scripts/10.2_train_multilabel_model.py

# scripts/10.2_train_multilabel_model.py

import os
import json
import numpy as np
from transformers import (
    AutoTokenizer,
    AutoModelForSequenceClassification,
    TrainingArguments,
    Trainer,
    EarlyStoppingCallback,
)
from datasets import load_from_disk
from torch.utils.data import default_collate
from sklearn.metrics import accuracy_score, f1_score, precision_score, recall_score

# === Konfiguracja
DATA_PATH = "data/processed/dataset_multilabel_top30"
OUTPUT_DIR = "models/multilabel"
MODEL_NAME = "microsoft/codebert-base"
NUM_LABELS = 30
NUM_EPOCHS = 12
SEED = 42

# === Ładowanie danych i tokenizera
print("📂 Ładowanie danych i tokenizera...")
ds = load_from_disk(DATA_PATH)
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)

# === Model
print("🧠 Inicjalizacja modelu...")
model = AutoModelForSequenceClassification.from_pretrained(
    MODEL_NAME,
    num_labels=NUM_LABELS,
    problem_type="multi_label_classification"
)

# === Funkcja metryk
def compute_metrics(pred):
    logits, labels = pred
    probs = 1 / (1 + np.exp(-logits))  # sigmoid
    preds = (probs > 0.5).astype(int)
    return {
        "accuracy": accuracy_score(labels, preds),
        "f1": f1_score(labels, preds, average="micro"),
        "precision": precision_score(labels, preds, average="micro"),
        "recall": recall_score(labels, preds, average="micro"),
    }

# === Batch collator: wymuszenie float32
def collate_fn(batch):
    batch = default_collate(batch)
    batch["labels"] = batch["labels"].float()
    return batch

# === Argumenty treningowe
args = TrainingArguments(
    output_dir=OUTPUT_DIR,
    evaluation_strategy="epoch",
    save_strategy="epoch",
    learning_rate=2e-5,
    per_device_train_batch_size=8,
    per_device_eval_batch_size=8,
    num_train_epochs=NUM_EPOCHS,
    weight_decay=0.01,
    load_best_model_at_end=True,
    save_total_limit=2,
    seed=SEED,
    logging_dir=os.path.join(OUTPUT_DIR, "logs"),
    logging_steps=50,
    metric_for_best_model="f1",
    greater_is_better=True,
    report_to="none"
)

# === Trener
trainer = Trainer(
    model=model,
    args=args,
    train_dataset=ds["train"].with_format("torch"),
    eval_dataset=ds["validation"].with_format("torch"),
    tokenizer=tokenizer,
    compute_metrics=compute_metrics,
    callbacks=[EarlyStoppingCallback(early_stopping_patience=2)],
    data_collator=collate_fn,
)

# === Trening
print("🚀 Start treningu...")
trainer.train()

# === Zapis modelu i logów
print("💾 Zapisuję model i logi...")
trainer.save_model(OUTPUT_DIR)

log_path = os.path.join(OUTPUT_DIR, "training_log.json")
with open(log_path, "w", encoding="utf-8") as f:
    json.dump(trainer.state.log_history, f, indent=2)

print(f"📝 Zapisano log treningu do {log_path}")
print("✅ Gotowe.")