This project demonstrates the use of RNNs for script generation. It will create a new "fake" TV script based on the 9 seasons of Seinfeld TV serries.
Install PyTorch (I used v1.2.0 for development):
pip install torch
Clone this repo:
git clone https://github.com/puhach/tv-script-gen.git
To start generating a new script specify the first word (should be the name of a character) and the script length:
scriptgen.py --prime-word elaine --script-len 333
Unless the model has already been trained, it will start the training process which make take a long time. Once training is finished, model's checkpoint is saved to 'train_rnn.pt' file and the script will be generated.
Further script generation will be using the pre-trained model, so the script should be printed shortly. In order to retrain the model, delete the 'trained_rnn.pt' file and launch generation with a different set of parameters.
| Parameter | Meaning | Example |
|---|---|---|
| prime-word | [Required] The first word of a new script (the name of a character). | --prime-word jerry |
| script-len | [Required] The length of a script to generate. | --script-len 123 |
| n-epochs | [Optional] The number of epochs to train the model for. Default is 2. | --n-epochs 10 |
| learning-rate | [Optional] The learning rate for gradient descent. Default is 0.0001. | --learning-rate 0.01 |
| sequence-len | [Optional] The number of words in a sequence. Default is 16. | --sequence-len 16 |
| batch-size | [Optional] The number of word sequences in a batch. Default is 32. | --batch-size 32 |
| n-layers | [Optional] The number of layers in the recurrent neural network. Default is 2. | --n-layers 2 |
| embedding-dim | [Optional] The embedding dimension. Default is 256. | --embedding-dim 128 |
| hidden-dim | [Optional] The hidden dimension of the recurrent neural network. Default is 512. | --hidden-dim 256 |
| stat-freq | [Optional] Prints loss statistics after the specified number of batches have been processed. Default is 200. | --stat-freq 100 |
elaine:.............(runs over to leave the door)
jerry:(to kramer) i can't go on the middle machine!
jerry: hey, that's right, i don't think so.
george: i know what i said, you know, i can't have to take it out. i don't want you to have to be any with me.
kramer:(frustrated) yeah, yeah. yeah.(jerry hangs up) oh, yeah.(to the bathroom.)
jerry:(pointing) oh, hi.
newman:(grabs) i told you i have to pick the other day.
elaine: i don't know. i think i know.
george: i got it.
jerry:(trying to himself) what do you think?
jerry: what are you doing?
elaine: i don't know what you think i was going to get it on your friend.
george: i got it.
jerry: i don't think i'm gonna get it.
george: no. no, no, no, no, it's a fine.
george: what are you doing here, you know?
jerry: what?
jerry: i don't know.
george: what?
jerry: what happened is a second.
kramer:(looking) oh, you don't know what you can take my god.
kramer:(to jerry) hey, what are you talking about?
jerry: i know what i was going to the one.
george: well, i know.
jerry:(looking at the table) you know what
The project was implemented as a part of Udacity Deep Learning nanodegree.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent