Using Deep Learning to Generate Complete Log Statements

We present LANCE(Log stAtemeNt reCommEnder), a DL-based approach for supporting the task of log statement generation and injection in the context of Java. LANCE is built on the recently proposed Text-To-Text Transfer Transformer (T5) architecture

How to experiment with LANCE

• How to train a new SentencePiece Model

  Before training the T5 small, namely the core of LANCE, it is important to also train a new tokenizer (sentencepiece model) to accomodate the expanded vocabulary given by the java programming language. For such, we used the raw pre-training instances(Java corpus) + English sentences from the well known C4 dataset

  Pythonic way

  pip install sentencepiece==0.1.96
  import sentencepiece as spm
  spm.SentencePieceTrainer.train('--input=all_sp.txt --model_prefix=LOG_SP --vocab_size=32000 --bos_id=-1  --eos_id=1 --unk_id=2 --pad_id=0 --shuffle_input_sentence=true --character_coverage=1.0 --user_defined_symbols=“<LOG_STMT>”') 
  

  Under this path we also provide our trained tokenizer: https://github.com/antonio-mastropaolo/LANCE/tree/main/Code

• Setup a Google Cloud Storage (GCS) Bucket

  To setup a new GCS Bucket for training and fine-tuning a T5 Model, please follow the original guide provided by Google: Here the link: https://cloud.google.com/storage/docs/quickstart-console

• Datasets

  The datasets for pre-training, fine-tuning, validating and finally testing LANCE can be found at this link: https://drive.google.com/drive/folders/1D12y-CIJTYLxMeSmGQjxEXjTEzQImgaH?usp=sharing

• Pre-training/Fine-tuning

  To pre-train and then, fine-tune LANCE, please use the following:

  • Pre-Training
  • Fine-Tuning

• Models

  • Pre-trained on the tasks mixture (Multi-Task)
  • Pre-trained on LogSTMT only Task
  • Pre-trained on Denoise only Task
  • No Pre-trained

• Results: 📂

  • Multi-Task
  • LogSTMT only Task
  • Denoising only Task
  • No Pre-trained

• Additional:

  Under Miscellaneous, you can find the additional script used for the data analysis and the exact hyper-parameters configuration we employed in the study.