute()


def _pickle(obj, filepath, serializer):
    if serializer == 'pickle':
        with open(filepath, 'wb') as f:
            pickle.dump(obj, f)
    elif serializer == 'joblib':
        joblib.dump(obj, filepath)
    elif serializer == 'cloudpickle':
        with open(filepath, 'wb') as f:
            cloudpickle.dump(obj, f)
    else:
        raise ValueError(f'Unrecognized serializer type: {serializer}')


def _unpickle(filepath, serializer):
    if serializer == 'pickle':
        with open(filepath, 'rb') as f:
            return pickle.load(f)
    elif serializer == 'joblib':
        return joblib.load(filepath)
    elif serializer == 'cloudpickle':
        with open(filepath, 'rb') as f:
            return cloudpickle.load(f)
    else:
        raise ValueError(f'Unrecognized serializer type: {serializer}')


@pytest.mark.parametrize('output', data_output)
@pytest.mark.parametrize('task', ['binary-classification', 'multiclass-classification'])
@pytest.mark.parametrize('boosting_type', boosting_types)
@pytest.mark.parametrize('tree_learner', distributed_training_algorithms)
def test_classifier(output, task, boosting_type, tree_learner, client):
    X, y, w, _, dX, dy, dw, _ = _create_data(
        objective=task,
        output=output
    )

    params = {
        "boosting_type": boosting_type,
        "tree_learner": tree_learner,
        "n_estimators": 50,
        "num_leaves": 31
    }
    if boosting_type == 'rf':
        params.update({
            'bagging_freq': 1,
            'bagging_fraction': 0.9,
        })
    elif boosting_type == 'goss':
        params['top_rate'] = 0.5

    dask_classifier = lgb.DaskLGBMClassifier(
        client=client,
        time_out=5,
        **params
    )
    dask_classifier = dask_classifier.fit(dX, dy, sample_weight=dw)
    p1 = dask_classifier.predict(dX)
    p1_proba = dask_classifier.predict_proba(dX).compute()
    p1_pred_leaf = dask_classifier.predict(dX, pred_leaf=True)
    p1_local = dask_classifier.to_local().predict(X)
    s1 = _accuracy_score(dy, p1)
    p1 = p1.compute()

    local_classifier = lgb.LGBMClassifier(**params)
    local_classifier.fit(X, y, sample_weight=w)
    p2 = local_classifier.predict(X)
    p2_proba = local_classifier.predict_proba(X)
    s2 = local_classifier.score(X, y)

    if boosting_type == 'rf':
        # https://github.com/microsoft/LightGBM/issues/4118
        assert_eq(s1, s2, atol=0.01)
        assert_eq(p1_proba, p2_proba, atol=0.8)
    else:
        assert_eq(s1, s2)
        assert_eq(p1, p2)
        assert_eq(p1, y)
        assert_eq(p2, y)
        assert_eq(p1_proba, p2_proba, atol=0.03)
        assert_eq(p1_local, p2)
        assert_eq(p1_local, y)

    # pref_leaf values should have the right shape
    # and values that look like valid tree nodes
    pred_leaf_vals = p1_pred_leaf.compute()
    assert pred_leaf_vals.shape == (
        X.shape[0],
        dask_classifier.booster_.num_trees()
    )
    assert np.max(pred_leaf_vals) <= params['num_leaves']
    assert np.min(pred_leaf_vals) >= 0
    assert len(np.unique(pred_leaf_vals)) <= params['num_leaves']

    # be sure LightGBM actually used at least one categorical column,
    # and that it was correctly treated as a categorical feature
    if output == 'dataframe-with-categorical':
        cat_cols = [
            col for col in dX.columns
            if dX.dtypes[col].name == 'category'
        ]
        tree_df = dask_classifier.booster_.trees_to_dataframe()
        node_uses_cat_col = tree_df['split_feature'].isin(cat_cols)
        assert node_uses_cat_col.sum() > 0
        assert tree_df.loc[node_uses_cat_col, "decision_type"].unique()[0] == '=='

    client.close(timeout=CLIENT_CLOSE_TIMEOUT)


@pytest.mark.parametrize('output', data_output)
@pytest.mark.parametrize('task', ['binary-classification', 'multiclass-classification'])
def test_classifier_pred_contrib(output, task, client):
    X, y, w, _, dX, dy, dw, _ = _create_data(
        objective=task,
        output=output
    )

    params = {
        "n_estimators": 10,
        "num_leaves": 10
    }

    dask_classifier = lgb.DaskLGBMClassifier(
        client=client,
        time_out=5,
        tree_learner='data',
        **params
    )
    dask_classifier = dask_classifier.fit(dX, dy, sample_weight=dw)
    preds_with_contrib = dask_classifier.predict(dX, pred_contrib=True).compute()

    local_classifier = lgb.LGBMClassifier(**params)
    local_classifier.fit(X, y, sample_weight=w)
    local_preds_with_contrib = local_classifier.predict(X, pred_contrib=True)

    if output == 'scipy_csr_matrix':
        preds_with_contrib = np.array(preds_with_contrib.todense())

    # be sure LightGBM actually used at least one categorical column,
    # and that it was correctly treated as a categorical feature
    if output == 'dataframe-with-categorical':
        cat_cols = [
            col for col in dX.columns
            if dX.dtypes[col].name == 'category'
        ]
        tree_df = dask_classifier.booster_.trees_to_dataframe()
        node_uses_cat_col = tree_df['split_feature'].isin(cat_cols)
        assert node_uses_cat_col.sum() > 0
        assert tree_df.loc[node_uses_cat_col, "decision_type"].unique()[0] == '=='

    # shape depends on whether it is binary or multiclass classification
    num_features = dask_classifier.n_features_
    num_classes = dask_classifier.n_classes_
    if num_classes == 2:
        expected_num_cols = num_features + 1
    else:
        expected_num_cols = (num_features + 1) * num_classes

    # * shape depends on whether it is binary or multiclass classification
    # * matrix for binary classification is of the form [feature_contrib, base_value],
    #   for multi-class it's [feat_contrib_class1, base_value_class1, feat_contrib_class2, base_value_class2, etc.]
    # * contrib outputs for distributed training are different than from local training, so we can just test
    #   that the output has the right shape and base values are in the right position
    assert preds_with_contrib.shape[1] == expected_num_cols
    assert preds_with_contrib.shape == local_preds_with_contrib.shape

    if num_classes == 2:
        assert len(np.unique(preds_with_contrib[:, num_features]) == 1)
    else:
        for i in range(num_classes):
            base_value_col = num_features * (i + 1) + i
            assert len(np.unique(preds_with_contrib[:, base_value_col]) == 1)

    client.close(timeout=CLIENT_CLOSE_TIMEOUT)


def test_find_random_open_port(client):
    for _ in range(5):
        worker_address_to_port = client.run(lgb.dask._find_random_open_port)
        found_ports = worker_address_to_port.values()
        # check that found ports are different for same address (LocalCluster)
        assert len(set(found_ports)) == len(found_ports)
        # check that the ports are indeed open
        for port in found_ports:
            with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
                s.bind(('', port))
    client.close(timeout=CLIENT_CLOSE_TIMEOUT)


def test_possibly_fix_worker_map(capsys, client):
    client.wait_for_workers(2)
    worker_addresses = list(client.scheduler_info()["workers"].keys())

    retry_msg = 'Searching for a LightGBM training port for worker'

    # should handle worker maps without any duplicates
    map_without_duplicates = {
        worker_address: 12400 + i
        for i, worker_address in enumerate(worker_addresses)
    }
    patched_map = lgb.dask._possibly_fix_worker_map_duplicates(
        client=client,
        worker_map=map_without_duplicates
    )
    assert patched_map == map_without_duplicates
    assert retry_msg not in capsys.readouterr().out

    # should handle worker maps with duplicates
    map_with_duplicates = {
        worker_address: 12400
        for i, worker_address in enumerate(worker_addresses)
    }
    patched_map = lgb.dask._possibly_fix_worker_map_duplicates(
        client=client,
        worker_map=map_with_duplicates
    )
    assert retry_msg in capsys.readouterr().out
    assert len(set(patched_map.values())) == len(worker_addresses)


def test_training_does_not_fail_on_port_conflicts(client):
    _, _, _, _, dX, dy, dw, _ = _create_data('binary-classification', output='array')

    lightgbm_default_port = 12400
    with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
        s.bind(('127.0.0.1', lightgbm_default_port))
        dask_classifier = lgb.DaskLGBMClassifier(
            client=client,
            time_out=5,
            n_estimators=5,
            num_leaves=5
        )
        for _ in range(5):
            dask_classifier.fit(
                X=dX,
                y=dy,
                sample_weight=dw,
            )
            assert dask_classifier.booster_

    client.close(timeout=CLIENT_CLOSE_TIMEOUT)


@pytest.mark.parametrize('output', data_output)
@pytest.mark.parametrize('boosting_type', boosting_types)
@pytest.mark.parametrize('tree_learner', distributed_training_algorithms)
def test_regressor(output, boosting_type, tree_learner, client):
    X, y, w, _, dX, dy, dw, _ = _create_data(
        objective='regression',
        output=output
    )

    params = {
        "boosting_type": boosting_type,
        "random_state": 42,
        "num_leaves": 31,
        "n_estimators": 20,
    }
    if boosting_type == 'rf':
        params.update({
            'bagging_freq': 1,
            'bagging_fraction': 0.9,
        })

    dask_regressor = lgb.DaskLGBMRegressor(
        client=client,
        time_out=5,
        tree=tree_learner,
        **params
    )
    dask_regressor = dask_regressor.fit(dX, dy, sample_weight=dw)
    p1 = dask_regressor.predict(dX)
    p1_pred_leaf = dask_regressor.predict(dX, pred_leaf=True)

    s1 = _r2_score(dy, p1)
    p1 = p1.compute()
    p1_local = dask_regressor.to_local().predict(X)
    s1_local = dask_regressor.to_local().score(X, y)

    local_regressor = lgb.LGBMRegressor(**params)
    local_regressor.fit(X, y, sample_weight=w)
    s2 = local_regressor.score(X, y)
    p2 = local_regressor.predict(X)

    # Scores should be the same
    assert_eq(s1, s2, atol=0.01)
    assert_eq(s1, s1_local)

    # Predictions should be roughly the same.
    assert_eq(p1, p1_local)

    # pref_leaf values should have the right shape
    # and values that look like valid tree nodes
    pred_leaf_vals = p1_pred_leaf.compute()
    assert pred_leaf_vals.shape == (
        X.shape[0],
        dask_regressor.booster_.num_trees()
    )
    assert np.max(pred_leaf_vals) <= params['num_leaves']
    assert np.min(pred_leaf_vals) >= 0
    assert len(np.unique(pred_leaf_vals)) <= params['num_leaves']

    assert_eq(p1, y, rtol=0.5, atol=50.)
    assert_eq(p2, y, rtol=0.5, atol=50.)

    # be sure LightGBM actually used at least one categorical column,
    # and that it was correctly treated as a categorical feature
    if output == 'dataframe-with-categorical':
        cat_cols = [
            col for col in dX.columns
            if dX.dtypes[col].name == 'category'
        ]
        tree_df = dask_regressor.booster_.trees_to_dataframe()
        node_uses_cat_col = tree_df['split_feature'].isin(cat_cols)
        assert node_uses_cat_col.sum() > 0
        assert tree_df.loc[node_uses_cat_col, "decision_type"].unique()[0] == '=='

    client.close(timeout=CLIENT_CLOSE_TIMEOUT)


@pytest.mark.parametrize('output', data_output)
def test_regressor_pred_contrib(output, client):
    X, y, w, _, dX, dy, dw, _ = _create_data(
        objective='regression',
        output=output
    )

    params = {
        "n_estimators": 10,
        "num_leaves": 10
    }

    dask_regressor = lgb.DaskLGBMRegressor(
        client=client,
        time_out=5,
        tree_learner='data',
        **params
    )
    dask_regressor = dask_regressor.fit(dX, dy, sample_weight=dw)
    preds_with_contrib = dask_regressor.predict(dX, pred_contrib=True).compute()

    local_regressor = lgb.LGBMRegressor(**params)
    local_regressor.fit(X, y, sample_weight=w)
    local_preds_with_contrib = local_regressor.predict(X, pred_contrib=True)

    if output == "scipy_csr_matrix":
        preds_with_contrib = np.array(preds_with_contrib.todense())

    # contrib outputs for distributed training are different than from local training, so we can just test
    # that the output has the right shape and base values are in the right position
    num_features = dX.shape[1]
    assert preds_with_contrib.shape[1] == num_features + 1
    assert preds_with_contrib.shape == local_preds_with_contrib.shape

    # be sure LightGBM actually used at least one categorical column,
    # and that it was correctly treated as a categorical feature
    if output == 'dataframe-with-categorical':
        cat_cols = [
            col for col in dX.columns
            if dX.dtypes[col].name == 'category'
        ]
        tree_df = dask_regressor.booster_.trees_to_dataframe()
        node_uses_cat_col = tree_df['split_feature'].isin(cat_cols)
        assert node_uses_cat_col.sum() > 0
        assert tree_df.loc[node_uses_cat_col, "decision_type"].unique()[0] == '=='

    client.close(timeout=CLIENT_CLOSE_TIMEOUT)


@pytest.mark.parametrize('output', data_output)
@pytest.mark.parametrize('alpha', [.1, .5, .9])
def test_regressor_quantile(output, client, alpha):
    X, y, w, _, dX, dy, dw, _ = _create_data(
        objective='regression',
        output=output
    )

    params = {
        "objective": "quantile",
        "alpha": alpha,
        "random_state": 42,
        "n_estimators": 10,
        "num_leaves": 10
    }

    dask_regressor = lgb.DaskLGBMRegressor(
        client=client,
        tree_learner_type='data_parallel',
        **params
    )
    dask_regressor = dask_regressor.fit(dX, dy, sample_weight=dw)
    p1 = dask_regressor.predict(dX).compute()
    q1 = np.count_nonzero(y < p1) / y.shape[0]

    local_regressor = lgb.LGBMRegressor(**params)
    local_regressor.fit(X, y, sample_weight=w)
    p2 = local_regressor.predict(X)
    q2 = np.count_nonzero(y < p2) / y.shape[0]

    # Quantiles should be right
    np.testing.assert_allclose(q1, alpha, atol=0.2)
    np.testing.assert_allclose(q2, alpha, atol=0.2)

    # be sure LightGBM actually used at least one categorical column,
    # and that it was correctly treated as a categorical feature
    if output == 'dataframe-with-categorical':
        cat_cols = [
            col for col in dX.columns
            if dX.dtypes[col].name == 'category'
        ]
        tree_df = dask_regressor.booster_.trees_to_dataframe()
        node_uses_cat_col = tree_df['split_feature'].isin(cat_cols)
        assert node_uses_cat_col.sum() > 0
        assert tree_df.loc[node_uses_cat_col, "decision_type"].unique()[0] == '=='

    client.close(timeout=CLIENT_CLOSE_TIMEOUT)


@pytest.mark.parametrize('output', ['array', 'dataframe', 'dataframe-with-categorical'])
@pytest.mark.parametrize('group', [None, group_sizes])
@pytest.mark.parametrize('boosting_type', boosting_types)
@pytest.mark.parametrize('tree_learner', distributed_training_algorithms)
def test_ranker(output, group, boosting_type, tree_learner, client):
    if output == 'dataframe-with-categorical':
        X, y, w, g, dX, dy, dw, dg = _create_data(
            objective='ranking',
            output=output,
            group=group,
            n_features=1,
            n_informative=1
        )
    else:
        X, y, w, g, dX, dy, dw, dg = _create_data(
            objective='ranking',
            output=output,
            group=group
        )

    # rebalance small dask.Array dataset for better performance.
    if output == 'array':
        dX = dX.persist()
        dy = dy.persist()
        dw = dw.persist()
        dg = dg.persist()
        _ = wait([dX, dy, dw, dg])
        client.rebalance()

    # use many trees + leaves to overfit, help ensure that Dask data-parallel strategy matches that of
    # serial learner. See https://github.com/microsoft/LightGBM/issues/3292#issuecomment-671288210.
    params = {
        "boosting_type": boosting_type,
        "random_state": 42,
        "n_estimators": 50,
        "num_leaves": 20,
        "min_child_samples": 1
    }
    if boosting_type == 'rf':
        params.update({
            'bagging_freq': 1,
            'bagging_fraction': 0.9,
        })

    dask_ranker = lgb.DaskLGBMRanker(
        client=client,
        time_out=5,
        tree_learner_type=tree_learner,
        **params
    )
    dask_ranker = dask_ranker.fit(dX, dy, sample_weight=dw, group=dg)
    rnkvec_dask = dask_ranker.predict(dX)
    rnkvec_dask = rnkvec_dask.compute()
    p1_pred_leaf = dask_ranker.predict(dX, pred_leaf=True)
    rnkvec_dask_local = dask_ranker.to_local().predict(X)

    local_ranker = lgb.LGBMRanker(**params)
    local_ranker.fit(X, y, sample_weight=w, group=g)
    rnkvec_local = local_ranker.predict(X)

    # distributed ranker should be able to rank decently well and should
    # have high rank correlation with scores from serial ranker.
    dcor = spearmanr(rnkvec_dask, y).correlation
    assert dcor > 0.6
    assert spearmanr(rnkvec_dask, rnkvec_local).correlation > 0.8
    assert_eq(rnkvec_dask, rnkvec_dask_local)

    # pref_leaf values should have the right shape
    # and values that look like valid tree nodes
    pred_leaf_vals = p1_pred_leaf.compute()
    assert pred_leaf_vals.shape == (
        X.shape[0],
        dask_ranker.booster_.num_trees()
    )
    assert np.max(pred_leaf_vals) <= params['num_leaves']
    assert np.min(pred_leaf_vals) >= 0
    assert len(np.unique(pred_leaf_vals)) <= params['num_leaves']

    # be sure LightGBM actually used at least one categorical column,
    # and that it was correctly treated as a categorical feature
    if output == 'dataframe-with-categorical':
        cat_cols = [
            col for col in dX.columns
            if dX.dtypes[col].name == 'category'
        ]
        tree_df = dask_ranker.booster_.trees_to_dataframe()
        node_uses_cat_col = tree_df['split_feature'].isin(cat_cols)
        assert node_uses_cat_col.sum() > 0
        assert tree_df.loc[node_uses_cat_col, "decision_type"].unique()[0] == '=='

    client.close(timeout=CLIENT_CLOSE_TIMEOUT)


@pytest.mark.parametrize('task', tasks)
def test_training_works_if_client_not_provided_or_set_after_construction(task, client):
    _, _, _, _, dX, dy, _, dg = _create_data(
        objective=task,
        output='array',
        group=None
    )
    model_factory = task_to_dask_factory[task]

    params = {
        "time_out": 5,
        "n_estimators": 1,
        "num_leaves": 2
    }

    # should be able to use the class without specifying a client
    dask_model = model_factory(**params)
    assert dask_model.client is None
    with pytest.raises(lgb.compat.LGBMNotFittedError, match='Cannot access property client_ before calling fit'):
        dask_model.client_

    dask_model.fit(dX, dy, group=dg)
    assert dask_model.fitted_
    assert dask_model.client is None
    assert dask_model.client_ == client

    preds = dask_model.predict(dX)
    assert isinstance(preds, da.Array)
    assert dask_model.fitted_
    assert dask_model.client is None
    assert dask_model.client_ == client

    local_model = dask_model.to_local()
    with pytest.raises(AttributeError):
        local_model.client
        local_model.client_

    # should be able to set client after construction
    dask_model = model_factory(**params)
    dask_model.set_params(client=client)
    assert dask_model.client == client

    with pytest.raises(lgb.compat.LGBMNotFittedError, match='Cannot access property client_ before calling fit'):
        dask_model.client_

    dask_model.fit(dX, dy, group=dg)
    assert dask_model.fitted_
    assert dask_model.client == client
    assert dask_model.client_ == client

    preds = dask_model.predict(dX)
    assert isinstance(preds, da.Array)
    assert dask_model.fitted_
    assert dask_model.client == client
    assert dask_model.client_ == client

    local_model = dask_model.to_local()
    with pytest.raises(AttributeError):
        local_model.client
        local_model.client_

    client.close(timeout=CLIENT_CLOSE_TIMEOUT)


@pytest.mark.parametrize('serializer', ['pickle', 'joblib', 'cloudpickle'])
@pytest.mark.parametrize('task', tasks)
@pytest.mark.parametrize('set_client', [True, False])
def test_model_and_local_version_are_picklable_whether_or_not_client_set_explicitly(serializer, task, set_client, tmp_path):

    with LocalCluster(n_workers=2, threads_per_worker=1) as cluster1, Client(cluster1) as client1:
        # data on cluster1
        X_1, _, _, _, dX_1, dy_1, _, dg_1 = _create_data(
            objective=task,
            output='array',
            group=None
        )

        with LocalCluster(n_workers=2, threads_per_worker=1) as cluster2, Client(cluster2) as client2:
            # create identical data on cluster2
            X_2, _, _, _, dX_2, dy_2, _, dg_2 = _create_data(
                objective=task,
                output='array',
                group=None
            )

            model_factory = task_to_dask_factory[task]

            params = {
                "time_out": 5,
                "n_estimators": 1,
                "num_leaves": 2
            }

            # at this point, the result of default_client() is client2 since it was the most recently
            # created. So setting client to client1 here to test that you can select a non-default client
            assert default_client() == client2
            if set_client:
                params.update({"client": client1})

            # unfitted model should survive pickling round trip, and pickling
            # shouldn't have side effects on the model object
            dask_model = model_factory(**params)
            local_model = dask_model.to_local()
            if set_client:
                assert dask_model.client == client1
            else:
                assert dask_model.client is None

            with pytest.raises(lgb.compat.LGBMNotFittedError, match='Cannot access property client_ before calling fit'):
                dask_model.client_

            assert "client" not in local_model.get_params()
            assert getattr(local_model, "client", None) is None

            tmp_file = str(tmp_path / "model-1.pkl")
            _pickle(
                obj=dask_model,
                filepath=tmp_file,
                serializer=serializer
            )
            model_from_disk = _unpickle(
                filepath=tmp_file,
                serializer=serializer
            )

            local_tmp_file = str(tmp_path / "local-model-1.pkl")
            _pickle(
                obj=local_model,
                filepath=local_tmp_file,
                serializer=serializer
            )
            local_model_from_disk = _unpickle(
                filepath=local_tmp_file,
                serializer=serializer
            )

            assert model_from_disk.client is None

            if set_client:
                assert dask_model.client == client1
            else:
                assert dask_model.client is None

            with pytest.raises(lgb.compat.LGBMNotFittedError, match='Cannot access property client_ before calling fit'):
                dask_model.client_

            # client will always be None after unpickling
            if set_client:
                from_disk_params = model_from_disk.get_params()
                from_disk_params.pop("client", None)
                dask_params = dask_model.get_params()
                dask_params.pop("client", None)
                assert from_disk_params == dask_params
            else:
                assert model_from_disk.get_params() == dask_model.get_params()
            assert local_model_from_disk.get_params() == local_model.get_params()

            # fitted model should survive pickling round trip, and pickling
            # shouldn't have side effects on the model object
            if set_client:
                dask_model.fit(dX_1, dy_1, group=dg_1)
            else:
                dask_model.fit(dX_2, dy_2, group=dg_2)
            local_model = dask_model.to_local()

            assert "client" not in local_model.get_params()
            with pytest.raises(AttributeError):
                local_model.client
                local_model.client_

            tmp_file2 = str(tmp_path / "model-2.pkl")
            _pickle(
                obj=dask_model,
                filepath=tmp_file2,
                serializer=serializer
            )
            fitted_model_from_disk = _unpickle(
                filepath=tmp_file2,
                serializer=serializer
            )

            local_tmp_file2 = str(tmp_path / "local-model-2.pkl")
            _pickle(
                obj=local_model,
                filepath=local_tmp_file2,
                serializer=serializer
            )
            local_fitted_model_from_disk = _unpickle(
                filepath=local_tmp_file2,
                serializer=serializer
            )

            if set_client:
                assert dask_model.client == client1
                assert dask_model.client_ == client1
            else:
                assert dask_model.client is None
                assert dask_model.client_ == default_client()
                assert dask_model.client_ == client2

            assert isinstance(fitted_model_from_disk, model_factory)
            assert fitted_model_from_disk.client is None
            assert fitted_model_from_disk.client_ == default_client()
            assert fitted_model_from_disk.client_ == client2

            # client will always be None after unpickling
            if set_client:
                from_disk_params = fitted_model_from_disk.get_params()
                from_disk_params.pop("client", None)
                dask_params = dask_model.get_params()
                dask_params.pop("client", None)
                assert from_disk_params == dask_params
            else:
                assert fitted_model_from_disk.get_params() == dask_model.get_params()
            assert local_fitted_model_from_disk.get_params() == local_model.get_params()

            if set_client:
                preds_orig = dask_model.predict(dX_1).compute()
                preds_loaded_model = fitted_model_from_disk.predict(dX_1).compute()
                preds_orig_local = local_model.predict(X_1)
                preds_loaded_model_local = local_fitted_model_from_disk.predict(X_1)
            else:
                preds_orig = dask_model.predict(dX_2).compute()
                preds_loaded_model = fitted_model_from_disk.predict(dX_2).compute()
                preds_orig_local = local_model.predict(X_2)
                preds_loaded_model_local = local_fitted_model_from_disk.predict(X_2)

            assert_eq(preds_orig, preds_loaded_model)
            assert_eq(preds_orig_local, preds_loaded_model_local)


def test_warns_and_continues_on_unrecognized_tree_learner(client):
    X = da.random.random((1e3, 10))
    y = da.random.random((1e3, 1))
    dask_regressor = lgb.DaskLGBMRegressor(
        client=client,
        time_out=5,
        tree_learner='some-nonsense-value',
        n_estimators=1,
        num_leaves=2
    )
    with pytest.warns(UserWarning, match='Parameter tree_learner set to some-nonsense-value'):
        dask_regressor = dask_regressor.fit(X, y)

    assert dask_regressor.fitted_

    client.close(timeout=CLIENT_CLOSE_TIMEOUT)


@pytest.mark.parametrize('tree_learner', ['data_parallel', 'voting_parallel'])
def test_training_respects_tree_learner_aliases(tree_learner, client):
    task = 'regression'
    _, _, _, _, dX, dy, dw, dg = _create_data(objective=task, output='array')
    dask_factory = task_to_dask_factory[task]
    dask_model = dask_factory(
        client=client,
        tree_learner=tree_learner,
        time_out=5,
        n_estimators=10,
        num_leaves=15
    )
    dask_model.fit(dX, dy, sample_weight=dw, group=dg)

    assert dask_model.fitted_
    assert dask_model.get_params()['tree_learner'] == tree_learner


def test_error_on_feature_parallel_tree_learner(client):
    X = da.random.random((100, 10), chunks=(50, 10))
    y = da.random.random(100, chunks=50)
    dask_regressor = lgb.DaskLGBMRegressor(
        client=client,
        time_out=5,
        tree_learner='feature_parallel',
        n_estimators=1,
        num_leaves=2
    )
    with pytest.raises(lgb.basic.LightGBMError, match='Do not support feature parallel in c api'):
        dask_regressor = dask_regressor.fit(X, y)

    client.close(timeout=CLIENT_CLOSE_TIMEOUT)


@gen_cluster(client=True, timeout=None)
def test_errors(c, s, a, b):
    def f(part):
        raise Exception('foo')

    df = dd.demo.make_timeseries()
    df = df.map_partitions(f, meta=df._meta)
    with pytest.raises(Exception) as info:
        yield lgb.dask._train(
            client=c,
            data=df,
            label=df.x,
            params={},
            model_factory=lgb.LGBMClassifier
        )
        assert 'foo' in str(info.value)


@pytest.mark.parametrize('task', tasks)
@pytest.mark.parametrize('output', data_output)
def test_training_succeeds_even_if_some_workers_do_not_have_any_data(client, task, output):
    if task == 'ranking' and output == 'scipy_csr_matrix':
        pytest.skip('LGBMRanker is not currently tested on sparse matrices')

    def collection_to_single_partition(collection):
        """Merge the parts of a Dask collection into a single partition."""
        if collection is None:
            return
        if isinstance(collection, da.Array):
            return collection.rechunk(*collection.shape)
        return collection.repartition(npartitions=1)

    X, y, w, g, dX, dy, dw, dg = _create_data(
        objective=task,
        output=output,
        group=None
    )

    dask_model_factory = task_to_dask_factory[task]
    local_model_factory = task_to_local_factory[task]

    dX = collection_to_single_partition(dX)
    dy = collection_to_single_partition(dy)
    dw = collection_to_single_partition(dw)
    dg = collection_to_single_partition(dg)

    n_workers = len(client.scheduler_info()['workers'])
    assert n_workers > 1
    assert dX.npartitions == 1

    params = {
        'time_out': 5,
        'random_state': 42,
        'num_leaves': 10
    }

    dask_model = dask_model_factory(tree='data', client=client, **params)
    dask_model.fit(dX, dy, group=dg, sample_weight=dw)
    dask_preds = dask_model.predict(dX).compute()

    local_model = local_model_factory(**params)
    if task == 'ranking':
        local_model.fit(X, y, group=g, sample_weight=w)
    else:
        local_model.fit(X, y, sample_weight=w)
    local_preds = local_model.predict(X)

    assert assert_eq(dask_preds, local_preds)

    client.close(timeout=CLIENT_CLOSE_TIMEOUT)


@pytest.mark.parametrize('task', tasks)
def test_network_params_not_required_but_respected_if_given(client, task, listen_port):
    client.wait_for_workers(2)

    _, _, _, _, dX, dy, _, dg = _create_data(
        objective=task,
        output='array',
        chunk_size=10,
        group=None
    )

    dask_model_factory = task_to_dask_factory[task]

    # rebalance data to be sure that each worker has a piece of the data
    client.rebalance()

    # model 1 - no network parameters given
    dask_model1 = dask_model_factory(
        n_estimators=5,
        num_leaves=5,
    )
    dask_model1.fit(dX, dy, group=dg)
    assert dask_model1.fitted_
    params = dask_model1.get_params()
    assert 'local_listen_port' not in params
    assert 'machines' not in params

    # model 2 - machines given
    n_workers = len(client.scheduler_info()['workers'])
    open_ports = [lgb.dask._find_random_open_port() for _ in range(n_workers)]
    dask_model2 = dask_model_factory(
        n_estimators=5,
        num_leaves=5,
        machines=",".join([
            "127.0.0.1:" + str(port)
            for port in open_ports
        ]),
    )

    dask_model2.fit(dX, dy, group=dg)
    assert dask_model2.fitted_
    params = dask_model2.get_params()
    assert 'local_listen_port' not in params
    assert 'machines' in params

    # model 3 - local_listen_port given
    # training should fail because LightGBM will try to use the same
    # port for multiple worker processes on the same machine
    dask_model3 = dask_model_factory(
        n_estimators=5,
        num_leaves=5,
        local_listen_port=listen_port
    )
    error_msg = "has multiple Dask worker processes running on it"
    with pytest.raises(lgb.basic.LightGBMError, match=error_msg):
        dask_model3.fit(dX, dy, group=dg)

    client.close(timeout=CLIENT_CLOSE_TIMEOUT)


@pytest.mark.parametrize('task', tasks)
def test_machines_should_be_used_if_provided(task):
    with LocalCluster(n_workers=2) as cluster, Client(cluster) as client:
        _, _, _, _, dX, dy, _, dg = _create_data(
            objective=task,
            output='array',
            chunk_size=10,
            group=None
        )

        dask_model_factory = task_to_dask_factory[task]

        # rebalance data to be sure that each worker has a piece of the data
        client.rebalance()

        n_workers = len(client.scheduler_info()['workers'])
        assert n_workers > 1
        open_ports = [lgb.dask._find_random_open_port() for _ in range(n_workers)]
        dask_model = dask_model_factory(
            n_estimators=5,
            num_leaves=5,
            machines=",".join([
                "127.0.0.1:" + str(port)
                for port in open_ports
            ]),
        )

        # test that "machines" is actually respected by creating a socket that uses
        # one of the ports mentioned in "machines"
        error_msg = "Binding port %s failed" % open_ports[0]
        with pytest.raises(lgb.basic.LightGBMError, match=error_msg):
            with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
                s.bind(('127.0.0.1', open_ports[0]))
                dask_model.fit(dX, dy, group=dg)

        # an informative error should be raised if "machines" has duplicates
        one_open_port = lgb.dask._find_random_open_port()
        dask_model.set_params(
            machines=",".join([
                "127.0.0.1:" + str(one_open_port)
                for _ in range(n_workers)
            ])
        )
        with pytest.raises(ValueError, match="Found duplicates in 'machines'"):
            dask_model.fit(dX, dy, group=dg)


@pytest.mark.parametrize(
    "classes",
    [
        (lgb.DaskLGBMClassifier, lgb.LGBMClassifier),
        (lgb.DaskLGBMRegressor, lgb.LGBMRegressor),
        (lgb.DaskLGBMRanker, lgb.LGBMRanker)
    ]
)
def test_dask_classes_and_sklearn_equivalents_have_identical_constructors_except_client_arg(classes):
    dask_spec = inspect.getfullargspec(classes[0])
    sklearn_spec = inspect.getfullargspec(classes[1])
    assert dask_spec.varargs == sklearn_spec.varargs
    assert dask_spec.varkw == sklearn_spec.varkw
    assert dask_spec.kwonlyargs == sklearn_spec.kwonlyargs
    assert dask_spec.kwonlydefaults == sklearn_spec.kwonlydefaults

    # "client" should be the only different, and the final argument
    assert dask_spec.args[:-1] == sklearn_spec.args
    assert dask_spec.defaults[:-1] == sklearn_spec.defaults
    assert dask_spec.args[-1] == 'client'
    assert dask_spec.defaults[-1] is None


@pytest.mark.parametrize(
    "methods",
    [
        (lgb.DaskLGBMClassifier.fit, lgb.LGBMClassifier.fit),
        (lgb.DaskLGBMClassifier.predict, lgb.LGBMClassifier.predict),
        (lgb.DaskLGBMClassifier.predict_proba, lgb.LGBMClassifier.predict_proba),
        (lgb.DaskLGBMRegressor.fit, lgb.LGBMRegressor.fit),
        (lgb.DaskLGBMRegressor.predict, lgb.LGBMRegressor.predict),
        (lgb.DaskLGBMRanker.fit, lgb.LGBMRanker.fit),
        (lgb.DaskLGBMRanker.predict, lgb.LGBMRanker.predict)
    ]
)
def test_dask_methods_and_sklearn_equivalents_have_similar_signatures(methods):
    dask_spec = inspect.getfullargspec(methods[0])
    sklearn_spec = inspect.getfullargspec(methods[1])
    dask_params = inspect.signature(methods[0]).parameters
    sklearn_params = inspect.signature(methods[1]).parameters
    assert dask_spec.args == sklearn_spec.args[:len(dask_spec.args)]
    assert dask_spec.varargs == sklearn_spec.varargs
    if sklearn_spec.varkw:
        assert dask_spec.varkw == sklearn_spec.varkw[:len(dask_spec.varkw)]
    assert dask_spec.kwonlyargs == sklearn_spec.kwonlyargs
    assert dask_spec.kwonlydefaults == sklearn_spec.kwonlydefaults
    for param in dask_spec.args:
        error_msg = f"param '{param}' has different default values in the methods"
        assert dask_params[param].default == sklearn_params[param].default, error_msg


@pytest.mark.parametrize('task', tasks)
def test_training_succeeds_when_data_is_dataframe_and_label_is_column_array(
    task,
    client,
):
    _, _, _, _, dX, dy, dw, dg = _create_data(
        objective=task,
        output='dataframe',
        group=None
    )

    model_factory = task_to_dask_factory[task]

    dy = dy.to_dask_array(lengths=True)
    dy_col_array = dy.reshape(-1, 1)
    assert len(dy_col_array.shape) == 2 and dy_col_array.shape[1] == 1

    params = {
        'n_estimators': 1,
        'num_leaves': 3,
        'random_state': 0,
        'time_out': 5
    }
    model = model_factory(**params)
    model.fit(dX, dy_col_array, sample_weight=dw, group=dg)
    assert model.fitted_

    client.close(timeout=CLIENT_CLOSE_TIMEOUT)


@pytest.mark.parametrize('task', tasks)
@pytest.mark.parametrize('output', data_output)
def test_init_score(task, output, client):
    if task == 'ranking' and output == 'scipy_csr_matrix':
        pytest.skip('LGBMRanker is not currently tested on sparse matrices')

    _, _, _, _, dX, dy, dw, dg = _create_data(
        objective=task,
        output=output,
        group=None
    )

    model_factory = task_to_dask_factory[task]

    params = {
        'n_estimators': 1,
        'num_leaves': 2,
        'time_out': 5
    }
    init_score = random.random()
    # init_scores must be a 1D array, even for multiclass classification
    # where you need to provide 1 score per class for each row in X
    # https://github.com/microsoft/LightGBM/issues/4046
    size_factor = 1
    if task == 'multiclass-classification':
        size_factor = 3  # number of classes

    if output.startswith('dataframe'):
        init_scores = dy.map_partitions(lambda x: pd.Series([init_score] * x.size * size_factor))
    else:
        init_scores = dy.map_blocks(lambda x: np.repeat(init_score, x.size * size_factor))
    model = model_factory(client=client, **params)
    model.fit(dX, dy, sample_weight=dw, init_score=init_scores, group=dg)
    # value of the root node is 0 when init_score is set
    assert model.booster_.trees_to_dataframe()['value'][0] == 0

    client.close(timeout=CLIENT_CLOSE_TIMEOUT)


def sklearn_checks_to_run():
    check_names = [
        "check_estimator_get_tags_default_keys",
        "check_get_params_invariance",
        "check_set_params"
    ]
    for check_name in check_names:
        check_func = getattr(sklearn_checks, check_name, None)
        if check_func:
            yield check_func


def _tested_estimators():
    for Estimator in [lgb.DaskLGBMClassifier, lgb.DaskLGBMRegressor]:
        yield Estimator()


@pytest.mark.parametrize("estimator", _tested_estimators())
@pytest.mark.parametrize("check", sklearn_checks_to_run())
def test_sklearn_integration(estimator, check, client):
    estimator.set_params(local_listen_port=18000, time_out=5)
    name = type(estimator).__name__
    check(name, estimator)
    client.close(timeout=CLIENT_CLOSE_TIMEOUT)


# this test is separate because it takes a not-yet-constructed estimator
@pytest.mark.parametrize("estimator", list(_tested_estimators()))
def test_parameters_default_constructible(estimator):
    name = estimator.__class__.__name__
    if sk_version >= parse_version("0.24"):
        Estimator = estimator
    else:
        Estimator = estimator.__class__
    sklearn_checks.check_parameters_default_constructible(name, Estimator)


@pytest.mark.parametrize('task', tasks)
@pytest.mark.parametrize('output', data_output)
def test_predict_with_raw_score(task, output, client):
    if task == 'ranking' and output == 'scipy_csr_matrix':
        pytest.skip('LGBMRanker is not currently tested on sparse matrices')

    _, _, _, _, dX, dy, _, dg = _create_data(
        objective=task,
        output=output,
        group=None
    )

    model_factory = task_to_dask_factory[task]
    params = {
        'client': client,
        'n_estimators': 1,
        'num_leaves': 2,
        'time_out': 5,
        'min_sum_hessian': 0
    }
    model = model_factory(**params)
    model.fit(dX, dy, group=dg)
    raw_predictions = model.predict(dX, raw_score=True).compute()

    trees_df = model.booster_.trees_to_dataframe()
    leaves_df = trees_df[trees_df.node_depth == 2]
    if task == 'multiclass-classification':
        for i in range(model.n_classes_):
            class_df = leaves_df[leaves_df.tree_index == i]
            assert set(raw_predictions[:, i]) == set(class_df['value'])
    else:
        assert set(raw_predictions) == set(leaves_df['value'])

    if task.endswith('classification'):
        pred_proba_raw = model.predict_proba(dX, raw_score=True).compute()
        assert_eq(raw_predictions, pred_proba_raw)

    client.close(timeout=CLIENT_CLOSE_TIMEOUT)