nl2ml/models_scripts/boosting_train.py

import argparse
import os
import pickle
import sys

from catboost import CatBoostClassifier
import numpy as np
import optuna
import pandas as pd
from sklearn.metrics import accuracy_score, f1_score
from sklearn.model_selection import KFold

from common.tools import *


parser = argparse.ArgumentParser()
parser.add_argument("GRAPH_VER", help="version of the graph you want regex to label your CSV with", type=str)
parser.add_argument("DATASET_PATH", help="path to your input CSV", type=str)
args = parser.parse_args()

GRAPH_VER = args.GRAPH_VER
DATASET_PATH = args.DATASET_PATH

MODEL_DIR = "../models/catboost_regex_graph_v{}.sav".format(GRAPH_VER)
TFIDF_DIR = "../models/tfidf_catboost_graph_v{}.pickle".format(GRAPH_VER)

TAGS_TO_PREDICT = get_graph_vertices(GRAPH_VER)

EXPERIMENT_DATA_PATH = ".."
CODE_COLUMN = "code_block"
TARGET_COLUMN = "graph_vertex_id"

RANDOM_STATE = 42
N_TRIALS = 30

HYPERPARAM_SPACE = {
    "tfidf_min_df": (1, 15),
    "tfidf_max_df": (0.2, 0.95),
    "catboost_lr": (0.1, 1.0),
    "catboost_iter": (100, 1500),
    "catboost_depth": (3, 6),
}


def cross_val_scores(kf, clf, X, y):
    f1s = []
    accuracies = []
    for i, (train_index, test_index) in enumerate(kf.split(X)):
        X_train, X_test = X[train_index], X[test_index]
        y_train, y_test = y[train_index], y[test_index]

        clf.fit(X_train, y_train)

        y_pred = clf.predict(X_test)
        f1s.append(f1_score(y_test, y_pred, average="weighted"))
        accuracies.append(accuracy_score(y_test, y_pred))

    f1s = np.array(f1s)
    accuracies = np.array(accuracies)
    return f1s.mean(), accuracies.mean()


class Objective:
    def __init__(self, df, kfold_params, tfidf_min_df, tfidf_max_df, catboost_lr, catboost_iter, catboost_depth):
        self.kf = KFold(**kfold_params)
        self.df = df

        self.min_df_range = tfidf_min_df
        self.max_df_range = tfidf_max_df

        self.lr_range = catboost_lr
        self.max_iter_range = catboost_iter
        self.depth_range = catboost_depth

    def __call__(self, trial):
        tfidf_params = {
            "min_df": trial.suggest_int("tfidf__min_df", *self.min_df_range),
            "max_df": trial.suggest_loguniform("tfidf__max_df", *self.max_df_range),
            "smooth_idf": True,
        }
        code_blocks_tfidf = tfidf_fit_transform(self.df[CODE_COLUMN], tfidf_params)
        X, y = code_blocks_tfidf, self.df[TARGET_COLUMN].values

        boosting_params = {
            "learning_rate": trial.suggest_loguniform("boosting__learning_rate", *self.lr_range),
            "iterations": trial.suggest_int("boosting__iterations", *self.max_iter_range),
            "depth": trial.suggest_int("boosting__depth", *self.depth_range),
            "random_seed": RANDOM_STATE,
            "verbose": False,
        }
        clf = CatBoostClassifier(**boosting_params)

        f1_mean, _ = cross_val_scores(self.kf, clf, X, y)
        return f1_mean


def select_hyperparams(df, kfold_params, tfidf_path, model_path):
    """
    Uses optuna to find hyperparams that maximize F1 score
    :param df: labelled dataset
    :param kfold_params: parameters for sklearn's KFold
    :param tfidf_dir: where to save trained tf-idf
    :return: dict with parameters and metrics
    """
    study = optuna.create_study(direction="maximize", study_name="boosting")
    objective = Objective(df, kfold_params, **HYPERPARAM_SPACE)

    study.optimize(objective, n_trials=N_TRIALS)

    best_tfidf_params = {
        "smooth_idf": True,
    }
    best_boosting_params = {
        "random_seed": RANDOM_STATE,
        "verbose": False,
    }
    for key, value in study.best_params.items():
        model_name, param_name = key.split("__")
        if model_name == "tfidf":
            best_tfidf_params[param_name] = value
        elif model_name == "boosting":
            best_boosting_params[param_name] = value

    code_blocks_tfidf = tfidf_fit_transform(df[CODE_COLUMN], best_tfidf_params, tfidf_path)
    X, y = code_blocks_tfidf, df[TARGET_COLUMN].values
    clf = CatBoostClassifier(**best_boosting_params)

    f1_mean, accuracy_mean = cross_val_scores(objective.kf, clf, X, y)

    clf.fit(X, y)
    pickle.dump(clf, open(model_path, "wb"))

    metrics = dict(test_f1_score=f1_mean, test_accuracy=accuracy_mean)

    return best_tfidf_params, best_boosting_params, metrics


if __name__ == "__main__":
    df = load_data(DATASET_PATH)

    print(df.columns)
    nrows = df.shape[0]
    print("loaded")

    kfold_params = {
        "n_splits": 3,
        "random_state": RANDOM_STATE,
        "shuffle": True,
    }
    data_meta = {
        "DATASET_PATH": DATASET_PATH,
        "nrows": nrows,
        "label": TAGS_TO_PREDICT,
        "model": MODEL_DIR,
        "script_dir": __file__,
    }

    print("selecting hyperparameters")
    tfidf_params, boosting_params, metrics = select_hyperparams(df, kfold_params, TFIDF_DIR, MODEL_DIR)
    print("hyperparams:", "\nboosting", boosting_params, "\ntfidf", tfidf_params)
    print("metrics:", metrics)
    print("finished")