ScrabbleGAN: Semi-Supervised Varying Length Handwritten Text Generation

This is a pytorch implementation of the paper "ScrabbleGAN: Semi-Supervised Varying Length Handwritten Text Generation"

Dependency

• This work was tested with PyTorch 1.2.0, CUDA 9.0, python 3.6 and Ubuntu 16.04.
• requirements can be found in the file environmentPytorch12.yml. The command to create the environment from the file is: conda env create --name pytorch1.2 --file=environmentPytorch12.yml
• To activate the environment use: source activate pytorch1.2

Training

• To view the results during the training process, you need to set up a visdom port: visdom -port 8192

Supervised Training

python train.py --name_prefix demo --dataname RIMEScharH32W16 --capitalize --display_port 8192 

• Main arguments:
  • --name: unless specified in the arguments, the experiment name is determined by the name_prefix, the dataset and parameters different from the default ones (see code in options/base_options.py).
  • --name_prefix: the prefix to the automatically generated experiment name.
  • --dataname: name of dataset which will determine the dataroot path according to data/dataset_catalog.py
  • --lex: the lexicon used to generate the fake images. There is a default lexicon for english/french data specified in options/base_options.py.
  • --capitalize: randomly capitalize first letters of words in the lexicon used.
  • --display_port: visdom display port
  • --checkpoints_dir: the networks weights and sample images are saved to checkpoints_dir/experiment_name.
  • --use_rnn: whether to use LSTM
  • --seed: determine the seed for numpy and pytorch instead of using a random one.
  • --gb_alpha: the balance between the recognizer and discriminator loss. Higher alpha means larger weight for the recognizer.
• Other arguments are explained in the file options/base_options.py and options/train_options.py.

Semi-Supervised Training

python train_semi_supervised.py --dataname IAMcharH32W16rmPunct --unlabeled_dataname CVLtrH32 --disjoint

• Main arguments:

  • --dataname: name of dataset which will determine the labeled dataroot path according to data/dataset_catalog.py. This data is used to train only the Recognizer (in the disjoint case) or the Recognizer and the Discriminator networks.
  • --unlabeled_dataname: name of dataset which will determine the unlabeled dataroot path according to data/dataset_catalog.py. This data is used to train only the Discriminator network.
  • --disjoint: Disjoint training of the discriminator and the recognizer (each sees only the unlabeled/labeled data accordingly).

• Other arguments are explained in the file options/base_options.py and options/train_options.py.

LMDB file generation for training data

Before generating an LMDB download the desired dataset into Datasets:

• IAM
  • Download the dataset from http://www.fki.inf.unibe.ch/databases/iam-handwriting-database
  • The partition into directories was downloaded from https://github.com/jpuigcerver/Laia/tree/master/egs/iam/data/part/lines/original.
• RIMES - download the dataset from http://www.a2ialab.com/doku.php?id=rimes_database:start
• CVL - download the dataset from https://cvl.tuwien.ac.at/research/cvl-databases/an-off-line-database-for-writer-retrieval-writer-identification-and-word-spotting/ The dictionaries used for training in the paper were not generated from the datasets, but rather downloaded and saved to the Datasets/lexicon.
• english_words.txt - https://github.com/dwyl/english-words/blob/master/words.txt
• Lexique383.tsv - https://github.com/AdrienVannson/Decorrecteur/blob/master/Lexique383

The structure of the directories should be:

• Datasets
  • IAM
    • wordImages (the downloaded words dataset)
    • lineImages (the downloaded lines dataset)
    • original (the downloaded xml labels data)
    • original_partition (the downloaded partition)
      • te.lst
      • tr.lst
      • va1.lst
      • va2.lst
  • RIMES
    • orig (the downloaded dataset)
      • training_WR
      • groundtruth_training_icdar2011.txt
      • testdataset_ICDAR
      • ground_truth_test_icdar2011.txt
      • valdataset_ICDAR
      • ground_truth_validation_icdar2011.txt
  • CVL
    • cvl-database-1-1 (the downloaded dataset)
      • trainset
      • testset
      • readme.txt
  • Lexicon
    • english_words.txt
    • Lexique383.tsv

To generate an LMDB file of one of the datasets CVL/IAM/RIMES/GW for training use the code:

cd data
python create_text_data.py

• Main arguments (determined inside the file):
  • create_Dict = False: create a dictionary of the generated dataset
  • dataset = 'IAM': CVL/IAM/RIMES/gw
  • mode = 'va2': tr/te/va1/va2/all
  • labeled = True: save the labels of the images or not.
  • top_dir = 'Datasets': The directory containing the folders with the different datasets.
  • words = False: parameter relevant for IAM/RIMES. Use words images, otherwise use lines
  • parameters relevant for IAM:
  • offline = True: use offline images
  • author_number = -1: use only images of a specific writer. If the value is -1, use all writers, otherwise use the index of this specific writer
  • remove_punc = True: remove images which include only one punctuation mark from the list ['.', '', ',', '"', "'", '(', ')', ':', ';', '!']
  • resize parameters:
  • resize='noResize': charResize|keepRatio|noResize - type of resize, char - resize so that each character's width will be in a specific range (inside this range the width will be chosen randomly), keepRatio - resize to a specific image height while keeping the height-width aspect-ratio the same. noResize - do not resize the image
  • imgH = 32: height of the resized image
  • init_gap = 0: insert a gap before the beginning of the text with this number of pixels
  • charmaxW = 18: The maximum character width
  • charminW = 10: The minimum character width
  • h_gap = 0: Insert a gap below and above the text
  • discard_wide = True: Discard images which have a character width 3 times larger than the maximum allowed character size (instead of resizing them) - this helps discard outlier images
  • discard_narr = True: Discard images which have a character width 3 times smaller than the minimum allowed charcter size.

The generated lmdb will be saved in the relevant dataset folder and the dictionary with be saved in Lexicon folder.

Generating an LMDB file with GAN data

python generate_wordsLMDB.py --dataname IAMcharH32rmPunct --results_dir ./lmdb_files/IAM_concat --n_synth 100,200 --name model_name 

• Main arguments:
  • --dataname: name of dataset which will determine the dataroot path according to data/dataset_catalog.py. note that will be concatenated to the generated image.
  • --no_concat_dataset: ignore “dataname” (previous parameter), do not concatenate
  • --results_dir: path to result, will be concatenated with "n_synth"
  • --n_synth: number of examples to generate in thousands
  • --name: name of model used to generate the images
  • --lex: lexicon used to generate the images

Main Folders

The structure of the code is based on the structure of the CycleGAN code.

1. data/ - Folder containing functions relating to the data, including generation, dataloading, alphabetes and a catalog which translates dataset names into folder location. The dataset_catalog should be updated according to the path to the lmdb you are using.
2. models/ - Folder containing the models (with the forward, backward and optimization functions) and the network architectures. The generator and discriminator architectures are based on BigGAN. The recognizer architecture is based on crnn.
3. options/ - Files containing the arguments for the training and data generation process.
4. plots/ - Python notebook files with visualizations of the data.
5. util/ - General function that are used in packages such as loss definitions.

Citation

If you use this code for your research, please cite our paper.

@inproceedings{fogel2020scrabblegan,
    title={ScrabbleGAN: Semi-Supervised Varying Length Handwritten Text Generation},
    author={Sharon Fogel and Hadar Averbuch-Elor and Sarel Cohen and Shai Mazor and Roee Litman},
    booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
    month = {June},
    year = {2020}
}

License

ScrabbleGAN is released under the MIT license. See the LICENSE and THIRD-PARTY-NOTICES.txt files for more information.

Contributing

Your contributions are welcome!
See CONTRIBUTING.md and CODE_OF_CONDUCT.md for more info.