This repository contains the code for the paper Analyzing the Potential of Active Learning for Document Image Classification by Saifullah Saifullah, Stefan Agne, Andreas Dengel, and Sheraz Ahmed.

Clone the repository along with sub-modules:

git clone https://github.com/saifullah3396/doc_al.git --recursive

Install the project dependencies.

pip install -r requirements.txt

Setup the environment variables for running the code.

export PYTHONPATH=`pwd`/src:`pwd`/external/xai_torch/src

Setup the output directory for generating checkpoints and logs.

export XAI_TORCH_OUTPUT_DIR=<path-to-output-dir>

Setup the output directory for dataset and models.

export XAI_TORCH_CACHE_DIR=<path-to-cache-dir>

To run an experiment, call the main training script and set args/al_args=. For example to run, active learning with entropy sampling on Tobacco3482 dataset, run:

./scripts/al_train.sh --config-path ../../../cfg/al_adv +experiment=active_learning/tobacco3482/resnet50 args/al_args=entropy_sampling args.data_args.dataset_dir=<path-to-dataset-dir>

To run different experiments, the helper scripts ./experiments/tobacco3482.sh, ./experiments/tobacco3482_pre.sh, ./experiments/rvlcdip.sh can be used. For example to run an experiment with tobacco3482 dataset, class imbalance m=2, and entropy sampling, the following script can be used.

./scripts/experiments/tobacco3482.sh --query=entropy_sampling --exp=imb_2 --data-path=<path-to-dataset-dir> --seed=0

If you find this paper helpful in your work, please consider citing:

Saifullah, S., Agne, S., Dengel, A. et al. Analyzing the potential of active learning for document image classification. IJDAR 26, 187–209 (2023). [https://doi.org/10.21203/rs.3.rs-2273654/v1](https://doi.org/10.1007/s10032-023-00429-8)]

This repository is released under the Apache 2.0 license as found in the LICENSE file.