Improving accuracy and speeding up Document Image Classification through parallel systems

Paper: Improving accuracy and speeding up Document Image Classification through parallel systems

Datasets

SmallTobacco files can be downloaded here. In Data folder we provide the scripts for getting ocr .txt files (ocr_tobacco.py) and for creating .hdf5 files (ST_hdf5_dataset_creation.py) with images and ocr data.

BigTobacco files can be downloaded here. ./Data/BT_hdf5_dataset_creation.py creates train, test and validation .hdf5 files based on the aforementioned link partition.

Repository structure

Image model

├── image_model
	├── eff_big_training.py # EfficientNet training in BigTobacco
	├── eff_small_training.py # EfficientNet training in SmallTobacco
	├── eff_utils.py # EfficientNet helper with common functions for Small and Big training
	├── H5Dataset.py # Dataset class reading hdf5 file
	├── tensorflow
		├── distr_effnet_shear.py # EfficientNet

Text model

├── text_model
	├── main.py # BERT training in SmallTobacco
	├── bert_utils.py # BERT helpers
	├── training_modules
		├── data_utils.py # data cleaning and H5Dataset class
		├── finetuned_models.py # BERT model definition
		├── model_utils.py # train and test procedures

Ensemble

├── text_model
	├── ensemble.py # ensemble image and text predictions
	├── bert_utils.py # BERT helpers
	├── ensemble_modules
		├── data_utils2.py # data cleaning and H5Dataset_ensemble class
		├── model_utils_ensemble.py # BERT and EfficientNet predictions and ensemble

efficientnet_pytorch library downloads the models in .cache/torch/checkpoints.

pytorch_transformers library does it in .cache/torch/pytorch_transformers. Make sure you previously download and store in those paths the models if your machine has no internet access.