mml2020-experiments

This repository contains scripts for reproducing the audioLIME experiments for our paper MML2020 paper (arxiv link).

Setup

• Head to the audioLIME repository and follow the installation instructions there
• Call pip install -r requirements.txt
• You need this list of audio files from the MillionSongDataset (which can not be shared)

Calls to reproduce plot

First you have to set the location of the spleeter models (or where you want them). Otherwise it will download you them in your current working directory.

export MODEL_PATH=/share/home/verena/experiments/spleeter/pretrained_models/

First compute and store the explanations (takes up a lot of space!):

python explain_samples_in_dir.py --samples_dir /share/home/verena/experiments/understanding_musictagging/selected_samples_100/ --out_dir /share/home/verena/experiments/understanding_musictagging/explanations_100/ --num_samples=16384 --batch_size=32 --n_temporal_segments=10 --model_type=fcn --n_display_components=3 --n_chunks=3

python explain_samples_in_dir.py --samples_dir /share/home/verena/experiments/understanding_musictagging/selected_samples_100/ --out_dir /share/home/verena/experiments/understanding_musictagging/explanations_100/ --num_samples=16384 --batch_size=32 --n_temporal_segments=9 --model_type=sample --n_display_components=3

python explain_samples_in_dir_slime.py --samples_dir /share/home/verena/experiments/understanding_musictagging/selected_samples_100/ --out_dir /share/home/verena/experiments/understanding_musictagging/explanations_100_slime/ --num_samples=16384 --batch_size=32 --n_temporal_segments=9 --model_type=sample --n_display_components=3

python explain_samples_in_dir_slime.py --samples_dir /share/home/verena/experiments/understanding_musictagging/selected_samples_100/ --out_dir /share/home/verena/experiments/understanding_musictagging/explanations_100_slime/ --num_samples=16384 --batch_size=32 --n_temporal_segments=10 --model_type=fcn --n_display_components=3 --n_chunks=3

Run analyse_components.py for both models (replace {model_type} by fcn or sample) and different {n_components} (3, 5, 7, 9, 11).

For audioLIME:

python analyse_components.py --input_dir /share/home/verena/experiments/understanding_musictagging/explanations_100/ --model_type {model_type} --analysis_path /share/home/verena/experiments/understanding_musictagging/analyse_components_100/ --n_random_samples 32 --n_components {n_components}

The same for SLIME:

python analyse_components.py --input_dir /share/home/verena/experiments/understanding_musictagging/explanations_100_slime/ --model_type {model_type} --analysis_path /share/home/verena/experiments/understanding_musictagging/analyse_components_100_slime/ --n_random_samples 32 --n_components {n_components}

Finally, create the figure:

python -m plot_figure_2 --analysis_path_audiolime /share/home/verena/experiments/understanding_musictagging/analyse_components_100/ --analysis_path_soundlime /share/home/verena/experiments/understanding_musictagging/analyse_components_100_slime/