Deep learning, DTU 02456. Autumn 2018.

The aim of this prokect is to test the hypothesis that: training with weakly labeled outputs to predict strongly labeled outputs.

• Weakly labeled output: only say if a label is present in the whole audio clip or not
• Strongly labeled output: per each label it says when, in time, each label appear in the audio clip (with a precision of 1s).

Final project working with Google AudioSet inspired in DCASE2018 Challange (Task 4).
See project documentation: https://drive.google.com/drive/folders/1wYKw9w9nIngUnUSXUytIEmZBjXbPZIWh

We recommend you to create a folder in your computer (e.g. 'your_project') and then clone it inside this folder.

1. Create a folder your_project

    your_project
   

2. Clone this repository inside the folder

    your_project/
        |_ 02456_project_audioset_attention (this repository)
   

3. Run 'setup.sh' to download and clone the files (it might take a while). It will automatically do the following:

   3.1. Uncomment some lines to download the packed_features.zip from Google AudioSet project.

   3.2. Clone audioset_classification from https://github.com/qiuqiangkong/audioset_classification

   3.3. Clone dcase18_baseline from https://github.com/DCASE-REPO/dcase2018_baseline/tree/master/ We are interested in the metadata of Task4 because it is strongly labeled data

   3.4. Create a symbolik link of main_3.py and core_3.py in the 'audioset_classification' repository. This are the files we modified to train the networks to be able to test our hypothesis.

   3.4. The directories should look like this:

    your_project/
        |_ 02456_project_audioset_attention (this repository)
        |_ audioset_classification
        |_ dcase2018_baseline
        |_ packed_features
   

1. If you want to train the network you are now ready to run the runme.sh file. It was already done and we provide you the data from the trained models in the data folder of this repository.

2. The results of this project are presented in the Jupyter notebook Results.ipynb.

If at some point you need to download the dataset again but the setup is already done you can just run the shell script 'data_generator.sh'

If you want to remove the folders created by during the setup you can just run the shell script 'remove_all.sh'. Then it will look like this again:

your_project/
    |_ 02456_project_audioset_attention (this repository)

Repositories: https://github.com/DTUComputeCognitiveSystems/AI_playground/blob/master/notebooks/experiments/Sound%20Demo%203%20-%20Multi-label%20classifier%20pretrained%20on%20audioset.ipynb

Contact us if you want more information sending an e-mail to [email protected]