Spotify provides millions of song tracks covering a full spectrum of music variety to users all over the world on multiple platforms. The key mission of Spotify is to help people find the right music at every moment through tailored song recommendations and playlists. We carry on this important spirit in our project and seek to further enhance users’ music enjoying experience. Specifically, as many research focus on acoustic features in music analysis, we aim to analyze the language component of songs, i.e. lyrics, to associate songs with relatable concepts such as moods and themes. We explore with supervised and unsupervised methods including Latent Dirichlet Allocation(LDA), Word2Vec and Long Short-Term Memory(LSTM) for language processing and predictive modeling. Our final model can be used to provide a two-way translation between songs and tags. In one direction, we have an automatic tagging system which suggests potential tags for any song. And conversely, we can recommend a list of songs associated with the keywords from user inputs. The lyrics based model yields satisfactory predictions and thus offers a new angle in analyzing song content. By leveraging the rich emotional and descriptive content encrypted in song lyrics, we provide Spotify a solution to create even more personalized music enjoying experience for its users.
Note: Please run all the functions under python 2
-
Install all the packages by running the following shell script:
./setup.sh -
Download pre-trained word2vec model from GoogleNews-vectors-negative300.bin.gz. Unzip the word2vec model and place the file inside this folder:
./Final_Deliverable/pipeline/1-LSTM_model_for_tag_prediction.
Note: if you have already installed django, gensim, tensorflow, or keras package, please make sure they have the following versions:
django: 1.8.13,
gensim: 2.0.0,
tensorflow: 1.1.0,
keras: 2.0.4
cd ./Final_Deliverable/pipeline/0-baseline_model_for_tag_prediction/ to run the baseline model
Run feature_engineering.py by:
python feature_engineering.py raw_data.csv [-f output.csv -n n_features]
Example:
python feature_engineering.py full_dataset.csv -f test.csv -n 150
Parameters explanation:
-
raw_data.csv(input, required) - the file contains the information about all the songs in database. Required informations are (in order): Current_index, original_index, song_id, song_name, artist, tag, path_to_the_song lyrics_file, song_lyrics. -
output.csv(output, optimal) - the file returns the hashing vectorizer representation of the lyrics and tags needed for the modeling step. default: "features.csv" -
n_features(int, input, optimal) - number of dimensions of the hashing vectorizer that we want for the modeling step. default: 150
Run model_training.py by:
python model_training.py input_feature.csv output_model_name [-e est -d depth]
Example:
python model_training.py features.csv output_model -e 20 -d 5
Parameters explanation:
-
input_feature.csv(input, required) - the hashing vectorizer of the raw data generated by feature_engineering.py. -
output_model_name(output, required) - the name of output model we want to save for the fitting part. -
est(int, input, optimal) - number of estimators required for Random Forest model. default: 15 -
depth(int, input, optimal) - max depth of the Random Forest model. default: 5
Run model_fitting.py by:
python model_fitting.py lyrics.txt -n n_features -m model
Example:
python model_fitting.py ../lyrics/If\ I\ die\ Young.txt -n 150 -m model.pickle
Parameters explanation:
-
lyrics.txt(input, required) - the lyrics txt file of the song we want to predict. -
n_features(input, optional) - number of dimensions of the hashing vectorizer, must equivalent to n_features in training step. default: 150 -
model(input, optional) - the model we trained in model training step. default: model.pickle - pre-trained model by the developer.
Return value:
- Terminal prints out 3 tags that have the top 3 probability.
- Example output:
The prediction results are: love memory sad
cd ./Final_Deliverable/pipeline/1-LSTM_model_for_tag_prediction/ to run the advanced model.
Note: Since the LSTM model is huge. It takes very long time to train and fit the model on a personal computer.
Run model_training.py by:
python model_training.py output_model_name
Example:
python model_training.py output_model
Parameters explanation:
output_model_name(output, required) - the name of the output model.
Run model_fitting.py by:
python model_fitting.py lyrics.txt model_name.h5
Example:
python model_fitting.py ../lyrics/If\ I\ die\ Young.txt output_model.h5
Parameters explanation:
-
lyrics.txt(input, required) - the lyrics txt file of the song we want to predict. -
output_model.h5(input, required) - the trained model from model_training.py
Return value:
- Terminal prints out 3 tags that have the top 3 probability.
- Example output:
The three top tags are: tag: funny tag: love tag: sad
cd ./Final_Deliverable/pipeline/2-Recommendation/ to run the Recommendation model
Run tag_to_song.py by:
python tag_to_song.py tag_name
Example:
python tag_to_song.py happy
Parameters explanation:
tag_name(input, required) - the name of tag or keywords that you want to get song playlist from.
Return value:
- the top 10 songs that most related to the input tag or keywords.
cd ./Final_Deliverable/website/mysite/ to start the web server.
python manager.py runserver
You will see similar output as below from terminal:
Performing system checks...
System check identified no issues (0 silenced).
May 14, 2017 - 08:54:41
Django version 1.8.13, using settings 'mysite.settings'
Starting development server at http://127.0.0.1:8000/
Quit the server with CONTROL-C.
type in chrome the following address:
http://127.0.0.1:8000/index/
You will see our webpage after loading for a while(3-5 min).
type the artist and song name in the corresponding boxes and hit predict.
The web will return the top 3 tags that associated with this song and the song lyrics
type in the keywords that you want the song playlist relates to.
The web will return list of 10 songs with the link to spotify song track.