In this project, I built a Recurrent Neural Network to generate my own Seinfeld TV scripts. The architecture uses 200 dimensions embedding layers at first, 2 stacked LSTM layers with 256 dimensions hidden state each, and a fully connected layer in the end. After model was trained, I used it to predict a new, "fake" TV script, based on patterns it recognized in this training data.
One main problem in this project is the hyper-parameters tuning. Because the dialogues are short, it is better to use a small sequence length. Also, if you are using a small dimension of embedding layers, do not use a high dropout probability like 0.5, because these word2vec contains valuable informations of the relationship between data.
The data for this project is located here, which has been crawled from the http://www.seinology.com/ website. I used part of the Seinfeld dataset of scripts from 9 seasons.
Python: 3.6
PyTorch: 0.4.0
GPU: NVIDIA GeForce GTX 970M
Windows: 8.1
jerry: well, i can't get this. i was thinking of a little problem.
jerry: oh, yeah, yeah.
kramer: well, i'm not going to be a little adjustment.
jerry: oh, yeah, i can't get this.
george: oh, no, no, no, no no no, no. no... i was thinking of that lady.
george: what do you mean?
hoyt: i don't want to get the money for you?
jerry: yeah.
jerry: you know what the hell is that?
hoyt: i don't know.
...
Check the dlnd_tv_script_generation.ipynb for all implementation.
Check the 5 example generated scripts in example folder.
The helper.py has functions for model saving and loading.
The problem_unittests.py has unit tests for functions.