🧪distillate is a Python package to aid with information extraction from PDFs.
It allows you to annotate PDF text using bounding boxes, and train multi-modal models.
For 🧪distillate to work, you must first install Poppler, a unix PDF annotation tool.
In future versions, I'd like to get OS X and Windows installation working.
sudo apt update
sudo apt install poppler
Once the above requirements are met, you can install via:
pip install -e .
There are 6 commands that make up the distillate core.
The first step is to create a new dataset from a collection of PDFs (and optionally annotations):
distillate create [DATASET] --pdfs=[PDFs] --base=[BASE]
If you add more PDFs or annotations, you can use distillate update to add them to the dataset:
distillate update [DATASET] --pdfs=[PDFs] --annotations=[ANNOTATIONS]
And if you would like to visualize annotations for the dataset (or for a specific file), you can use distillate visualize:
distillate visualize [DATASET] --filename=[FILE] --base=[BASE][DATASET]
To train a LayoutLM model, use distillate train with a specific dataset:
distillate train [DATASET] --model=[MODEL] --split=[SPLIT_PERCENTAGE]
If you'd like to use a pretrained model instead, you can use distillate download to fetch one of the available models:
distillate download [MODEL]
Using a pretrained model, you can run over a PDF using distillate predict, passing an optional serialization directory to store the intermediate outputs:
distillate predict [FILE] --model=[MODEL] --inspect/--no-inspect --serialization-dir=[SERIALIZATION_DIRECTORY]
We use the CSV annotations from MakeSense.ai.