scripts/11.2_evaluate_multilabel.py

# 11.2_evaluate_multilabel.py

import os
import json
import torch
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt

from pathlib import Path
from datasets import load_from_disk
from transformers import AutoModelForSequenceClassification, AutoTokenizer, Trainer
from torch.utils.data import default_collate
from sklearn.metrics import classification_report, multilabel_confusion_matrix

# === Ścieżki
MODEL_DIR = Path("models/multilabel/")
DATASET_DIR = Path("data/processed/dataset_multilabel_top30")
TOP_RULES_PATH = Path("data/metadata/top_rules.json")

OUT_DIR = MODEL_DIR
REPORT_CSV = OUT_DIR / "classification_report.csv"
REPORT_JSON = OUT_DIR / "metrics.json"
CONF_MATRIX_PNG = OUT_DIR / "confusion_matrix_multilabel.png"

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

# === Wczytanie top_rules
with open(TOP_RULES_PATH) as f:
    top_rules = json.load(f)

# === Wczytaj model + tokenizer
print("📂 Wczytywanie modelu...")
model = AutoModelForSequenceClassification.from_pretrained(MODEL_DIR)

try:
    tokenizer = AutoTokenizer.from_pretrained(MODEL_DIR)
except:
    print("⚠️ Brak tokenizera w modelu — pobieram z microsoft/codebert-base")
    tokenizer = AutoTokenizer.from_pretrained("microsoft/codebert-base")
    tokenizer.save_pretrained(MODEL_DIR)

# === Wczytaj dane i stwórz Trainer
dataset = load_from_disk(str(DATASET_DIR))
trainer = Trainer(model=model, data_collator=collate_fn)

# === Predykcja
print("🔍 Predykcja na zbiorze testowym...")
predictions = trainer.predict(dataset["test"].with_format("torch"))
probs = torch.sigmoid(torch.tensor(predictions.predictions)).numpy()
y_pred = (probs > 0.5).astype(int)
y_true = predictions.label_ids

# === Raport klasyfikacji
print("📊 Raport klasyfikacji:")
report_dict = classification_report(
    y_true,
    y_pred,
    target_names=top_rules,
    zero_division=0,
    output_dict=True
)
report_text = classification_report(y_true, y_pred, target_names=top_rules, zero_division=0)
print(report_text)

# === Zapis raportów
pd.DataFrame(report_dict).transpose().to_csv(REPORT_CSV)
with open(REPORT_JSON, "w") as f:
    json.dump(report_dict, f, indent=2)

print(f"💾 Zapisano raport CSV: {REPORT_CSV}")
print(f"💾 Zapisano metryki JSON: {REPORT_JSON}")

# === Macierz błędów (sumaryczna)
print("🧱 Generuję multilabel confusion matrix...")
mcm = multilabel_confusion_matrix(y_true, y_pred)
support = y_true.sum(axis=0).astype(int)

fig, ax = plt.subplots(figsize=(12, 8))
bars = plt.barh(range(len(top_rules)), support)
plt.yticks(range(len(top_rules)), top_rules)
plt.xlabel("Liczba wystąpień w zbiorze testowym")
plt.title("🔢 Rozkład występowania reguł w testowym zbiorze")

for i, bar in enumerate(bars):
    width = bar.get_width()
    plt.text(width + 1, bar.get_y() + bar.get_height() / 2, str(support[i]), va='center')

plt.tight_layout()
plt.savefig(CONF_MATRIX_PNG)
plt.close()
print(f"🖼️ Zapisano confusion matrix jako PNG: {CONF_MATRIX_PNG}")