ciaua / clip2frame

Codes for "Event localization in music auto-tagging"

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Clip2frame

Clip2frame provides Python codes for the models in "Event Localization in Music Auto-tagging" by Jen-Yu Liu and Yi-Hsuan Yang, published in ACM Multimedia 2016.

Demo

http://clip2frame.ciaua.com

You can put any youtube link into the UI and see the frame-level prediction of our model. The Gaussian filter is removed from the model to produce more sharp prediction (Gaussian is still used in training)

Scripts

The following lists the scripts in the Scripts folder.

Training

Scripts/train_model.py

Testing -- Clip level

Scripts/test_model.clip.py

Testing -- Frame level

Scripts/test_model.frame.py

Network structure

Scripts/network_structure.py

Feature extraction

Scripts/extract_feats.py

Standardize data

Scripts/standardize_data.py

Process annotations

Scripts/process_annotations.top_n_tags.py

Make experimental data in Numpy format

Scripts/make_exp_data.py

Threshods derivation

Scripts/get_thresholds.with_magnatagatune.py Scripts/get_thresholds.with_medleydb.py

Data for experiments in the paper

The learned parameters of the best model with Gaussian filter in the paper is included in the package:

data/models/model.20160309_111546.npz

The network structure can be found here: Scripts/network_structure.py

This model has 188D output. The corresponding tags is here:

data/data.magnatagatune/tag_list.top188.txt

The correspondance between the 9 instruments for test in MedleyDB and the 188 tags in MagnaTagATune is here:

data/data.medleydb/instrument_list.medleydb_magnatagatune.top9.csv

The training, validation and testing features can be downloaded from the following links.

MagnaTagATune: clip-level training and testing

http://mac.citi.sinica.edu.tw/~liu/data/exp_data.MagnaTagATune.188tags.zip

Decompress the files into a folder

unzip exp_data.MagnaTagATune.188tags.zip -d exp_data.MagnaTagATune

MedleyDB: frame-level testing

Feature: http://mac.citi.sinica.edu.tw/~liu/data/feature.MedleyDB.zip

Annotations: http://mac.citi.sinica.edu.tw/~liu/data/annotation.medleydb.top9.zip

Data for deriving thresholds: http://mac.citi.sinica.edu.tw/~liu/data/exp_data.MedleyDB.length_320.zip

Decompress the files

unzip feature.MedleyDB.zip -d feature.MedleyDB
unzip annotation.medleydb.top9.zip

About

Codes for "Event localization in music auto-tagging"

License:ISC License


Languages

Language:Python 100.0%