train_indobert.py

"""
train_indobert.py — Fine-tuning IndoBERT untuk klasifikasi sentimen.
Jalankan secara lokal dengan GPU, BUKAN saat startup app.
"""

if __name__ == "__main__":
    import pandas as pd
    from datasets import Dataset
    from transformers import (
        AutoTokenizer,
        AutoModelForSequenceClassification,
        Trainer,
        TrainingArguments,
        EarlyStoppingCallback,
    )

    MODEL_NAME  = "indobenchmark/indobert-base-p1"
    OUTPUT_DIR  = "model/final_model"
    DATA_PATH   = "data/train.csv"   # kolom: text, label (0/1/2)

    # ── LOAD DATA ──
    df      = pd.read_csv(DATA_PATH)
    dataset = Dataset.from_pandas(df)

    # ── TOKENIZER ──
    tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)

    def tokenize(batch):
        return tokenizer(batch["text"], truncation=True, padding=True, max_length=128)

    dataset = dataset.map(tokenize, batched=True)
    dataset = dataset.rename_column("label", "labels")
    dataset.set_format("torch", columns=["input_ids", "attention_mask", "labels"])

    # ── SPLIT ──
    split   = dataset.train_test_split(test_size=0.1, seed=42)
    train_d = split["train"]
    eval_d  = split["test"]

    # ── MODEL ──
    num_labels = df["label"].nunique()
    model = AutoModelForSequenceClassification.from_pretrained(
        MODEL_NAME, num_labels=num_labels
    )

    # ── TRAINING ARGS ──
    args = TrainingArguments(
        output_dir=OUTPUT_DIR,
        per_device_train_batch_size=16,
        per_device_eval_batch_size=16,
        num_train_epochs=5,
        evaluation_strategy="epoch",
        save_strategy="epoch",
        load_best_model_at_end=True,
        metric_for_best_model="eval_loss",
        logging_steps=20,
        fp16=True,              # matikan jika CPU-only
        report_to="none",
    )

    trainer = Trainer(
        model=model,
        args=args,
        train_dataset=train_d,
        eval_dataset=eval_d,
        callbacks=[EarlyStoppingCallback(early_stopping_patience=2)],
    )

    trainer.train()

    # ── SAVE ──
    model.save_pretrained(OUTPUT_DIR)
    tokenizer.save_pretrained(OUTPUT_DIR)
    print(f"✅ Model saved: {OUTPUT_DIR}")