Skip to content

Latest commit

 

History

History
115 lines (102 loc) · 5.93 KB

Dataset.md

File metadata and controls

115 lines (102 loc) · 5.93 KB
Raw

Instruction

The examples of datasets are in (./data_example) Note that, we only provide the example font files; not the font files used for the training the provided weight (generator.pth). The example font files are downloaded from here(chinese) and here(korean).

  • This code can treat both dataset in image files(.png, .jpg ...) and truetype font files(.ttf).
  • To use ttf data, see 1.1.
  • To use image data, see 1.2.

1. Prepare dataset

1.1. Prepare TTF files

  • Prepare the TrueType font files(.ttf) to use for the training and the validation.

  • Put the training font files and validation font files into separate directories.

  • In this case, the available characters list is needed in txt format. (described in 1.1.a.)

  • If you want to split characters to seen characters and unseen characters, the character split information is needed in json format. (described in 2.)

    1.1.a. Preparing the available characters list

    • If you have the available character list of a .ttf file, save its available characters list to a text file (.txt) with the same name in the same directory with the ttf file.
      • (example)
      • TTF file: data/ttfs/train/MaShanZheng-Regular.ttf
      • its available characters: data/ttfs/train/MaShanZheng-Regular.txt
    • You can also generate the available characters files automatically using the get_chars_from_ttf.py
    • How to use:
      • python get_chars_from_ttf.py path/to/ttf/dir
      • path/to/ttf/dir: The root directory to find the .ttf files. All the .ttf files under this directory and its subdirectories will be processed.

1.2. Prepare Image files

  • The images are should be placed in this format:
    * data_dir
    |-- font1
        |-- char1.png
        |-- char2.png
        |-- char3.png
    |-- font2
        |-- char1.png
        |-- char2.png
            .
            .
            .
  • You can see the example at data_example/chn/png.
  • The images with the same style are should be grouped with the same directory.
  • The name of each image file should be its character.

2. Split characters to train and validation set

(required for 1.1, optional for 1.2)

  • Save the list of characters to use for the training as a json file.
    • This step can be skipped if you want to use all the available characters without splitting for the training.
    • The characters both existing in this list and available from the dataset will be used for the training.
    • Our example is in data/chn/train_chars.json(Chinese) and data/kor/train_chars.json(Korean)
  • Save the list of characters to use for the validation as a json file.
    • The characters both existing in this list and available from the dataset will be used for the validation.
    • The characters that we used for the training is in data/chn/val_unseen_chars.json(Chinese) and data/kor/val_unseen_chars.json(Korean)

3. Prepare decomposition information

(not required for FUNIT)

  • The files that we used are provided.

    • Chinese: data/chn/decomposition.json, data/chn/primals.json
    • Korean: data/kor/decomposition.json, data/kor/primals.json, data/kor/decomposition_DM.json(only for DM-Font)

  • See 3.1 to build your own decomposition file:

    3.1 The format of decomposition rule and primals file

    • Decomposition rule
      • structure: dict (in json format)
      • format: {char: [list of components]}
      • example: {'㐬': ['亠', '厶', '川'], '㐭': ['亠', '囗', '口']}
    • Primals
      • structure: list (in json format)
      • format: [All the components in the decomposition rule file]
      • example: ['亠', '厶', '川', '囗', '口']

4. Prepare source font or source images

(not required for DM-Font)

4.1 Using ttf file as source

* You can get the Chinese source font which we used from [here](https://chinesefontdesign.com/font-housekeeper-song-ming-typeface-chinese-font-simplified-chinese-fonts.html).
  * Download this file and save to `data/chn/source.ttf`.
* The Korean source font is attached at `data/kor/source.ttf`

4.2 Using image files as source

* Put all the source image files in a single directory. 
* The name of each image file should be its character.

5. Modify the data configuration file (cfgs/data/train/custom.yaml)

Please do not modify the indentation, because the indentation rule is very important in these configuration files.

  • dset: (leave blank)
    • train: (leave blank)
      • data_dir: path to training data.
      • extension: extension of training data.
        • Set this to "ttf" if you are using ttf files.
        • If you are using image files set this to the image files' extension.
      • chars: The json file containing the training characters.
        • If this is blank, all the available characters will be used for the training.
        • If you generated the file in step 2, set this to its location.
    • val: (leave blank)
      • name: Anything is okay; this will be the name of the validation set.
        • data_dir : path to validation data.
        • extension: extension of validation data.
        • n_gen: How many characters to generate during validation. The characters will be randomly selected from characters.
        • n_font How many fonts to generate during validation. The fonts will be randomly selected from available fonts.
        • chars: The json file containing the entire validation characters.
          • If this is blank, the validation characters will be chosen from all the available characters.
        • source_path : path to the source font or source directory to use for the validation.
        • source_ext: extension of the source data.
          • If you are using a ttf file, set this to "ttf".
          • If you are using image files, set this to their extension ("png", "jpg" ...).