DCASE2019 Task 1b

DCASE2019 Task 1b - Acoustic Scene Classification with mismatched recording devices

Description

This repository contains CP-JKU Student team's submission for DCASE Challenge 2019. A technical report describing this system will be available on the DCASE homepage as soon as official evaluation results are available. We need to stress that self trained results might differ slightly from the ones described in the report, since we do not seed the random number generator manually. We therefore additionally provide all the files necessary to recreate our submissions in folders tmp/data/{no_da, mse_da_0, mse_da_1, mi_da}.

For a detailed description of task, data set, and baseline, see: http://dcase.community/challenge2019/task-acoustic-scene-classification

Step-by-Step Guide for Prediction & Training

As a first step we need to set up the environment:

• Clone this project to you local machine:

  git clone https://github.com/OptimusPrimus/dcase2019_task1b.git
  

• Download and place data set into the data folder:

  .
  │
  └───data
  .   └───raw
  .       └───dcase20191b
  .           |   create_folds.py
  .           |   meta.csv 
  .           └───submission                           # submission data
  .           .   |   1.wav
  .           .   |   ...
  .           └───test                                # leaderboard data
  .           .   |   1.wav
  .           .   |   ...
  .           └───audio                               # training
  .           .   |   tram-vienna-285-8639-a.wav
  .           .   |   ...
  .           └───evaluation_setup                    # evaluation split
  .           .   |   fold1_evaluate.csv 
  .           .   |   fold1_test.csv
  .           .   |   fold1_train.csv
  .           └───training_setup                      # empty, 4-flod CV split
  |   ...
  
  

• Import conda environment & activate it:

  conda env create -f environment.yaml && source activate slap
  

• Create folds:

  cd data/raw/dcase20191b
  python create_folds.py
  cd ../../..
  

• Install Omniboard:

  npm install -g omniboard
  

• Install & Setup MongoDB (more here)

• Create Database (e.g. use dcase20191b)

• Add file mongodb.json with MongoDB credentials to root:

  {
    "user": "username",
    "pwd": "password123",
    "ip": "127.0.0.1",
    "db": "dcase20191b",
    "port": 27017
  }

• [OPTIONAL] Add file telegram.json with Telegram Bot credentials to root:

  {
    "token": "SecretToken",
    "chat_id": "ChatID"
  }

Predict

Use the provided training files to create predictions for leaderboard and evaluation set. The command to create predictions is:

python predict.py model_id_0 [model_id_1 model_id_2 ...]

To create all submission files simply run:

python predict.py no_da && python predict.py mse_da_0 && python predict.py mi_da && python predict.py mse_da_0 mse_da_1

The submission files can be found in directory data\tmp.

Train the System Yourself

• Start MongoDB:

  mongod
  

• Activate conda environment:

  source activate slap
  

• Edit configs/dcase20191b.json

• Train MSE, MI, and NoDA models with:

  OMP_NUM_THREADS=1 CUDA_VISIBLE_DEVICES=0 python main.py
  

  OMP_NUM_THREADS=1 CUDA_VISIBLE_DEVICES=0 python main.py with training.domain_adaptation.class=domain_adaptation.MSE
  

  OMP_NUM_THREADS=1 CUDA_VISIBLE_DEVICES=0 python main.py with training.domain_adaptation.class=domain_adaptation.MutualInformationDA
  

See Sacred CLI Description on how to manipulate training parameters. Trained models can be found in folder data/tmp. To see training progress, start Omniboard:

omniboard -m 127.0.0.1:27017:dcase20191b

Citation

If you use the model or the model implementation please cite the following paper:

@inproceedings{Koutini2019Receptive,
    author      =   {Koutini, Khaled and Eghbal-zadeh, Hamid and Dorfer, Matthias and Widmer, Gerhard},
    title       =   {{The Receptive Field as a Regularizer in Deep Convolutional Neural Networks for Acoustic Scene Classification}},
    booktitle   =   {Proceedings of the European Signal Processing Conference (EUSIPCO)},
    address     =   {A Coru\~{n}a, Spain},
    year        =   2019
}

If you use other parts of the implementation please cite:

@techreport{Primus2019DCASE,
    Author      =   {Primus, Paul and Eitelsebner, David},
    institution =   {{DCASE2019 Challenge}},
    title       =   {Acoustic Scene Classification with mismatched recording devices},
    month       =   {June},
    year        =   2019
}

References

[1] Annamaria Mesaros, Toni Heittola, and Tuomas Virtanen. A multi-device dataset for urban acoustic scene classification. In Proceedings of the Detection and Classification of Acoustic Scenes and Events 2018 Workshop (DCASE2018), 9–13. November 2018. URL: https://arxiv.org/abs/1807.09840.

[2] Khaled Koutini, Hamid Eghbal-zadeh, Matthias Dorfer, and Gerhard Widmer. The Receptive Field as a Regularizer in Deep Convolutional Neural Networks for Acoustic Scene Classification. In Proceedings of the European Signal Processing Conference (EUSIPCO), June 2019.

Links

• Sacred Documentation
• Sacred GitHub
• Omniboard