MLM4DNN (Element-Based Automated DNN Repair with Fine-Tuned Masked Language Model)

Overview of MLM4DNN

Main Components

1. Fine-tuning
2. Patch Generation
3. Patch Filtering
4. Patch Validation

Preparation

1. Clone repository

   git clone https://github.com/PGZXB/MLM4DNN.git

2. Benchmark preparation

   • Refer to benchmark/README.md to get our benchmark (i.e., $Benchmark_{APR4DNN}$)

3. Dataset preparation

   • Refer to dataset/README.md to get our dataset (i.e., $Dataset_{MLM}$)

4. Model preparation

   • Refer to models/README.md to get our fine-tuned model

5. Main conda env preparation (for component 1,2,3)

   • Create: conda create --name mlm4dnn_main python=3.11.0
   • Install libraries for this env
     • pytorch 2.1.0
     • transformers 4.36.2
     • requests 2.31.0
     • tqdm 4.66.1
   • See mlm4dnn_main_env.yml for details

6. PV conda env preparation (for component 4)

   • Create: conda create --name mlm4dnn_pv python=3.6.13
   • Install libraries for this env
     • keras 2.3.1
     • tensorflow 2.1.0
     • numpy 1.18.5
     • pandas 1.0.5
     • scikit-learn 0.23.1
     • astunparse 1.6.3
   • See mlm4dnn_pv_env.yml for details

Repro on $Benchmark_{APR4DNN}$

conda activate mlm4dnn_main
python mlm4dnn.py repro --dnn-train-env-name mlm4dnn_pv --output-dir /path/to/output

Re-Finetune Model & Perform MLM4DNN on New Model

1. Model Fine-tuning

   • conda activate mlm4dnn_main
   • python mlm4dnn.py train
   • The log and checkpoints will be saved in models/

2. MLM4DNN Performing

   • Create a config file (refer to configs/infill_api_config.json)
     • Change output_dir to new model's
   • To Repro: python mlm4dnn.py repro ... --infill-api-config-file /path/to/your_config