spongebob transcript generator

Project Components

1. Beautiful soup web scraper that parses the fan wiki for episode transcripts. Located in directory web_scraper/

2. PyTorch code that processes a text file, trains a RNN, and saves checkpoints for predictions. Located in directory rnn/

• This folder has one rnnBob.py which has all the necessary code
• And one spongebob_txt_generation.ipynb which is a jupyter notebook that does some explaining of the code

4. Python script to predict text based on saved model located at rnn/predict_from_saved_model.py
5. 1000 words of predicted text plus an entire predicted script at rnn/example_predictions/

Example Prediction Text

SpongeBob: Hey, you can have the greatest thing you go, Mr. Krabs.

Squidward: You don't know. I have the little old little more thing is a few friend, I can't do it! I need the Krusty Krab!

Mr. Krabs: And you were the little idea! Mr. Krabs the animals are we have a bunch and you have you two idea! Mr. Krabs?

Squidward: And that's the Krusty Krab. I got you two one more idea!

SpongeBob: Oh, that's not a great idea!

SpongeBob: Hey, hey, hey, hey, that's the Krusty Krab.

Squidward: You don't know.

SpongeBob: And this isn't the little animals is not to do it. I got you were to get to my old new king and you can have a great of the king is you can do this.

Project Notes

• PyTorch was significantly easier to use than I imagined. I spent a lot of time learning about recurrent neural networks and still trying to grasp the math but when it came down to it, pytorch made the hard parts easy.
• Training was significantly slower than I imagined on my desktop. I use a budget machine with no standalone graphics card. Let the model train for over 24 hours and got some interesting results, but only made it 1/10 through the total training loop. Easy to see how machines with graphics cards dedicated for this might greatly improve the experience of a NLP dev. Probably not going to let the model train for 10 days.