app.py

import os
os.system('pip install detectron2 -f https://dl.fbaipublicfiles.com/detectron2/wheels/cu102/torch1.9/index.html')

# work around: https://discuss.huggingface.co/t/how-to-install-a-specific-version-of-gradio-in-spaces/13552
os.system("pip uninstall -y gradio")
os.system("pip install gradio==3.4.1")

from os import getcwd, path, environ
import deepdoctection as dd
from deepdoctection.dataflow.serialize import DataFromList

import gradio as gr


_DD_ONE = "conf_dd_one.yaml"
_DETECTIONS = ["table", "ocr"]

dd.ModelCatalog.register("layout/model_final_inf_only.pt",dd.ModelProfile(
            name="layout/model_final_inf_only.pt",
            description="Detectron2 layout detection model trained on private datasets",
            config="dd/d2/layout/CASCADE_RCNN_R_50_FPN_GN.yaml",
            size=[274632215],
            tp_model=False,
            hf_repo_id=environ.get("HF_REPO"),
            hf_model_name="model_final_inf_only.pt",
            hf_config_file=["Base-RCNN-FPN.yaml", "CASCADE_RCNN_R_50_FPN_GN.yaml"],
            categories={"1": dd.LayoutType.text,
                        "2": dd.LayoutType.title,
                        "3": dd.LayoutType.list,
                        "4": dd.LayoutType.table,
                        "5": dd.LayoutType.figure},
        ))

# Set up of the configuration and logging. Models are globally defined, so that they are not re-loaded once the input
# updates
cfg = dd.set_config_by_yaml(path.join(getcwd(),_DD_ONE))
cfg.freeze(freezed=False)
cfg.DEVICE = "cpu"
cfg.freeze()

# layout detector
layout_config_path = dd.ModelCatalog.get_full_path_configs(cfg.CONFIG.D2LAYOUT)
layout_weights_path = dd.ModelDownloadManager.maybe_download_weights_and_configs(cfg.WEIGHTS.D2LAYOUT)
categories_layout = dd.ModelCatalog.get_profile(cfg.WEIGHTS.D2LAYOUT).categories
assert categories_layout is not None
assert layout_weights_path is not None
d_layout = dd.D2FrcnnDetector(layout_config_path, layout_weights_path, categories_layout, device=cfg.DEVICE)

# cell detector
cell_config_path = dd.ModelCatalog.get_full_path_configs(cfg.CONFIG.D2CELL)
cell_weights_path = dd.ModelDownloadManager.maybe_download_weights_and_configs(cfg.WEIGHTS.D2CELL)
categories_cell = dd.ModelCatalog.get_profile(cfg.WEIGHTS.D2CELL).categories
assert categories_cell is not None
d_cell = dd.D2FrcnnDetector(cell_config_path, cell_weights_path, categories_cell, device=cfg.DEVICE)

# row/column detector
item_config_path = dd.ModelCatalog.get_full_path_configs(cfg.CONFIG.D2ITEM)
item_weights_path = dd.ModelDownloadManager.maybe_download_weights_and_configs(cfg.WEIGHTS.D2ITEM)
categories_item = dd.ModelCatalog.get_profile(cfg.WEIGHTS.D2ITEM).categories
assert categories_item is not None
d_item = dd.D2FrcnnDetector(item_config_path, item_weights_path, categories_item, device=cfg.DEVICE)

# word detector
det = dd.DoctrTextlineDetector()

# text recognizer
rec = dd.DoctrTextRecognizer()


def build_gradio_analyzer(table, table_ref, ocr):
    """Building the Detectron2/DocTr analyzer based on the given config"""

    cfg.freeze(freezed=False)
    cfg.TAB = table
    cfg.TAB_REF = table_ref
    cfg.OCR = ocr
    cfg.freeze()

    pipe_component_list = []
    layout = dd.ImageLayoutService(d_layout, to_image=True, crop_image=True)
    pipe_component_list.append(layout)

    if cfg.TAB:
        cell = dd.SubImageLayoutService(d_cell, dd.LayoutType.table, {1: 6}, True)
        pipe_component_list.append(cell)

        item = dd.SubImageLayoutService(d_item, dd.LayoutType.table, {1: 7, 2: 8}, True)
        pipe_component_list.append(item)

        table_segmentation = dd.TableSegmentationService(
            cfg.SEGMENTATION.ASSIGNMENT_RULE,
            cfg.SEGMENTATION.IOU_THRESHOLD_ROWS
            if cfg.SEGMENTATION.ASSIGNMENT_RULE in ["iou"]
            else cfg.SEGMENTATION.IOA_THRESHOLD_ROWS,
            cfg.SEGMENTATION.IOU_THRESHOLD_COLS
            if cfg.SEGMENTATION.ASSIGNMENT_RULE in ["iou"]
            else cfg.SEGMENTATION.IOA_THRESHOLD_COLS,
            cfg.SEGMENTATION.FULL_TABLE_TILING,
            cfg.SEGMENTATION.REMOVE_IOU_THRESHOLD_ROWS,
            cfg.SEGMENTATION.REMOVE_IOU_THRESHOLD_COLS,
        )
        pipe_component_list.append(table_segmentation)

        if cfg.TAB_REF:
            table_segmentation_refinement = dd.TableSegmentationRefinementService()
            pipe_component_list.append(table_segmentation_refinement)

    if cfg.OCR:
        d_layout_text = dd.ImageLayoutService(det, to_image=True, crop_image=True)
        pipe_component_list.append(d_layout_text)

        d_text = dd.TextExtractionService(rec, extract_from_roi="WORD")
        pipe_component_list.append(d_text)

        match = dd.MatchingService(
            parent_categories=cfg.WORD_MATCHING.PARENTAL_CATEGORIES,
            child_categories=dd.LayoutType.word,
            matching_rule=cfg.WORD_MATCHING.RULE,
            threshold=cfg.WORD_MATCHING.IOU_THRESHOLD
            if cfg.WORD_MATCHING.RULE in ["iou"]
            else cfg.WORD_MATCHING.IOA_THRESHOLD,
        )
        pipe_component_list.append(match)
        order = dd.TextOrderService(
            text_container=dd.LayoutType.word,
            floating_text_block_names=[dd.LayoutType.title, dd.LayoutType.text, dd.LayoutType.list],
            text_block_names=[
                dd.LayoutType.title,
                dd.LayoutType.text,
                dd.LayoutType.list,
                dd.LayoutType.cell,
                dd.CellType.header,
                dd.CellType.body,
            ],
        )
        pipe_component_list.append(order)

    pipe = dd.DoctectionPipe(pipeline_component_list=pipe_component_list)

    return pipe


def prepare_output(dp, add_table, add_ocr):
    out = dp.as_dict()
    out.pop("image")

    layout_items = dp.items
    if add_ocr:
        layout_items.sort(key=lambda x: x.reading_order)
    layout_items_str = ""
    for item in layout_items:
        layout_items_str += f"\n {item.layout_type}: {item.text}"
    if add_table:
        html_list = [table.html for table in dp.tables]
        if html_list:
            html = html_list[0]
        else:
            html = None
    else:
        html = None

    return dp.viz(show_table_structure=False), layout_items_str, html, out


def analyze_image(img, pdf, attributes):

    # creating an image object and passing to the analyzer by using dataflows
    add_table = _DETECTIONS[0] in attributes
    add_ocr = _DETECTIONS[1] in attributes

    analyzer = build_gradio_analyzer(add_table, add_table, add_ocr)

    if img is not None:
        image = dd.Image(file_name="input.png", location="")
        image.image = img[:, :, ::-1]

        df = DataFromList(lst=[image])
        df = analyzer.analyze(dataset_dataflow=df)
    elif pdf:
        df = analyzer.analyze(path=pdf.name, max_datapoints=3)
    else:
        raise ValueError

    df.reset_state()
    df_iter = iter(df)

    dp = next(df_iter)

    return prepare_output(dp, add_table, add_ocr)


demo = gr.Blocks(css="scrollbar.css")

with demo:
    with gr.Box():
        gr.Markdown("<h1><center>deepdoctection - A Document AI Package</center></h1>")
        gr.Markdown("<strong>deep</strong>doctection is a Python library that orchestrates document extraction"
                    " and document layout analysis tasks using deep learning models. It does not implement models"
                    " but enables you to build pipelines using highly acknowledged libraries for object detection,"
                    " OCR and selected NLP tasks and provides an integrated frameworks for fine-tuning, evaluating"
                    " and running models.\n This pipeline consists of a stack of models powered by <strong>Detectron2"
                    "</strong> for layout analysis and table recognition and <strong>DocTr</strong> for OCR.")
    with gr.Box():
        gr.Markdown("<h2><center>Upload a document and choose setting</center></h2>")
        with gr.Row():
            with gr.Column():
                with gr.Tab("Image upload"):
                    with gr.Column():
                        inputs = gr.Image(type='numpy', label="Original Image")
                with gr.Tab("PDF upload (only first image will be processed)"):
                    with gr.Column():
                        inputs_pdf = gr.File(label="PDF")
            with gr.Column():
                gr.Examples(
                    examples=[path.join(getcwd(), "sample_1.jpg"), path.join(getcwd(), "sample_2.png")],
                    inputs = inputs)

        with gr.Row():
            tok_input = gr.CheckboxGroup(
                _DETECTIONS, value=_DETECTIONS, label="Additional extractions", interactive=True)
        with gr.Row():
            btn = gr.Button("Run model", variant="primary")

    with gr.Box():
        with gr.Row():
            with gr.Column():
                gr.Markdown("<h2><center>Text output</center></h2>")
                gr.Markdown("Will only show contiguous text from text blocks, titles and lists")
                image_text = gr.Textbox()
                gr.Markdown("<h2><center>First table</center></h2>")
                html = gr.HTML()
                gr.Markdown("<h2><center>JSON output</center></h2>") 
                json = gr.JSON()
            with gr.Column():
                gr.Markdown("<h2><center>Layout detection</center></h2>")
                image_output = gr.Image(type="numpy", label="Output Image")

    btn.click(fn=analyze_image, inputs=[inputs, inputs_pdf, tok_input], outputs=[image_output, image_text, html, json])

demo.launch()