Lyrics-Generation-using-RNNs

This project generates new lyrics which did'nt exist before of your favourite artists , using Recurrent-Neural-Networks (RNNs) from a data-set created by web scrapping The GENIUS website using its API.

After vectorizing the text and creating training examples and targets , a four layer model was used. The Four layers used are :

1. Embedding layer -> The input layer. A trainable lookup table that will map the numbers of each character to a vector with embedding_dim dimensions

2. GRU layer -> A type of RNN with size units=rnn_units (Birectional-LSTM could also be used here.)

3. Dense layer -> The output layer, with vocab_size outputs and 'RELU' as the activation fuction

4. Dropout layer -> Benifits regularisation and prevents overfitting

The final prediction loop for text generation works as follows

Steps to implement this project :

1. Fork the repository.

2. Put lyrics_scrapper_from_GENIUS_website.py and lyrics_dataset_creator.py in the same directory.

3. Head over to https://genius.com/api-clients/new , create your GENIUS account and then generate a client api and get your 'CLIENT ACCESS TOKEN'.

4. Install 'lyricsgenius' module in your local terminal. Execute lyrics_scrapper_from_GENIUS_website.py for each individual artist , to fetch their songs from The GENIUS website in a JSON file. (number of songs , title , popularity and be mentioned). Repeat execution for many artists.

5. Execute lyrics_dataset_creator.py to parse the JSON files which selects just the 'lyrics' key. It also creates the dataset named 'lyrics_dataset.txt' after doing some data-cleansing.

6. Run the Lyrics_Generation.ipynb notebook (preferably in Google Collab with GPU support for faster execution) , uploading the 'lyrics_dataset.txt' file , getting the new lyrics being generated.

7. Tweak the hyper-parameters according to "YOUR" best results

My dataset used to train and get the output is => lyrics_dataset.txt

Some lyrics generated :

`Love in the world`

`Happens each time to keep me warm`

`Love in the rain`

`And we'll stop eating food and I'll say the words that I want in this cold`

`I don't have to think that I want to do`

`I can't think of anyone, anyone else`

`And we gotta save ourselves`

`If we want it to be`

`I'll be taking my time`

`As she looks lovely in the wood`

This model can be used for any type of text prediction , not being limited to only lyrics.

Scope for improvements :

1. Using a LSTM-GAN based model, it has the best potential for most coherent output.

2. Adding another layer of Birectional-LSTM and lowering the EPOCHS keeping the loss within 1.5 to 1.1 , for most meaningful text generated.