AutoShakespeare

This program generates shakespear-style text from a piece of starting text that you provide. This is a pytorch implementation of the tensorflow tutorial on generating text with an RNN. The shakespeare.txt file is used as the dataset for training.

Usage

You just need to run main.py script to run the program. Here are the arguments the program can receive:

usage:
       [-h] [--starting_text STARTING_TEXT] [--output_size OUTPUT_SIZE]
       [--use_pretrained {none,rnn,gru,lstm}]
       [--recurrent_module {rnn,gru,lstm}] [--dropout DROPOUT]

optional arguments:
  -h, --help            show this help message and exit
  --starting_text STARTING_TEXT, -s STARTING_TEXT
                        piece of text used as starting point for generation
  --output_size OUTPUT_SIZE, -o OUTPUT_SIZE
                        length of text to generate (in characters)
  --use_pretrained {none,rnn,gru,lstm}
                        which pretrained model to use?
  --recurrent_module {rnn,gru,lstm}, -r {rnn,gru,lstm}
                        type of recurrent module used for training
  --dropout DROPOUT, -d DROPOUT
                        how much dropout to use for training (between 0 and 1)

If --use_pretrained is none, then a new model will be trained before generation begins, otherwise the specified pre-trained model is used.

Dependencies

* PyTorch - 1.5
* matplotlib - 3.2
* numpy - 1.18

The versions provided above are those the program was tested with. The dependency is very light, however, since the features used are very basic.

Sample Output

Here are three 500-character long texts generated by the pretrained models, using MENENIUS as the starting text:

RNN:

MENENIUS:
Ha! Marcius coming home!

VOLUMNIA:
Ay, worthy Menenius; and with most prosperous
approbate
approbied is not unknown to the senate; they have
had inkling this fortnight what we interior survey of your good selves!
O that you could!

BRUTUS:
What then?
'Fore me, this fellow speaks! who come to destre consuberand, if they are trust?

COMINIUS:
As I guess, Marcius,
Their bands i' the vaward are the Antiate, who cell not
Be grafted to have the general food at will with me trigure
There as 

LSTM:

MENENIUS:
Ha! Marcius coming home!

VOLUMNIA:
Ay, worthy Menenius; and with most prosperous
approbate in this afternoob.

MENENIUS:
There was a netth doust;
And make my wars on you: look to't: come on;
If not this sound arm that have
fall enemy take him seek
danger where he was like to find fame. To a cruel
war I sent him; from whence he receive
But it proceed an hour timps: the noble senate, who,
answere yet deed to doubh to your worships have
delivered the matter well, when I find the air of t

GRU:

MENENIUS:
O, true-bred!

First Senator:
Your company to the Capitol; where, I know,
Our greatest friends! Come, blow thy blast.
Tutus Aufidius, is he within your walls?

First Senator:
Speak, good Cominius:
Leave nothing out for length, and his actions
in their hearts, that for their tongues to be
silent, and not confess shall plaines; and their pluckings; well forth proverbs,
That hunger broke stone walls, that dogs must eat,
That meat fit in it.

MENENIUS:
A lay our good word, that is as enemy

It can be seen that all three models have a solid understanding of the structure of a play: they keep creating blocks of text which are preceded by the name of the speaker in all-capitals (except when there's a title instead of a name, in which case it is merely capitalized). Also, most of the sentences are grammatical. But, since the model is character-based, it hasn't understood the meaning of words, and as a result most sentences are semantically incorrect.

It can also be observed that the first few sentences are always memorized from the dataset, but as the model progress onwards, it starts becoming more spontaneous and generates novel sentences of its own.

Authors

• Ahmad Pourihosseini - ahmad-PH

Acknowledgments

• The idea of this project is inspired by the tensorflow tutorial on generating text with an RNN and that is also were the dataset is downloaded from.