This example implements a character-level language model heavily inspired by char-rnn.
- For training, the model takes as input a text file and learns to predict the next character following a given sequence.
- For generation, the model uses a fixed seed, returns the next character distribution from which a character is randomly sampled and this is iterated to generate a text.
At the end of each training epoch, some sample text is generated and printed.
Any text file can be used as an input, as long as it's large enough for training.
A typical example would be the
tiny Shakespeare dataset.
The training text file should be stored in data/char-rnn/input.txt.
To run the example:
cd example/char-rnn
go run .
Here is an example of generated data when training on the Shakespeare dataset after only 5 epochs.
rom Dogs on the gamper, as their rear and
A king your revenues Numilitul. Your hanged
like awaket from me Mucafion as even sit best.
HENRY BOLINGBROKE:
O, I know'st me not his princess
RIVERS:
Bidgened Walter, march is their beadeful,
To be full yiel successes him a brother
And treason wid ought to do the haughty likes
Keep lay issued-formoners?
LUCIO:
Nay, but he's very were I live, I have seen.
KING EDWARD IV:
Be indeed.
What? what new and past your knee, how now!
LADY GREY:
To their own lidight right.
Will you joy?
MERCUTIO:
'Tis it that dear Bianca and York; some it
But my consent. What's the king so plant! 'LYears,
My personage of thy Lady Grumio!
KATHARINA:
Reblong, my lord;
For-sun, therefore thou high'd me to me.
ESCALUS:
But they are gone.
Second Servant:
O, they do seek out the day.
CLARENCE:
Alas; for father, lords, take they be little state,
His business: and therefore fit it should
Being well seating lie out.
Second Servant:
I know no one of purpose.
CAPULET:
Make me us.