experiments.py

"""Exeperiments of XGBDistribution vs NGBRegressor and XGBRegressor
"""

import glob
import logging
import os
import shutil
import sys
import tempfile
import time
import zipfile
from argparse import ArgumentParser
from dataclasses import dataclass
from datetime import datetime
from functools import partial, wraps
from pathlib import Path

import numpy as np
import pandas as pd
import requests
import sqlalchemy as sa
from ngboost import NGBRegressor
from scipy.stats import norm
from sklearn.metrics import mean_squared_error
from sklearn.model_selection import KFold, train_test_split
from xgboost import XGBRegressor

from xgboost_distribution import XGBDistribution

_logger = logging.getLogger(__name__)


# -------------------------------------------------------------------------------------
# Datasets defintions
# -------------------------------------------------------------------------------------

DATASETS = {}  # each instantiated dataset will be stored here, keyed by name
DATA_DIR = Path(__file__).parent.parent.absolute().joinpath("data")


@dataclass
class Dataset:
    name: str
    url: str
    file_to_unpack: str = None  # if url refers to zip, we mark the file to unpack
    load_func: callable = pd.read_csv
    processing_func: callable = None

    def __post_init__(self):
        DATASETS.update({self.name: self})


Dataset(
    name="housing",
    url="https://archive.ics.uci.edu/ml/machine-learning-databases/housing/housing.data",  # noqa: E501
    load_func=partial(pd.read_csv, header=None, delim_whitespace=True),
)

Dataset(
    name="concrete",
    url="https://archive.ics.uci.edu/ml/machine-learning-databases/concrete/compressive/Concrete_Data.xls",  # noqa: E501
    load_func=pd.read_excel,
)

Dataset(
    name="energy",
    url="https://archive.ics.uci.edu/ml/machine-learning-databases/00242/ENB2012_data.xlsx",  # noqa: E501
    load_func=partial(pd.read_excel, usecols=[f"X{i}" for i in range(1, 9)] + ["Y1"]),
)

Dataset(
    name="naval",
    url="https://archive.ics.uci.edu/ml/machine-learning-databases/00316/UCI%20CBM%20Dataset.zip",  # noqa: E501
    file_to_unpack="data.txt",
    load_func=partial(pd.read_csv, header=None, delim_whitespace=True),
    processing_func=lambda x: x.iloc[:, :-1],
)

Dataset(
    name="power",
    url="https://archive.ics.uci.edu/ml/machine-learning-databases/00294/CCPP.zip",
    file_to_unpack="Folds5x2_pp.xlsx",
    load_func=pd.read_excel,
)

Dataset(
    name="protein",
    url="https://archive.ics.uci.edu/ml/machine-learning-databases/00265/CASP.csv",
    processing_func=lambda x: x.loc[
        :, ["F1", "F2", "F3", "F4", "F5", "F6", "F7", "F8", "F9", "RMSD"]
    ],
)

Dataset(
    name="wine",
    url="https://archive.ics.uci.edu/ml/machine-learning-databases/wine-quality/winequality-red.csv",  # noqa: E501
    load_func=partial(pd.read_csv, delimiter=";"),
)

Dataset(
    name="yacht",
    url="http://archive.ics.uci.edu/ml/machine-learning-databases/00243/yacht_hydrodynamics.data",  # noqa: E501
    load_func=partial(pd.read_csv, header=None, delim_whitespace=True),
)

Dataset(
    name="msd",
    url="https://archive.ics.uci.edu/ml/machine-learning-databases/00203/YearPredictionMSD.txt.zip",  # noqa: E501
    file_to_unpack="YearPredictionMSD.txt",
    load_func=pd.read_csv,
    processing_func=lambda x: x.iloc[:, ::-1],
)

# -------------------------------------------------------------------------------------
# Datasets loading functions
# -------------------------------------------------------------------------------------


def load_dataset(name, data_dir=DATA_DIR):
    dataset = DATASETS[name]
    local_path = f"{data_dir}/{name}.data"

    if os.path.exists(local_path):
        _logger.info(f"Loading dataset from local {local_path}...")
        return load_local_dataset(dataset, local_path)
    else:
        _logger.info("Dataset not locally cached, downloading from url...")
        download_dataset(dataset, local_path)
        return load_dataset(name)


def load_local_dataset(dataset, local_path):
    df = dataset.load_func(local_path)
    if dataset.processing_func:
        df = dataset.processing_func(df)
    return df.iloc[:, :-1].values, df.iloc[:, -1].values


def download_dataset(dataset, local_path):
    r = requests.get(dataset.url)

    if dataset.file_to_unpack is not None:
        with tempfile.TemporaryFile() as fp:
            fp.write(r.content)
            unpack_file_from_zip(fp, dataset.file_to_unpack, local_path)
    else:
        with open(local_path, "wb") as f:
            f.write(r.content)
    _logger.info(f"Downloaded {dataset.name} to {local_path}")


def unpack_file_from_zip(zip_file, to_unpack, path):
    with zipfile.ZipFile(zip_file, "r") as inzipfile:
        with tempfile.TemporaryDirectory() as temp_dir:
            inzipfile.extractall(path=temp_dir)
            data_file = glob.glob(f"{temp_dir}/**/{to_unpack}", recursive=True)[0]
            shutil.copyfile(data_file, path)


# -------------------------------------------------------------------------------------
# Metrics to evaluate
# -------------------------------------------------------------------------------------


def root_mean_squared_error(y_pred, y_test):
    return np.sqrt(mean_squared_error(y_pred, y_test))


def normal_nll(loc, scale, y_test):
    return -norm.logpdf(y_test.flatten(), loc=loc, scale=scale).mean()


@dataclass
class EvalResult:
    total_time: float
    rmse: float
    nll: float


# -------------------------------------------------------------------------------------
# Evaluate decorator, which evaluates the pattern function(split_data) -> predictions
# -------------------------------------------------------------------------------------


@dataclass
class SplitData:
    X_train: np.ndarray
    y_train: np.ndarray
    X_val: np.ndarray
    y_val: np.ndarray
    X_test: np.ndarray
    y_test: np.ndarray


evaluations = []


def evaluate(evaluation_func):
    @wraps(evaluation_func)
    def measured(data):
        t0 = time.perf_counter()
        preds = evaluation_func(data)
        elapsed = time.perf_counter() - t0

        if isinstance(preds, np.ndarray):
            rmse = root_mean_squared_error(preds, data.y_test)
            nll = None
        else:
            rmse = root_mean_squared_error(preds.loc, data.y_test)
            nll = normal_nll(preds.loc, preds.scale, data.y_test)
        return EvalResult(rmse=rmse, nll=nll, total_time=elapsed)

    evaluations.append(measured)
    return measured


# -------------------------------------------------------------------------------------
# Model functions, everything decorated by @evaluate will be evaluated
# -------------------------------------------------------------------------------------


@evaluate
def ngb_regressor(data):
    ngb = NGBRegressor(verbose=False)
    ngb.fit(
        data.X_train,
        data.y_train,
        X_val=data.X_val,
        Y_val=data.y_val,
        early_stopping_rounds=10,
    )
    return ngb.pred_dist(data.X_test, max_iter=ngb.best_val_loss_itr)


@evaluate
def xgb_distribution(data):
    xgbd = XGBDistribution(max_depth=3, natural_gradient=True, n_estimators=500)
    xgbd.fit(
        data.X_train,
        data.y_train,
        eval_set=[(data.X_val, data.y_val)],
        early_stopping_rounds=10,
        verbose=False,
    )
    return xgbd.predict(data.X_test)


@evaluate
def xgb_regressor(data):
    xgb = XGBRegressor(max_depth=3, n_estimators=500)
    xgb.fit(
        data.X_train,
        data.y_train,
        eval_set=[(data.X_val, data.y_val)],
        early_stopping_rounds=10,
        verbose=False,
    )
    return xgb.predict(data.X_test)


# -----------------------------------------------------------------------------
# Functions for running cross-validation becnhmark experiment
# -----------------------------------------------------------------------------


def parse_args():
    argparser = ArgumentParser()
    argparser.add_argument("--data-dir", type=str, default=DATA_DIR)
    argparser.add_argument("--dataset", type=str, default="concrete")
    argparser.add_argument("--random-seed", type=int, default=42)
    argparser.add_argument("--n-folds", type=int, default=10)
    argparser.add_argument("--db-name", type=str, default="results.db")
    return argparser.parse_args()


def run_experiment():
    args = parse_args()
    setup_logging()

    _logger.info(f"Storing datasets and results in {args.data_dir}")
    os.makedirs(args.data_dir, exist_ok=True)

    db = DataBase(data_dir=args.data_dir, db_name=args.db_name)
    _logger.info("Connected to database")

    X, y = load_dataset(args.dataset, data_dir=args.data_dir)
    _logger.info(f"Loaded dataset: `{args.dataset}`, X.shape={X.shape}")

    _logger.info(f"Setting random seed to {args.random_seed}")
    np.random.seed(args.random_seed)

    kf = KFold(n_splits=args.n_folds, shuffle=True, random_state=args.random_seed)
    _logger.info(f"Cross-validation with {args.n_folds} folds...")
    results = _cross_validate_evaluations(kf, X, y, args)

    _logger.info("Inserting results into database...")
    for eval_func, result in zip(evaluations, results):
        df_metrics = aggregate_results(result)
        df_metrics["dataset"] = args.dataset
        df_metrics["model"] = eval_func.__name__
        db.insert_metrics_pdf(df_metrics)

    df_metrics = db.get_metrics_pdf()
    df_summary = summarize_metrics(df_metrics)
    _logger.info(f"\n{df_summary}")


def _cross_validate_evaluations(kfold, X, y, args):
    results = [list() for eval in evaluations]

    for ii, (train_index, test_index) in enumerate(kfold.split(X)):
        _logger.info(f"Fold {ii+1} / {args.n_folds}...")

        X_train, X_test = X[train_index], X[test_index]
        y_train, y_test = y[train_index], y[test_index]
        X_train, X_val, y_train, y_val = train_test_split(
            X_train, y_train, test_size=0.1, random_state=args.random_seed
        )
        split_data = SplitData(
            X_train=X_train,
            y_train=y_train,
            X_val=X_val,
            y_val=y_val,
            X_test=X_test,
            y_test=y_test,
        )
        for eval_func, result in zip(evaluations, results):
            result.append(eval_func(split_data))
    return results


def aggregate_results(results):
    return (
        pd.DataFrame(results)
        .agg(["mean", "std"], axis=0)
        .melt(ignore_index=False, var_name="metric")
        .rename_axis("agg_func")
        .reset_index()
    )


def summarize_metrics(df_metrics):
    return df_metrics.pivot_table(
        values=["value"],
        index=["dataset", "agg_func"],
        columns=["model", "metric"],
    )


# -------------------------------------------------------------------------------------
# Database for storing results of each experiment run
# -------------------------------------------------------------------------------------


class DataBase:
    def __init__(self, data_dir=DATA_DIR, db_name="results.db"):
        _logger.info(f"Using SQLite database at {data_dir}/{db_name}")
        self.metadata = sa.MetaData()
        self.engine = sa.create_engine(f"sqlite:///{data_dir}/{db_name}")
        self.connection = self.engine.connect()
        self._create_tables()

    def _create_tables(self):
        self.metrics = sa.Table(
            "metrics",
            self.metadata,
            sa.Column("dataset", sa.String(), nullable=False),
            sa.Column("model", sa.String(), nullable=False),
            sa.Column("metric", sa.String(), nullable=False),
            sa.Column("agg_func", sa.String(), nullable=False),
            sa.Column("value", sa.Float(), nullable=True),
            sa.Column("created_on", sa.DateTime(), default=datetime.now),
        )
        self.metadata.create_all(self.engine)

    def insert_records(self, records, table="metrics"):
        ins = getattr(self, table).insert()
        self.connection.execute(ins, records)

    def insert_metrics_pdf(self, df):
        self.insert_records(df.to_dict("records"), table="metrics")

    def get_metrics_pdf(self):
        return pd.read_sql(
            sa.select(self.metrics).order_by(self.metrics.c.created_on.desc()),
            self.connection,
        )


def setup_logging(loglevel=logging.INFO):
    logformat = "[%(asctime)s] %(levelname)s:%(name)s:%(message)s"
    logging.basicConfig(
        level=loglevel, stream=sys.stdout, format=logformat, datefmt="%Y-%m-%d %H:%M:%S"
    )


if __name__ == "__main__":
    run_experiment()