FilmyKeeda

A simple text generation package aimed to allow a user to easily explore results produced by State of the Art NLP models presented to us in the last few months.

The package currently supports two models -

1. ULMFiT - From Fast.AI
2. GPT2 - From OpenAI

Download -

pip install filmykeeda

Usage -

This repository currently holds 4 example.py files demonstrating the use of this package.

example0.py - This will demonstrate on how you can download the corpus

example1.py - This will demonstrate on how to generate script using ULMFiT + Default Tokenizer

example2.py - This will demonstrate on how to generate script using ULMFiT + SentencePiece Tokenizer

example3.py - This will demonstrate on how to generate script using GPT2

Data Used -

Both the models have been trained on a extensive corpus of scripts written by Aaron Sorkin. The works included in this project are -

• The West Wing (Series)
• The Social Network (Movie)
• A Few Good Men (Movie)
• The American President (Movie)

Model Details -

ULMFiT

ULMFiT Model has been trained using the FastAI library.

ULMFiT allows the user to train a model using a custom tokenizer and 
therefore this package includes two different ULMFiT models -

        1. Trained with default FastAI Tokenizer
        2. Trained with SentencePiece Tokenizer

GPT2

GPT2 Model has been trained using the gpt_2_simple library

GPT2 does not allow an external tokenizer to be used and hence
the model has been simply finetuned to our corpus.

TODO -

• Package this project as PIP Library
• Add evaluation scheme's for generated scripts such as ROUGE and Perplexity
• Clean ULMFiT generated script
• User to have the ability to train his model