TraTra

Setup

Environment

• Set up the pipenv environment via the following command

$ pipenv install --skip-lock --dev

• Download IAMONLINE dataset and extract it to PoC_TraTra/data/iam_online/.

└── iam_online
    ├── ascii
    ├── gen_iam_256 ( <- we'll generate this later)
    ├── lineImages
    ├── lineStrokes
    ├── org_unp ( <- we'll generate this later)
    └── original

Data Setup

• 1. First, convert the structured xml file into a more readable unp file

$ pipenv run iam_cvt_xml2unp

• 2. Next, for the PoC of the TraTra model, extracting the single stroke data and generating images.

$ pipenv run iam_gen_single_stroke

• 3. Resampling is used because the method of taking points does not seem always suitable for deep learning.

$ pipenv run iam_resampling

Training

$ pipenv run iam_train --gpu_id 0

Evaluation

• Download pre-trained PoC TraTra model from gdrive link and save it wherever you want.
• Change load_model config from null to your path in config/exp001.yaml.
• Run script (change test_data_path config from null to directory path that contains test images in config/exp001.yaml)

$ pipenv run iam_test --gpu_id 0

Results

References

• [Placeholder] TraTra Idea Slides

TODOs

• IAM ONLINE multi stroke infra
• resampling multi processing
• data augmentation (resize, rotate, noise, CutUp, MixUp, etc.)