CP JKU Submission for DCASE 2019

UPDATE 2020: a new version of this repo with support to more datasets and architectures has been moved to https://github.com/kkoutini/cpjku_dcase20

Requirements

Conda should be installed on the system.

install_dependencies.sh installs the following:

• Python 3.6.3
• PyTorch
• torchvision
• tensorboard-pytorch
• etc..

Installation

• Install Anaconda or conda

• Run the install dependencies script:

./install_dependencies.sh

This creates conda environment cpjku_dcase19 with all the dependencies.

Running source activate cpjku_dcase19 is needed before running exp*.py

Usage

After installing dependencies:

• Activate Conda environment created by ./install_dependencies.sh

  $ source activate cpjku_dcase19

• Download the dataset:

  $ python download_dataset.py --version 2019

  You can also download previous versions of DCASE --version year, year is one of 2018,2017,2016,2019.

  Alternatively, if you already have the dataset downloaded:

  • You can make link to the dataset:

  ln -s ~/some_shared_folder/TAU-urban-acoustic-scenes-2019-development ./datasets/TAU-urban-acoustic-scenes-2019-development

  • Change the paths in config/[expermient_name].json.

• Run the experiment script:

  $ CUDA_VISIBLE_DEVICES=0 python exp_[expeirment_name].py 
  

• The output of each run is stored in outdir, you can also monitor the experiments with TensorBoard, using the logs stored in the tensorboard runs dir runsdir. Example:

  tensorboard --logdir   ./runsdir/cp_resnet/exp_Aug20_14.11.28

The exact commmand is printed when you run the experiment script.

Example runs

DCASE 2019 DCASE 1A

CP_ResNet

default adapted receptive field RN1,RN1 (in Koutini2019Receptive below):

$ CUDA_VISIBLE_DEVICES=0 python exp_cp_resnet.py 

Large receptive Field

$ CUDA_VISIBLE_DEVICES=0 python exp_cp_resnet.py  --rho 15

very small max receptive Field:

$ CUDA_VISIBLE_DEVICES=0 python exp_cp_resnet.py  --rho 2

Loading pretrained models

Download the evaluation set:

$ python download_dataset.py --version 2019eval

Download the trained models (from zando)

Run the experiment with the load the correct rho value, because the rho value changes the network weights shape)

$ CUDA_VISIBLE_DEVICES=0 python exp_cp_resnet.py  --rho 5 --load=path_to_model.pth

In case that you want to predict on a different dataset, you should add the dataset to the config file. For example look at the eval dataset in configs/cp_resnet_eval.json.

Missing Features

This repo is used to publish for our submission to DCASE 2019 and MediaEval 2019. If some feauture/architecture/dataset missing feel free to contact the authors or to open an issue.

Citation

If use this repo please cite The Receptive Field as a Regularizer , Receptive-Field-Regularized CNN Variants for Acoustic Scene Classification :

@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
}


@inproceedings{KoutiniDCASE2019CNNVars,
  title = {Receptive-Field-Regularized CNN Variants for Acoustic Scene Classification},
  booktitle = {Preprint},
  date = {2019-10},
  author = {Koutini, Khaled and Eghbal-zadeh, Hamid and Widmer, Gerhard},
}