app.py

import os

os.environ["TOKENIZERS_PARALLELISM"] = "false"

from PIL import ImageDraw
import streamlit as st

st.set_page_config(layout="wide")

import torch
from docquery.pipeline import get_pipeline
from docquery.document import load_bytes, load_document


def ensure_list(x):
    if isinstance(x, list):
        return x
    else:
        return [x]


CHECKPOINTS = {
    "LayoutLMv1 🦉": "impira/layoutlm-document-qa",
    "Donut 🍩": "naver-clova-ix/donut-base-finetuned-docvqa",
}


@st.experimental_singleton(show_spinner=False)
def construct_pipeline(model):
    device = "cuda" if torch.cuda.is_available() else "cpu"
    ret = get_pipeline(checkpoint=CHECKPOINTS[model], device=device)
    return ret


@st.cache(show_spinner=False)
def run_pipeline(model, question, document, top_k):
    pipeline = construct_pipeline(model)
    return pipeline(question=question, **document.context, top_k=top_k)


# TODO: Move into docquery
# TODO: Support words past the first page (or window?)
def lift_word_boxes(document):
    return document.context["image"][0][1]


def expand_bbox(word_boxes):
    if len(word_boxes) == 0:
        return None

    min_x, min_y, max_x, max_y = zip(*[x[1] for x in word_boxes])
    return [min(min_x), min(min_y), max(max_x), max(max_y)]


# LayoutLM boxes are normalized to 0, 1000
def normalize_bbox(box, width, height):
    pct = [c / 1000 for c in box]
    return [pct[0] * width, pct[1] * height, pct[2] * width, pct[3] * height]


st.markdown("# DocQuery: Query Documents w/ NLP")

if "document" not in st.session_state:
    st.session_state["document"] = None

input_col, model_col = st.columns([2, 1])

with input_col:
    input_type = st.radio("Pick an input type", ["Upload", "URL"], horizontal=True)

with model_col:
    model_type = st.radio("Pick a model", list(CHECKPOINTS.keys()), horizontal=True)


def load_file_cb():
    if st.session_state.file_input is None:
        return

    file = st.session_state.file_input
    with loading_placeholder:
        with st.spinner("Processing..."):
            document = load_bytes(file, file.name)
            _ = document.context
            st.session_state.document = document


def load_url_cb():
    if st.session_state.url_input is None:
        return

    url = st.session_state.url_input
    with loading_placeholder:
        with st.spinner("Downloading..."):
            document = load_document(url)
        with st.spinner("Processing..."):
            _ = document.context
        st.session_state.document = document


if input_type == "Upload":
    file = st.file_uploader(
        "Upload a PDF or Image document", key="file_input", on_change=load_file_cb
    )

elif input_type == "URL":
    # url = st.text_input("URL", "", on_change=load_url_callback, key="url_input")
    url = st.text_input("URL", "", key="url_input", on_change=load_url_cb)

question = st.text_input("QUESTION", "")

document = st.session_state.document
loading_placeholder = st.empty()
if document is not None:
    col1, col2 = st.columns(2)
    image = document.preview


colors = ["blue", "red", "green"]
if document is not None and question is not None and len(question) > 0:
    col2.header(f"Answers ({model_type})")
    with col2:
        answers_placeholder = st.empty()
        answers_loading_placeholder = st.empty()

        with answers_loading_placeholder:
            # Run this (one-time) expensive operation outside of the processing
            # question placeholder
            with st.spinner("Constructing pipeline..."):
                construct_pipeline(model_type)

            with st.spinner("Processing question..."):
                predictions = run_pipeline(
                    model=model_type, question=question, document=document, top_k=1
                )

        with answers_placeholder:
            image = image.copy()
            draw = ImageDraw.Draw(image)
            for i, p in enumerate(ensure_list(predictions)):
                col2.markdown(f"#### { p['answer'] }: ({round(p['score'] * 100, 1)}%)")
                if "start" in p and "end" in p:
                    x1, y1, x2, y2 = normalize_bbox(
                        expand_bbox(
                            lift_word_boxes(document)[p["start"] : p["end"] + 1]
                        ),
                        image.width,
                        image.height,
                    )
                    draw.rectangle(((x1, y1), (x2, y2)), outline=colors[i], width=3)

if document is not None:
    col1.image(image, use_column_width="auto")

"DocQuery uses LayoutLMv1 fine-tuned on DocVQA, a document visual question answering dataset, as well as SQuAD, which boosts its English-language comprehension. To use it, simply upload an image or PDF, type a question, and click 'submit', or click one of the examples to load them."

"[Github Repo](https://github.com/impira/docquery)"