This project includes my solution for TV script generation project in Udacity deep learning course. Instructions for set up can be found in its original repo
This project builds a RNN model training on 9 seasons of Seinfeld dataset to generate "fake" TV script.
There are a few steps we take to prepare the training data from the raw TV script:
- tokenize punctuation (convert punctuation as an individual word)
- collect vocabulary of words from the data set and convert each word into a int index in the vocabulary
- Given a sequence length, the input features are the number of sequence length words, and the prediction target is the word after.
The model is implemented through pytorch, constructed by an embedding layer, a LSTM layer with 2 hidden layers, and a fully connected linear layer to generate the final output.
Cross Entropy.
We use "jerry" as start word and use our model to generate a script with length of 400 words to give people more idea about how well the model is performing currently:
jerry: 13.
hoyt: so?
jerry: oh, come on.
jerry: hey!
hoyt: so what are you doing?
kramer: well, it's a very dangerous wind.
hoyt: and basically the defendants- less frosting.
chiles: objection overruled?
george: i was carjacked.
kramer: hey, what do you know?
kramer: yeah yeah, well...
hoyt: so, essentially, you know what?
george: yeah, yeah.
hoyt: and you don't want to be the only thing i've ever had, but i was on the plane.
george: i think i know how to make it out.
chiles: so what?
george: i don't know what it means.
jerry: you know, this is elizabeth 4th.
george: i know.
hoyt: so, you know, i don't want to talk to jill about this testimony.
jerry: what?
hoyt: this flavor is killing me.
hoyt: what do you think?
elaine: no, it's not.
jerry: what?
jerry: i don't want to talk to you.
george: what?
hoyt: and the judge is a saab-- nguh war- absorption- absorption?!
george: no.
hoyt: no, no, no. no, no, no. no. donald, you know, the bubble boy is the nature- tub.
[new witness: police detective].
hoyt: so, what do ya mean?
jerry: no. i can't believe that.
[new witness: police detective], 000 crespi?
hoyt: no further questions?
chiles: i think it's a saab. i was wondering if you can be a kid.
george: i can't.
chiles: you know what?
chiles: objection stu.
george: what?
jerry: oh, yeah, yeah. i was carjacked, and i was in
As shown above, most of the sentences are complete and makes sense. It does look like tv script structurally, though many of the conversation turns look weird and far from a real conversations on TV.