最近在复现 LineVul 这篇文章的实验，本文主要用于简化文章复现流程和记录复现过程中出现的问题。

1 安装依赖环境

pip install gdown
pip install transformers
pip install captum
pip install torch torchvision torchaudio
pip install numpy
pip install tqdm
pip install scikit-learn
pip install pandas
pip install tokenizers

2 下载数据

#### 下载数据集 #### (需要科学上网)
cd data/big-vul_dataset
gdown https://drive.google.com/uc?id=1h0iFJbc5DGXCXXvvR6dru_Dms_b2zW4V
gdown https://drive.google.com/uc?id=1ldXyFvHG41VMrm260cK_JEPYqeb6e6Yw
gdown https://drive.google.com/uc?id=1yggncqivMcP0tzbh8-8Eu02Edwcs44WZ
cd ../..
#### 下载模型 ####
cd linevul/saved_models/checkpoint-best-f1
gdown https://drive.google.com/uc?id=1oodyQqRb9jEcvLMVVKILmu8qHyNwd-zH
cd ../../..
#### 下载 codebert-base ####
cd linevul
mkdir microsoft
cd microsoft
git lfs install
git clone https://huggingface.co/microsoft/codebert-base

3 实验复现

根据GPU显存大小合理设置batchsize。

3.1 RQ1复现

cd linevul
python linevul_main.py \--output_dir=./saved_models \--model_type=roberta \--tokenizer_name=microsoft/codebert-base \--model_name_or_path=microsoft/codebert-base \--do_train \--do_test \--train_data_file=../data/big-vul_dataset/train.csv \--eval_data_file=../data/big-vul_dataset/val.csv \--test_data_file=../data/big-vul_dataset/test.csv \--epochs 10 \--block_size 512 \--train_batch_size 16 \--eval_batch_size 16 \--learning_rate 2e-5 \--max_grad_norm 1.0 \--evaluate_during_training \--seed 123456  2>&1 | tee train.log

3.2 RQ2复现

cd linevul
python linevul_main.py \--model_name=12heads_linevul_model.bin \--output_dir=./saved_models \--model_type=roberta \--tokenizer_name=microsoft/codebert-base \--model_name_or_path=microsoft/codebert-base \--do_test \--do_local_explanation \--top_k_constant=10 \--reasoning_method=all \--train_data_file=../data/big-vul_dataset/train.csv \--eval_data_file=../data/big-vul_dataset/val.csv \--test_data_file=../data/big-vul_dataset/test.csv \--block_size 512 \--eval_batch_size 512

3.3 RQ3复现

cd linevul
python linevul_main.py \--model_name=12heads_linevul_model.bin \--output_dir=./saved_models \--model_type=roberta \--tokenizer_name=microsoft/codebert-base \--model_name_or_path=microsoft/codebert-base \--do_test \--do_sorting_by_line_scores \--effort_at_top_k=0.2 \--top_k_recall_by_lines=0.01 \--top_k_recall_by_pred_prob=0.2 \--reasoning_method=all \--train_data_file=../data/big-vul_dataset/train.csv \--eval_data_file=../data/big-vul_dataset/val.csv \--test_data_file=../data/big-vul_dataset/test.csv \--block_size 512 \--eval_batch_size 512

4 报错及解决办法

1. 出现 Unexpected key(s) in state_dict: "roberta.embeddings.position_ids", "encoder.roberta.embeddings.position_ids" 错误。

报错原因：transformers库的版本过高，回退版本。

2. 出现 ImportError: cannot import name 'SAVE_STATE_WARNING' from 'torch.optim.lr_scheduler' 错误。

报错原因：transformers库的版本过低，更新版本。

经实际测试，版本为 transformers==4.30.0 时可正常运行。

5 参考

GitHub - awsm-research/LineVul: A Transformer-based Line-Level Vulnerability Prediction