codecommentclassification/modeling/train.py

"""Module for training different types of models for code comment classification."""

import argparse
import logging
import os

import dagshub
from datasets import Dataset
import mlflow
import yaml

from .utils import load_dataset_splits, parse_labels_column

logging.basicConfig(
    level=logging.INFO,
    format="%(asctime)s - %(name)s - %(levelname)s - %(message)s",
)
logger = logging.getLogger(__name__)


dagshub.init(repo_owner="se4ai2526-uniba", repo_name="TheClouds", mlflow=True)


def train_model(lang, model_type, data_path, model_output_path, params):
    """Trains and saves a model for a specific language and model type."""
    print(f"--- Starting training for language: {lang} with model: {model_type} ---")

    ds = load_dataset_splits(data_path)

    train_df = ds[f"{lang}_train"]
    eval_df = ds[f"{lang}_test"]

    train_df = parse_labels_column(train_df)
    eval_df = parse_labels_column(eval_df)

    # converto i DataFrame in HuggingFace Dataset
    train_dataset = Dataset.from_pandas(train_df, preserve_index=False)
    eval_dataset = Dataset.from_pandas(eval_df, preserve_index=False)

    if model_type == "setfit":
        from setfit import SetFitModel, Trainer, TrainingArguments

        mlflow.set_experiment("SetFit Training")
        with mlflow.start_run(run_name=f"train-{lang}-{model_type}"):
            mlflow.log_param("language", lang)
            mlflow.log_param("model_type", model_type)
            model = SetFitModel.from_pretrained(
                "sentence-transformers/paraphrase-MiniLM-L6-v2",
                multi_target_strategy="multi-output",
            )
            args = TrainingArguments(**params)
            trainer = Trainer(
                model=model,
                args=args,
                train_dataset=train_dataset,
                eval_dataset=eval_dataset,
                column_mapping={"combo": "text", "labels": "label"},
            )

            mlflow.log_param("num_epochs", args.num_epochs)
            mlflow.log_param("num_iterations", args.num_iterations)

            trainer.train()

            eval_metrics = trainer.evaluate()
            for metric_name, metric_value in eval_metrics.items():
                mlflow.log_metric(metric_name, metric_value)

            trainer.model.save_pretrained(model_output_path)

            mlflow.transformers.log_model(
                transformers_model=model_output_path,
                artifact_path=f"{lang}_setfit_model",
                task="text-classification",
            )
            mlflow.end_run()

    elif model_type == "random_forest":
        import joblib
        import numpy as np
        from sklearn.ensemble import RandomForestClassifier
        from sklearn.feature_extraction.text import TfidfVectorizer
        from sklearn.multioutput import MultiOutputClassifier
        from sklearn.pipeline import Pipeline

        mlflow.set_experiment("Random Forest Training")
        with mlflow.start_run(run_name=f"train-{lang}-{model_type}"):
            mlflow.log_param("language", lang)
            mlflow.log_param("model_type", model_type)
            mlflow.log_params(params)

            tfidf_params = {
                "ngram_range": tuple(params.pop("ngram_range", (1, 1))),
                "max_features": params.pop("max_features", None),
                "min_df": params.pop("min_df", 1),
                "max_df": params.pop("max_df", 1.0),
            }

            rf_params = params
            pipeline = Pipeline(
                [
                    ("tfidf", TfidfVectorizer(**tfidf_params)),
                    (
                        "clf",
                        MultiOutputClassifier(
                            RandomForestClassifier(
                                random_state=42, class_weight="balanced", **rf_params
                            )
                        ),
                    ),
                ]
            )

            X_train = train_dataset["combo"]
            y_train = np.array(train_dataset["labels"])

            pipeline.fit(X_train, y_train)

            X_test = eval_dataset["combo"]
            y_test = np.array(eval_dataset["labels"])

            score = pipeline.score(X_test, y_test)
            mlflow.log_metric("accuracy", score)

            os.makedirs(os.path.dirname(model_output_path), exist_ok=True)
            joblib.dump(pipeline, f"{model_output_path}.joblib")

            mlflow.sklearn.log_model(
                sk_model=pipeline, artifact_path=f"{lang}_random_forest_model"
            )
            mlflow.end_run()

    elif model_type == "transformer":
        from .transformer import (
            TransformerConfig,
            TransformerTrainer,
        )

        mlflow.set_experiment("Transformer Training")
        with mlflow.start_run(run_name=f"train-{lang}-{model_type}"):
            mlflow.log_param("language", lang)
            mlflow.log_param("model_type", model_type)
            mlflow.log_params(params)

            cfg = TransformerConfig(
                lang=lang,
                raw_data_dir="data/raw",
                processed_data_dir="data/processed/transformer",
                model_output_path=model_output_path,
                pretrained_model_name=params.get(
                    "pretrained_model_name", "microsoft/codebert-base"
                ),
                max_length=params.get("max_length", 128),
                batch_size=params.get("batch_size", 16),
                lr=params.get("lr", 2e-5),
                num_epochs=params.get("num_epochs", 5),
                warmup_ratio=params.get("warmup_ratio", 0.1),
                pos_weight_cap=params.get("pos_weight_cap", 30.0),
                threshold=params.get("threshold", 0.5),
                preprocessing=params.get("preprocessing", False),
                preprocessing_factor=params.get("preprocessing_factor", 1.0),
            )

            logger.info(
                "Starting transformer training for language '%s' with config: %s",
                lang,
                cfg,
            )

            trainer = TransformerTrainer(cfg)
            metrics = trainer.run()

            logger.info("Final transformer metrics for %s: %s", lang, metrics)

            for name, value in metrics.items():
                mlflow.log_metric(f"final_{name}", value)

            mlflow.end_run()

    else:
        raise ValueError(f"Unsupported model_type: {model_type}")

    print(f"Model for {lang}-{model_type} saved to {model_output_path}")


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--lang", type=str, required=True)
    parser.add_argument("--model_type", type=str, required=True)
    args = parser.parse_args()

    with open("params.yaml", "r") as f:
        all_params = yaml.safe_load(f)

    model_params = all_params[args.model_type].copy()

    train_model(
        lang=args.lang,
        model_type=args.model_type,
        data_path="data/raw",
        model_output_path=f"models/{args.lang}/{args.model_type}",
        params=model_params,
    )