tv-script-generator

Example of using LSTM RNN network to generate random TV scripts. In this project the Neural Network model will generate scripts for a popular TV series, Seinfeld. It gets trained from the scripts of season 9 and try to generate some randomd dialogues. Question is how much those random dialogues makes sense!

Notebook

dlnd_tv_script_generation.ipynb is the main notebook.

Required files

• Seinfeld season 9 script (https://www.kaggle.com/thec03u5/seinfeld-chronicles#scripts.csv)
• helper.py (originally supplied in Udacity project)
• problem_unittest.py (originally supplied in Udacity project)

Glimpse at the generated script

Following is a part of script generated by model. The script is saved in generated_script_1.txt file. Here goes a snippet of that,

elaine:(leaving) hey.
jerry: what is that?
elaine: what happened to the last story?
george: i thought i could do this.
kramer: i think i wanted a little effeminate.
george: i have to say that.
george: i was convinced.
jerry: yeah, i thought we were gotten dating in the contest.
elaine: i was afraid i sent to say you on.
elaine: hey!
george: what?
george: i mean, i don't understand.(starts opening the sleeve employee, and the lopper skank, and i was convinced.

You see the generated script is not contextually perfect but the sentence constructions are not that bad.

Steps I followed

The steps were instructed by the Udacity course instructor. My responsibility was to implement the corresponding functions.

create_lookup_tables

Creates a numeric identifier of the vocabulary and returns two follwing maps,

1. word -> identifier
2. identifier -> word

token_lookup

Was to replace the puctuation with special keywords.

batch_data

Convert the whole training script into training dataset which will be fed to the model in multiple batches. Based on the given sequence_length I use DataLoader from pytorch to create the batches

RNN Model

The Neural Network model comprises:

• Embedding layer
• LSTM RNN layer
• Linear classifcation layer In the notebook you will see how I tried with different hyperparamters. The best result produced a loss score 2.88 over the last batch.

generate

Finally used the generate function to generate another random script by using the trained model.