app.py

import os
os.system('pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cpu')

import gradio as gr
import numpy as np
from transformers import AutoModelForTokenClassification
from datasets.features import ClassLabel
from transformers import AutoProcessor
from datasets import Features, Sequence, ClassLabel, Value, Array2D, Array3D
import torch
from datasets import load_metric
from transformers import LayoutLMTokenizer
from transformers.data.data_collator import default_data_collator


from transformers import LayoutLMForTokenClassification
from datasets import load_dataset
from PIL import Image, ImageDraw, ImageFont


tokenizer = LayoutLMTokenizer.from_pretrained("microsoft/layoutlm-base-uncased")
model = LayoutLMForTokenClassification.from_pretrained("microsoft/layoutlm-base-uncased", num_labels=13)




id2label = {0: 'O', 1: 'B-HEADER', 2: 'I-HEADER', 3: 'B-QUESTION', 4: 'I-QUESTION', 5: 'B-ANSWER', 6: 'I-ANSWER'}
label2color = {'question':'blue', 'answer':'green', 'header':'orange', 'other':'violet'}


def unnormalize_box(bbox, width, height):
     return [
         width * (bbox[0] / 1000),
         height * (bbox[1] / 1000),
         width * (bbox[2] / 1000),
         height * (bbox[3] / 1000),
     ]


def iob_to_label(label):
    return label



def process_image(image):

    print(type(image))
    width, height = image.size

    # encode
    encoding = tokenizer(image, truncation=True, return_offsets_mapping=True, return_tensors="pt")
    offset_mapping = encoding.pop('offset_mapping')

    # forward pass
    outputs = model(**encoding)

    # get predictions
    predictions = outputs.logits.argmax(-1).squeeze().tolist()
    token_boxes = encoding.bbox.squeeze().tolist()

    # only keep non-subword predictions
    is_subword = np.array(offset_mapping.squeeze().tolist())[:,0] != 0
    true_predictions = [id2label[pred] for idx, pred in enumerate(predictions) if not is_subword[idx]]
    true_boxes = [unnormalize_box(box, width, height) for idx, box in enumerate(token_boxes) if not is_subword[idx]]

    # draw predictions over the image
    draw = ImageDraw.Draw(image)
    font = ImageFont.load_default()
    for prediction, box in zip(true_predictions, true_boxes):
        predicted_label = iob_to_label(prediction)
        draw.rectangle(box, outline=label2color[predicted_label])
        draw.text((box[0]+10, box[1]-10), text=predicted_label, fill=label2color[predicted_label], font=font)
    
    return image


title = "Extraction d'informations de factures en utilisant le modèle LayoutLM"
description = "J'utilise LayoutLM de Microsoft formé sur un ensemble de données de factures pour prédire le nom de l'émetteur de factures, l'adresse de l'émetteur de factures, le code postal de l'émetteur de factures, la date d'échéance, la TPS, la date de facturation, le numéro de facture, le sous-total et le total. Pour l'utiliser, il suffit de télécharger une image ou d'utiliser l'exemple d'image ci-dessous. Les résultats seront affichés en quelques secondes."

article="<b>References</b><br>[1] Y. Xu et al., “LayoutLMv3: Pre-training for Document AI with Unified Text and Image Masking.” 2022. <a href='https://arxiv.org/abs/2204.08387'>Paper Link</a><br>[2]  <a href='https://github.com/NielsRogge/Transformers-Tutorials/tree/master/LayoutLMv3'>LayoutLMv3 training and inference</a>" 


css = """.output_image, .input_image {height: 600px !important}"""

iface = gr.Interface(fn=process_image, 
                     inputs=gr.inputs.Image(type="pil"), 
                     outputs=gr.outputs.Image(type="pil", label="annotated image"),
                     title=title,
                     description=description,
                     article=article,
                     css=css,
                     analytics_enabled = True, enable_queue=True)

iface.launch(inline=False, share=False, debug=False)