Adding the ability to "prime" the network with an author's style

Grzego / handwriting-generation

Implementation of handwriting generation with use of recurrent neural networks in tensorflow. Based on Alex Graves paper (https://arxiv.org/abs/1308.0850).

MIT License

519 stars 107 forks source link

Adding the ability to "prime" the network with an author's style #4

Closed kristofbc closed 6 years ago

kristofbc commented 6 years ago

Hello! I've added the ability to prime the network with eight pre-defined styles (you can see them in the image folder). The implementation of "priming" is the one described in Graves's paper: join a character sequence and null vector of pen positions to a real character sequence and its associated pen positions. The output is then clipped because we only keep the synthetic sequence.

Grzego commented 6 years ago

Hi!

Thanks for PR! I've spotted some minor bugs in code, like not cutting of the priming sequence from synthetic one (I will push fixes after merging your PR).

Thanks again! :)

kristofbc commented 6 years ago

Sure, no problem, glad I could help :). The general idea behind priming is there, feel free to change stuff!