2017
2017 copied to clipboard
Something Something [char-rnn-tensorflow]
I decided to try using char-rnn-tensorflow and I've been downloading and cleaning up a selection of books from Project Gutenberg as training input.
Ingredients so far:
- Isaac Asimov
- Jane Austen
- Ray Bradbury
- Philip K. Dick
- Charles Dickens
- Frank Herbert
- Herman Hesse
- Communist Manifesto
- H.G. Wells
Just finished training on 5MB of text. I think I'm going to add some more books, remove others, do a bit more manual cleaning of the input text and then re-train.
Here are some fun samples:
- my face. A matter rose, imagining, had verial to my faculty.
- the darkening of never was a full of few of them?
- Why, the evils of worth.
- He seems his man and the advantage of trap it. The glorious. He busted wis which sher, but set and in things. But she broke men.
UPDATE 1:
I removed war of the worlds, added some childrens books and cleaned up the source text a little more.
Here are some samples:
- "There blessed to three wonderful meat! Don’t be from his sister.”
- Miss Lucy and Mrs. Smally from the sauce.
UPDATE 2:
Added the Sherlock Holmes books and after retraining, things got pretty weird. Maybe because of the size of the input and formatting variations? I cleaned up the text more, removing margins (they were only in the sherlock holmes books) and Chapter/ Part headers and going to re-train again.
Trying to follow some general tips from Karpathy. I'm using a GTX1080 so I increased the rnn_size
to 700, num_layers
to 3 to try to improve the model and increased the batch_size
to 1000 to speed up the process.
Samples are much more interesting:
- Then I will not ask what you have; but I am very sorry for my plagued questions.
- The bear was delightedly round, when very free trails of debris were to go, could eat all the blackened slopes of strength to do something like blood from the fireplace.