RapLyrics-Back

Part of the raplyrics.eu project: generate powerful AI-powered rap lyrics.

This repository contains:

• a hand-curated, human-cleaned rap lyrics dataset
• the implementation of a generative rnn model (based on minimaxir work)
• a custom generative function and a serving implementation

Getting started

Setup

Get the repo. Clone from GitHub:

$ git clone https://github.com/cyrilou242/RapLyrics-Back  

Setup a virtualenv and install the required libraries.

Note: Make sure you have a python --version == python3.6 otherwise some library (especially tensorflow may not be available or behave as expected)

$ cd RapLyrics-Back  
$ mkvirtualenv --python `which python3.6` -r requirements.txt RapLyrics_Back  

Train the model

The training is based on minimaxir's textgenrnn with small pre-processing tweaks. The development of this library is still active, you may want to use it instead of our fork:

$ rm -r textgenrnn  
$ pip install textgenrnn    

train.py contains the parameters and the path of the training dataset.
Have a look at it and launch your own training. Don't change anything to train from the rap lyric dataset provided. This will create a word model and a char model.

Parameters available:
new_model: True to create a new model or False to train based on an existing one.
rnn_layers: Number of recurrent LSTM layers in the model (default: 2)
rnn_size: Number of cells in each LSTM layer (default: 128)
rnn_bidirectional: Whether to use Bidirectional LSTMs, which account for sequences both forwards and backwards. Recommended if the input text follows a specific schema. (default: False)
max_length: Maximum number of previous characters/words to use before predicting the next token. This value should be reduced for word-level models (default: 40)
max_words: Maximum number of words (by frequency) to consider for training (default: 10000)
dim_embeddings: Dimensionality of the character/word embeddings (default: 100)
num_epochs: number of epochs. (model is save at eah epoch by default)
word_level: Whether to train the model at the word level (default: False)
dropout: ratio of character to ignore on each sentence, may lead to better results. Don't use with word model
train_size: ratio to use as training set. The remaining ratio will make a validation set. Default is 1. (no validation set)
gen_epochs: during the training, a generation test will be made at each multiple of gen_epochs. Default is 1 (generation at each epoch)
max_gen_length: length of the generation during training

3 files will be generated by the training:
Model_name_config.json : parameters given for the training
Model_name_vocab.json : vocabulary of the dataset
Model_name_weights.hdf5 : model weights
Put these file in the weights folder.

Serve the model

The paths of the model files are defined in the first lines of server.py.
Change them to use your custom model.

The parameters of the custom generative function can also be tweaked.
(documentation fot this function will be issued in another document)

To test the serving, launch the development flask server:

$ python app.py

Make a request to the server in a different terminal:

$ curl '127.0.0.1:5000/apiUS' -X POST -H "Content-Type: application/x-www-form-urlencoded" -d "input=low life"  

{"output":"low life down the whip\nand i don't mean it 'cause i never had a death\ni guess i have a story of the bitches change\nfor the life\nfor the life that i know will you get through\n'cause i don't give a fuck i got the way before i started inhale"}

For further serving in production, use a production WSGI server. We recommend gunicorn.

Next steps:

• Further researches and testing with character level models. Seems very promising.