In Codice Ratio

Synthetic dataset generation for sequence prediction models

Requirements

• numpy
• cv2
• networkx
• matplotlib

Usage

Set the dataset, corpus and destination paths in the generate_textlines.py main, then run it to generate synthetic line images and their transcription.

Dataset folder expects structure dataset_folder/{character classes}/character_images.png.

Corpus folder expects structure corpus_folder/{text files}.txt

Files in the destination folder will be of the type destination_folder/{i.png, i.txt} for each line generated.

json file abbr_matchings.json maps text to sequences of symbols in the dataset.

Model training

Requirements

• tensor2tensor

Usage

Preprocess generated data (synthetic dataset in the form of i.png, i.txt couples must be in $TMP_DIR/ocr) and put it into $DATA_DIR, using custom problem definition (in t2t_usr):

$ t2t-datagen \
    --t2t_usr_dir=t2t_usr \
    --problem=ocr_latin \
    --tmp_dir=$TMP_DIR \
    --data_dir=$DATA_DIR

Train the transformer_sketch model on the generated dataset, using custom problem definition (in t2t_usr):

$ t2t-trainer \
    --t2t_usr_dir=t2t_usr \
    --problem=ocr_latin \
    --tmp_dir=$TMP_DIR/ocr \
    --data_dir=$DATA_DIR \
    --model=transformer_sketch \
    --hparams_set=transformer_small_sketch \
    --output_dir=$OUTPUT_DIR

Author

This project was developed by Elena Nieddu during Pi School's AI programme in Fall 2017.