sushentsev / dapstep

DapStep: Deep Assignee Prediction for Stack trace Error rePresentation

DapStep

DapStep is an algorithm for predicting assignees presented in our paper "DapStep: Deep Assignee Prediction for Stack Trace Error rePresentation".

Dataset

The public dataset presented in the paper can be found here.

Format

The Assignee Prediction Dataset consists of error stack traces and annotations to them. All annotations were collected from the IntelliJ IDEA Community using the git blame command. The dataset is anonymized, each entity is encoded with a unique identifier, all the temporal data has been shifted by a fixed timestamp. The data is presented as follows:

• the reports directory contains error reports, each report includes a unique report identifier (id), error time (timestamp), and a stack trace (elements). Each stack frame consists of the method name (name), file name (file_name), line number (line_number), annotation commit hash (commit_hash), and subsystem (subsystem).
• the files directory contains annotations where the name of each annotation is {commit_hash}:{file_name}:annotation.csv.
• the labels.csv file associates each error report (rid) with the developer (uid) who fixed the given error .

The bug report is presented in the following JSON:

{
  "id": 4524,
  "timestamp": "1235540669458",
  "elements": [
    {
      "name": "javax.swing.Timer.fireActionPerformed",
      "file_name": "Timer.java",
      "line_number": 313,
      "commit_hash": 8138bd7e342810cfdc5174dcb222d136bd229288,
      "subsystem": "javax.swing"
    },
    {
      "name": "java.awt.event.InvocationEvent.dispatch",
      "file_name": "InvocationEvent.java",
      "line_number": 311,
      "commit_hash": eb2vb45e75108e0c8d0c44e80c78e9d061a11448,
      "subsystem": "java.awt.event"
    },
  ]
}

where

• id – identifier of the report
• timestamp – timestamp of the report creation (Unix time)
• elements – sequence of stack frames starting from the top of the stack

The annotation is presented in the CSV file:

commit_hash,author,timestamp
580dsbcfd18374a14575b65085ba46adea8b015d,24,1108830319000
9d78b417254b099c9d3ae349cfbeed0e4d8efa42,631,1341273406000
9d26b217254b099c9d3ae919cfbvbd0e4d8efa42,123,1641999406000

where

• commit_hash – commit hash of the last edit
• author – identifier of the developer
• timestamp – timestamp of the last edit (Unix time)

The labels CSV file looks like this:

rid,uid
234,116
54,47
4,116

where

• rid – identifier of the report
• uid – identifier of the developer

How to use

Install

pip install -r requirements.txt

Train ranking DL-based models

The example of a train config can be found in src/scripts/configs/dl_ranking.yaml and has the following form:

data_dir: data_dir
features_dir: features_dir
save_dir: save_dir

data_split:
  val_size: 0
  test_size: 1500

coder:
  entry_coder_type: file_name
  cased: True
  trim_len: 0
  rem_equals: False

model:
  emb_type: cnn

optimizer:
  lr: 0.001
  weight_decay: 0.001

train:
  epochs: 10
  update_every: 4

where

• data_dir – directory where the data is stored The directory should have the following structure:
  • reports folder with all reports in JSON format
  • files folder with all annotations in CSV format
  • labels.csv file with report labels
• features_dir – directory with features The example of such directory can be found in src/scripts/features_examples The lists of all features are presented in src/features/features_maps.py
• save_dir – directory where the model files will be saved to
• data_split – time-based data splitting
• coder – parameters of the sequence coder
• model – model type (cnn or rnn)
• optimizer – parameters of the optimizer
• train – parameters of the training

Run train:

cd src/scripts/train/
python dl_manual_features_ranking.py

or

cd src/scripts/train/
python dl_neural_features_ranking.py

The trained model will be saved in save_dir.

Run ranking DL-based models

Run eval:

cd src/scripts/eval/
python eval.py --data_dir=<DATA_DIR> 
--model_dir=<MODEL_DIR> --features_dir=<FEATURES_DIR> 
--test_size=<TEST_SIZE>

How this works

TBD: add the link to the paper.

Contacts

If you have any questions or suggestions, feel free to create an issue or reach Denis Sushentsev at [email protected].