Nice! Did you find any weird/funny case while training the model? [I remember autotranlation/readers inventing stuff a few years ago.]
> Drive a real mechanical pen
> I think it would be fun to hook the output of this model up to a real pen. I'm not sure of the use case, but maybe you're an artist and have some ideas?
If you get someone to make a few samples and photos, it would be very nice.
> 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
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.
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.
> Drive a real mechanical pen
> I think it would be fun to hook the output of this model up to a real pen. I'm not sure of the use case, but maybe you're an artist and have some ideas?
If you get someone to make a few samples and photos, it would be very nice.
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
P.S. I checked the issue: when I input '다.', it outputs '다', but when I input '다', it outputs '디'.
PS: I saw your feedback report. 화이팅! :)