Show HN: Happy Hangul Day! An RNN for Generating Korean Handwriting Strokes

Thank you for the feedback. I’ll incorporate this into the new training run.

I’m sorry about your father. Reproducing his handwriting would be a beautiful project. What kind of samples do you have from him?

Thank you, you're right! It definitely seems to have a problem when the end of the string ends with 다. It seems to work if you end the sentence with "다." instead of "다". I'll look into this more. There should be plenty of "다" data in the training set, maybe there's something deeper going on.

PS: I saw your feedback report. 화이팅! :)

> Nice! Did you find any weird/funny case while training the model? [I remember autotranlation/readers inventing stuff a few years ago.]

Mostly, it was me trying things the naive way, running into problems, and then solving them, usually by copying what Alex Graves did in 2013 (but not always!)

For example, to make things simple I first built the model to simply predict the next x, y direction of the pen. This worked for simple pen strokes, but I noticed the model had a difficult time turning corners.

To fix this, I changed the model to instead predict an Mixture Density Network (Same as Graves). This is explained better in his paper, but essentially, instead of predicting one x, y direction, you predict 10, and then randomly sample one of those predictions. Also instead of predicting scalar x, y values, you predict a parameters for a gaussian distribution, and then sample from that.

It always amazes me how much randomness is involved in intelligence.

> If you get someone to make a few samples and photos, it would be very nice.

I'm in Seoul, if you know anyone who would be interested in collaborating, please send me a note! jon@jonb.org

Noticed that 펲시 triggers odd behavior.

I think this would work well for Chinese, but you’d probably need to modify the text embedding system. Let me know if you’d like to try, happy to collaborate.

The text embedding structure is very specific to Hangul. Someone that knows more about Chinese than me could probably design a suitable text embedding structure that would work well. I’m happy to collaborate if you want to fork my project and try.

That's an interesting idea. Velocity is something that can be derived from the x, y inputs that the model already has. So, if the velocity contained useful information, I would expect the model to be able to learn this during the training phase. The pressure, on the other hand, was new information that the model didn't have before.

Oh thank you, good catch. "쀍" is definitely not a common character in my training set. I'll add it to the next run!