from transformers import pipeline

label2id = {
    'Task_Solution': 0,
    'Creative_Generation': 1,
    'Knowledge_Explanation': 2,
    'Analytical_Reasoning': 3,
    'Information_Extraction': 4,
    'Step_by_Step_Calculation': 5,
    'Role_Play_Response': 6,
    'Opinion_Perspective': 7
}

def preprocess_text_classification(examples: dict[str, list]) -> BatchEncoding:
    """バッチ処理用に修正"""
    encoded_examples = tokenizer(
        examples["questions"],  # バッチ処理なのでリストで渡される
        max_length=512,
        padding=True,
        truncation=True,
        return_tensors=None  # バッチ処理時はNoneを指定
    )
    
    # ラベルをバッチで数値に変換
    encoded_examples["labels"] = [label2id[label] for label in examples["labels"]]
    return encoded_examples

# 使用するデータセット
test_data = test_data.to_pandas()
test_data["labels"] = test_data["labels"].apply(lambda x: label2id[x])
test_data

model_name = "hiroki-rad/bert-base-classification-ft"
classify_pipe = pipeline(model=model_name, device="cuda:0")

class_label = dataset["labels"].unique()
label2id = {label: id for id, label in enumerate(class_label)}
id2label = {id: label for id, label in enumerate(class_label)}

results: list[dict[str, float | str]] = []
for i, example in tqdm(enumerate(test_data.itertuples())):
    # モデルの予測結果を取得
    model_prediction = classify_pipe(example.questions)[0]
    # 正解のラベルIDをラベル名に変換
    true_label = id2label[example.labels]
    results.append(
        {
            "example_id": i,
            "pred_prob": model_prediction["score"],
            "pred_label": model_prediction["label"],
            "true_label": true_label,
        }
    )