app.py

import os
from pathlib import Path
import pandas as pd
import gradio as gr
from collections import OrderedDict
from PIL import Image, ImageDraw, ImageFont
from io import BytesIO
import PyPDF2
import pdf2image

MAX_PAGES = 50
MAX_PDF_SIZE = 100000000  # almost 100MB
MIN_WIDTH, MIN_HEIGHT = 150, 150


def equal_image_grid(images):
    def compute_grid(n, max_cols=6):
        equalDivisor = int(n**0.5)
        cols = min(equalDivisor, max_cols)
        rows = equalDivisor
        if rows * cols >= n:
            return rows, cols
        cols += 1
        if rows * cols >= n:
            return rows, cols
        while rows * cols < n:
            rows += 1
        return rows, cols

    # assert len(images) == rows*cols
    rows, cols = compute_grid(len(images))

    # rescaling to min width [height padding]
    images = [im for im in images if (im.height > 0) and (im.width > 0)]  # could be NA

    min_width = min(im.width for im in images)
    images = [im.resize((min_width, int(im.height * min_width / im.width)), resample=Image.BICUBIC) for im in images]

    w, h = max([img.size[0] for img in images]), max([img.size[1] for img in images])

    grid = Image.new("RGB", size=(cols * w, rows * h))
    grid_w, grid_h = grid.size

    for i, img in enumerate(images):
        grid.paste(img, box=(i % cols * w, i // cols * h))
    return grid


def add_pagenumbers(im_list, height_scale=40):
    def add_pagenumber(image, i):
        width, height = image.size
        draw = ImageDraw.Draw(image)
        fontsize = int((width * height) ** (0.5) / height_scale)
        font = ImageFont.truetype("Arial.ttf", fontsize)
        margin = int(2 * fontsize)
        draw.text(
            (width - margin, height - margin),
            str(i + 1),
            fill="#D00917",
            font=font,
            spacing=4,
            align="right",
        )

    for i, image in enumerate(im_list):
        add_pagenumber(image, i)


def pdf_to_grid(pdf_path):
    reader = PyPDF2.PdfReader(pdf_path)
    reached_page_limit = False
    images = []
    try:
        for p, page in enumerate(reader.pages):
            if reached_page_limit:
                break
            for image in page.images:
                im = Image.open(BytesIO(image.data))
                if im.width < MIN_WIDTH and im.height < MIN_HEIGHT:
                    continue
                images.append(im)
    except Exception as e:
        print(f"{pdf_path} PyPDF get_images {e}")
        images = pdf2image.convert_from_path(pdf_path)

    # simpler but slower
    # images = pdf2image.convert_from_path(pdf_path)

    if len(images) == 0:
        return None
    add_pagenumbers(images)
    return equal_image_grid(images)


def main(complexity, evidence, form, operation, type):
    # need to write a query on diagnostic test and sample from it based on slider values
    # then return the sample
    query = " and ".join(
        [
            f"{cat}_{val} == {True}"
            for cat, val in zip(meta_cats.keys(), [complexity, evidence, form, operation, type])
            if val
        ]
    )
    results = DIAGNOSTIC_TEST.query(query)
    if len(results) == 0:
        return f"No results found for query {query}", "", "", "", ""

    for i, sample in results.sample(frac=1).iterrows():
        if not sample['nhash']:
            continue
        print("Sampled: ", sample["nhash"])

        # first get PDF file
        PDF, grid = None, None
        pdf_path = PDF_PATH / "test" / (sample["nhash"] + ".pdf")
        if not os.path.exists(pdf_path):
            continue
        PDF = pdf_path
        grid = pdf_to_grid(pdf_path)
        if not grid:
            continue
        question, answer = sample["label"] #might need to translate
        
        return label, grid, PDF

_CLASSES = [
    "letter",
    "form",
    "email",
    "handwritten",
    "advertisement",
    "scientific report",
    "scientific publication",
    "specification",
    "file folder",
    "news article",
    "budget",
    "invoice",
    "presentation",
    "questionnaire",
    "resume",
    "memo",
]
# test
# l, im, f = main(*slider_defaults)

#load both datasets in memory? --> easier retrieval afterwards with seed index based on pressing button
DATASETS = {
    'rvl_cdip': load_dataset(
        "bdpc/rvl_cdip_mp",
        split="test"), 
    'rvl_cdip_N': load_dataset(
        "bdpc/rvl_cdip_mp",
        split="test")
}

meta_cats = {'dataset': ['rvl_cdip', 'rvl_cdip_N'],
             'label': _CLASSES
            }
sliders = [gr.Dropdown(choices=choices, value=choices[-1], label=label) for label, choices in meta_cats.items()]
slider_defaults = [slider.value for slider in sliders]

outputs = [
    gr.Textbox(label="label"),
    gr.Image(label="image grid of PDF"),
    gr.File(label="PDF"),
]

iface = gr.Interface(fn=main, inputs=sliders, outputs=outputs, description="Visualize PDF samples from multi-page (PDF) document classification datasets", title='Beyond Document Page Classification: Examples')
iface.launch(share=True)