Python project with the aim of analyzing Hit Song Prediction.

The dataset from which I've started to train and validate the system proposed is SpotGenTrack. This dataset contains:

• Spotify data of 101940 tracks between songs and podcasts (spotify_track.csv , spotify_album.csv, spotify_artists.csv)
• Low level features computed on each song
• Lyrics features computed on each song
• Preview url for wach one

A cleaning process has been conducted to remove not properly annotated songs, duplicates, songs that contain podcasts and also to update the lyrics using the MusixMatch API since sometime title and lyrics of songs did not match.

I've ended up with two different datasets: one with only English songs and another with multi-lingual songs that can be found in the Datasets directory. Starting from the Spotify ID, in spotify_track.csv the url to download the audio file is available. In these datasets information stored are: spotify_ID, musixmatch_ID, release year, lyrics and popularity score assigned by Spotify in 2019.

The model consists in three main components:

• Audio Embedding extractor
• Text Embedding extractor
• Final Multi-Layer Perceptron to predict songs popularity

• datasets: it contains .csv / .parquet files with english songs and multilingual songs. Additional information of songs (title, artist ecc) are stored in /nas/home/ecastelli/Data Sources folder and divided into 3 files: spotify_album.csv, spotify_artists.csv and spotify_tracks.csv. A join can be done using as key the spotify_id and track_id columns.
• models: it contains the models used for this project.
  • podcast_discriminator contains a .ipynb notebook used to implement a model for distinguishing audio that contains podcasts and podcasts that conatins music
  • genre_classificator is a model pre-trained on GTZAN Genre that is used as audio feature extractor
  • hsp_model is the final multi-layer perceptron used to predict the song popularity

Starting from the train.py file the training can be started passing three parameters:

• Problem to solve: classification (c) or regression (r)
• Language to consider: english (en) or multilingual (mul)
• Number of popularity classes to be considered in case of classification

To start the training process:

• Install the requirements.txt in your virtual environment
• Check the connection to the ISPL servers to be able to have access to the audio files stored there (/nas/home/ecastelli/thesis/Audio) and to the checkpoint of the pre-trained model ("/nas/home/ecastelli/thesis/models/Model/checkpoint/pretraining.ckpt").
• Change the NeptuneLogger parameters if you want to be able to see logs on Neptune or substitute it with whatever you want
• Change the device type used by the trainer in GTZANPreTrained.py and HSPModel.py with your kind of gpu
• Launch from the main folder python train.py --help and follow the instructions