sjappig/mldn-capstone

AudioSet classification using RNN

Capstone Project for Machine Learning Nanodegree

Acquiring AudioSet samples

When in directory dataset/audioset/:

1. Check following dependencies: bash, ffmpeg and youtube-dl
2. Training dataset: mkdir train_data; cd train_data; cat ../balanced_train_segments.csv | ../download/download.sh
3. Test dataset: mkdir test_data; cd test_data; cat ../eval_segments.csv | ../download/download.sh

Downloading can be canceled by pressing Ctrl+C (few times if one does not seem to work). When started the next time, it will skip the already downloaded files (you might need to remove .part-files manually). Note that the below scripts won't work if there is very small number of samples (there seems to be some bug in pandas), so download at least ~100 samples for both training and testing before proceeding.

Preprocessing the samples and training models

When in repository root:

1. Create virtual environment: virtualenv --python=python2.7 sandbox
2. Activate the environment: source sandbox/bin/activate
3. Install required packages: pip install -r requirements.txt
4. Preprocess training data: python -m audiolabel.preprocess pp/train.h5 dataset/audioset/balanced_train_segments.csv dataset/audioset/train_data
5. Preprocess test data: python -m audiolabel.preprocess --normalize-using pp/train.h5 pp/test.h5 dataset/audioset/eval_segments.csv dataset/audioset/test_data
6. Train and test model with small number of samples and small epoch count: python -m audiolabel.fit_and_predict pp/train.h5 --N 100 --epochs 100 --validation-size 0 --test pp/test.h5 (this might take several minutes)

Check available command line parameters for both scripts with --help.