Font Design GAN

Code for training/generating GAN, that is for font design.

Samples

Generated fonts, they have various styles and they are readable as a character.

Random walk in style-input space. You can watch gradually transformed fonts.

How to use

Environment

The auther checked running codes in only following environment:

• Ubuntu 16.04 LTS
• GeForce GTX 1080 & Driver (Driver Version: 384.111)
• CUDA 8.0
• cuDNN v6.0 (Membership required)

If you'd like to run with latest environment, revise some files as necessary.

Firstly clone this repository. Add --recursive because this repository contains submodule (font2img)

git clone --recursive https://github.com/uchidalab/fontdesign_gan
cd fontdesign_gan

with Docker (GPU required)

If you use GPU and Docker, it's easy to set up your environment. Requirement libraries are written in Dockerfile.
Install GPU Driver/NVIDIA Docker and run following commands:

docker build -t fontdesign_gan .
docker run --runtime=nvidia -it --rm -p 6006:6006 --volume `pwd`:/workdir -w /workdir/ fontdesign_gan

no Docker

with GPU: Install GPU Driver/CUDA/cuDNN and run following command:

pip install -r requirements_gpu.txt

with CPU: Run following command:

pip install -r requirements_cpu.txt

Prepare

Convert font files (.ttf) into image files (.png), and pack them into a HDF5 file (.h5)

Make a directory (ex. ./ttfs) and copy font files. Set destination path (ex. ./src/myfonts.h5) and run.

python main.py --ttf2png --png2h5 --font_ttfs ./ttfs --font_h5 ./src/myfonts.h5

--ttf2png is the option for converting, image files are saved in ./src/pngs/{YYYY-MM-DD_HHmmss}. If you use --font_pngs option, you can set path yourself.
--png2h5 is the option for packing, packed file is saved in the path you set with --font_h5 option.

Train

Set packed file's path and run.

python main.py --train --font_h5 ./src/myfonts.h5

Results are saved in ./result/{YYYY-MM-DD_HHmmss}. You can set destination with --gan_dir.

In log directory, saved same files: flags' log, latest/kept TensorFlow's dumps (.ckpt*), TensorBoard's log.
In sample directory, fonts that generated temporary are saved.

While training, TensorBoard is also running. Access URL that will shown in command line.

Generate

Set a path of trained GAN's output directory and a JSON file.

The JSON file have to be written style/character IDs. A style ID correspond with a random matrix. A character ID correspond with a character.

IDs are loaded in order, and you can use some operands. For example, if JSON file is following:

{
  "style_ids": [
    "0", "5", "21", "37", "17..39:4"
  ],
  "char_ids": [
    "0-3", "6*4"
  ],
  "col_n": 4
}

Input IDs will be like this:

style_ids = [0, 5, 21, 37, (between 17 and 39 with 4 steps)]
char_ids = [0, 1, 2, 3, 6, 6, 6, 6]

and, # of result columns is 4. Also check sample files (./jsons/sample*.json).

After preparing JSON file, run like this:

python main.py --generate --gan_dir ./result/{trained} --ids ./jsons/sample01.json

Generated fonts are saved in ./result/{trained}/generated/{YYYY-MM-DD_HHmmss}.png. You can set output file name with --gen_name.

Generated image is like this:

If you want random walking fonts, use --generate_walk. A JSON file is needless.

python main.py --generate_walk --gan_dir ./result/{YYYY-MM-DD_HHmmss} --char_img_n 256

256 styles' fonts will be generated, and they are transformed gradually.

Options

There are many options. Check following table.

For	Option	Description	Default
Prepare, Train	--img_width	width of images.	64
Prepare, Train	--img_height	height of images.	64
Prepare, Train	--img_dim	dimension of images.	3
Prepare, Train	--chars_type	you can choose characters type, "caps", "hiragana" or "caps,hiragana"	"caps"
Train, Generate	--gpu_ids	GPU IDs you use. this type is string (ex. "0, 1")	(set automatically)
Train, Generate	--batch_size	batch size for GAN	256
Train	--arch	architecture of GAN models. choose "DCGAN" or "ResNet"	"DCGAN"
Train	--style_ids_n	# of style IDs	256
Train	--style_z_size	size of style_z	100
Train	--gan_epoch_n	# of epoch iterations	10000
Train	--critic_n	# of critic iterations	5
Train	--sample_imgs_interval	interval of saving sample images	10
Train	--sample_col_n	# of sample image's columns	26
Train	--keep_ckpt_interval	interval of keeping TensorFlow's dumps	250
Train	--run_tensorboard	run tensorboard or not	True
Train	--tensorboard_port	port for tensorboard page	6006