Tool for classification of documents based on their layout.
Steps: gather some training data and label this by putting images in their respective folders. For example make two folders: letters and forms.
${traindir}/letters
${traindir}/forms
optionally split the training data into train and validation so you have the following folders:
${traindir}/letters
${traindir}/forms
${validationdir}/letters
${validationdir}/forms
Make sure none of the images are duplicated between train and validation
python3.6 main.py --do_train \
--train_set ${traindir} \
--do_validation \
--validation_set ${validationdir} \
--seed 42 \
--gpu 0
Look at the numbers of the validation. They should be getting more correct each few epochs.
The call will look like this:
python3 main.py --do_inference --inference_set /path/to/inference/dir/ --existing_model /path/to/model/best_val
if you have data that is not balanced (different numbers of items per class) it might make sense to add
--use_class_weights