Skip to content
/ tango Public
generated from ncoop57/nb_template

Tango is a research tool for automatically detecting duplicate video-based bug reports by combining visual and textual information present in the videos.

License

Notifications You must be signed in to change notification settings

ncoop57/tango

Repository files navigation

Tango 🕺💃

Tango is a research tool for automatically detecting duplicate video-based bug reports by combining visual and textual information present in the videos.

DOI PyPI - Python Version PyPI Status PyPI Status license

Data

We provide all of our data, models, and results. You can download the zip here: https://doi.org/10.5281/zenodo.4439661

In this data package, you will find two folders: artifacts and outputs

artifacts contains the videos we collected in our user study, the model files for the different models we evaluated, and the detailed results that we generated (See Detailed Results section for more information). The videos folder is broken down by user, where each user has a folder contain the apps they were given to create videos for. Each of the apps contain folders that denote the bug they generated a report for. Finally, instead these bug folders there is the actual video-based bug report as an mp4 file. The user_assignment.csv just contains the finalized assignments of users to corresponding bug reports.

In the models folder, you will find the two models we evaluated (SIFT, SimCLR, and OCR+IR). In each folder you will find the corresponding trained codebook files that we generated for SIFT and SimCLR. These codebook files are pickle files that contain the binary representation of the trained codebooks. Additionally, in the SimCLR folder, you will find a checkpoint and pytorch model file that contains all the necessary information for reloading our trained SimCLR model. For the OCR+IR folder, you will find all of the code for the OCR+IR model as well as the intermediate output for this particular model, other models' outputs are stored in the outputs folder.

The outputs folder contains all of the intermediate outputs of our code, except for OCR+IR. In the results folder, you will find all of the raw rankings and metrics for the SIFT and SimCLR model for all combinations of video-based bug reports per app. NOTE: SIFT is missing the 10k raw ranking and metrics, but will be provided in a future version. evaluation_setting contains a json file that contains all of the duplicate detection tasks we used for evaluating our models, i.e. setting 2 (See paper for more details). user_rankings_weighted_all and user_results_weighted_all contain converted version of the raw rankings and metrics for the SIFT and SimCLR model to match setting 2. extracted_text contains the output of running the OCR model, i.e. the frames of the videos and the text from each frame. Lastly, combined contains the results of the combined tango approach.

Reproduce Results

All code has only been tested on Ubuntu 20.04. Additionally, there is an issue when attempting to reproduce tango via Docker on MacOS. If you are using MacOS, please look at our steps for how to reproduce without Docker.

The prefered method to reproduce our paper's results is to use Docker. Please install Docker if you do not already have it install.

git clone https://github.com/ncoop57/tango.git
cd tango

Reproduce via Docker:

docker run -v <out_path>:/data semerulab/tools:tango <vis_model>
  • out_path: The absolute path on your machine you want all files to be saved to
  • vis_model: The type of visual model. Can be either SimCLR or SIFT, taking ~6 hours or >2 weeks, respectively, for all apps on our machine with 755G of RAM and 72 CPUs.

You can also build the docker image yourself:

cd docker_build
docker build -f Dockerfile.prod -t tango .
cd ..
docker run -v <out_path>:/data tango <vis_model>

Reproduce without Docker:

pip install two-to-tango
tango_reproduce <down_path> <out_path> <vis_model>
  • down_path: The directory where all the files will be downloaded and extracted to.
  • out_path: The output path to place all results in.
  • vis_model: The type of visual model. Can be either SimCLR or SIFT, taking ~6 hours or >2 weeks, respectively, for all apps on our machine with 755G of RAM and 72 CPUs.

You can also use this Google Colab notebook, however, due to the timeout associated with Google Colab, you will most likely be unable to generate all of our results in the alotted time.

Detailed Results

You can find a spreadsheet containing the results for all of the different configurations we tested at tango_reproduction_package/artifacts/detailed_results.xlsx.

In this excel file, we have multiple sheets. overall shows the performance of the different model configurations averaged across all apps. overall_comb shows the combined performance of the visual and textual model configurations averaged across all apps. Additionally, per-app and per-app-comb has the performance of the single and combined model configurations per app, respectively. Lastly, we provide the overall performance in sheet overall_user_study and overall_user_study_comb of the single and combined model configurations on the settings (used only APOD app) given to the users for evaluating how much time and effort tango can save developers.

All sheets show the performance in terms of mRR (avg_rr), the standard deviation of recipical rank, median (med_rr), and quartile 1 and 3 (q#_rr). The same is true for mAP (avg_ap). We also show the performance in terms of average rank including standard deviation, and quartiles. Lastly, we providing HIT@1-5, 7, and 10 (h#).

Sheets that contain the weight column have information regarding how much weight is given to the visual and textual information. A value of 0.1 means that the textual information received a weight of 0.1 while the visual information was given a weight of 0.9. For values containing two numbers, e.g. 0.1-0.0, refers to the weighting scheme introduced in the paper for when there may be high overlap in vocabulary between duplicate and non-duplicates (See paper for more details). If an app does not have high overlap, then a weight of 0.1 is used for the textual information, else the textual information is not considered, i.e., weight of 0.0.

Install

From pypi:

pip install two-to-tango

From source:

git clone https://github.com/ncoop57/tango.git
cd tango
pip install .

How to use

Download Data CLI Tool

tango_download <out_path>

  • out_path: The output path to save and unzip all files.

Duplicate Detection CLI Tool

tango <query_path> <corpus_path> <simclr_path>

  • query_path: The path to a video you want to detect duplicates for.
  • Example: <out_path>/tango_reproduction_package/artifacts/cli_videos/U1/APOD/CC1/APOD-CC1_fixed_30.mp4
  • corpus_path: The path to a video you want to detect duplicates for.
  • Example: <out_path>/tango_reproduction_package/artifacts/cli_videos
  • simclr_path: The path to a video you want to detect duplicates for.
  • Example: <out_path>/tango_reproduction_package/artifacts/models/SimCLR

Example Output:

OrderedDict([   (('APOD', 'CC1', 'U1'), 0.9838350837260246),
                (('APOD', 'CC1', 'U12'), 0.9193482983504456),
                (('APOD', 'CC1', 'U2'), 0.3723964243572911),
                (('APOD', 'CC6', 'U12'), 0.3718521026630344),
                (('APOD', 'CC9', 'U12'), 0.36803837161265085),
                (('APOD', 'CC6', 'U8'), 0.33589710905277315),
                (('APOD', 'CC4', 'U12'), 0.3118613303188616),
                (('APOD', 'CC9', 'U5'), 0.2718403622668689),
                (('APOD', 'CC4', 'U7'), 0.25082093055745974),
                (('APOD', 'CC9', 'U9'), 0.22580393621884165),
                (('APOD', 'CC5', 'U12'), 0.1768510685792533),
                (('APOD', 'CC6', 'U7'), 0.1682816804179776),
                (('APOD', 'CC4', 'U8'), 0.13915926428362999),
                (('APOD', 'CC2', 'U2'), 0.1354447367818957),
                (('APOD', 'CC3', 'U2'), 0.12004454785432789),
                (('APOD', 'CC5', 'U7'), 0.11235793525631509),
                (('APOD', 'CC5', 'U8'), 0.10867946897348428),
                (('APOD', 'CC7', 'U8'), 0.09014217805772731),
                (('APOD', 'RB', 'U1'), 0.08324154319710894),
                (('APOD', 'CC8', 'U12'), 0.08179046960502091),
                (('APOD', 'CC8', 'U5'), 0.07290831091450554),
                (('APOD', 'RB', 'U5'), 0.07194441953180176),
                (('APOD', 'CC3', 'U12'), 0.06729098674201965),
                (('APOD', 'CC7', 'U7'), 0.06326635817907807),
                (('APOD', 'CC7', 'U12'), 0.05922061313241868),
                (('APOD', 'CC8', 'U10'), 0.05328420969145727),
                (('APOD', 'CC2', 'U12'), 0.04707548776015297),
                (('APOD', 'CC3', 'U1'), 0.04342630487280919),
                (('APOD', 'CC2', 'U1'), 0.04211602057931267),
                (('APOD', 'RB', 'U12'), 0.029766244020504186)])

If you would like to do this on your own data, look at the tango_reproduction_package/artifacts/cli_videos directory structure from our dataset on Zenodo for how you need to formate your data to work with tango

Training SimCLR

For training the SimCLR model we used the RICO dataset and this repository for training a SimCLR model using Pytorch Lightning

To cite our work, please use the following bibtex:

@inproceedings{Cooper:ICSE21,
      title={It Takes Two to Tango: Combining Visual and Textual Information for Detecting Duplicate Video-Based Bug Reports}, 
      author={Nathan Cooper and Carlos Bernal-Cárdenas and Oscar Chaparro and Kevin Moran and Denys Poshyvanyk},
      year={2021},
      booktitle={ICSE’21},
}

About

Tango is a research tool for automatically detecting duplicate video-based bug reports by combining visual and textual information present in the videos.

Resources

License

Stars

Watchers

Forks