app.py

import io
import pandas as pd
import plotly.express as px
import streamlit as st

import torch
import torch.nn.functional as F
from easyocr import Reader
from PIL import Image
from torch.utils.data import Dataset, DataLoader

from transformers import (
    LayoutLMv3FeatureExtractor, 
    LayoutLMv3TokenizerFast, 
    LayoutLMv3Processor, 
    LayoutLMv3ForSequenceClassification
    )

DEVICE = "cuda:0" if torch.cuda.is_available() else "cpu"
MICROSOFT_MODEL_NAME = "microsoft/layoutlmv3-base"
MODEL_NAME = "curiousily/layoutlmv3-financial-document-classification"

def creat_bounding_box(bbox_data, width_scale: float, height_scale: float):
    xs = []
    ys = []
    for x, y in bbox_data:
        xs.append(x)
        ys.append(y)

        left = int(min(xs) * width_scale)
        top = int(min(ys) * height_scale)
        right = int(max(xs) * width_scale)
        bottom = int(max(ys) * height_scale)
    
    return [left, top, right, bottom]

@st.experimental_singleton
def create_ocr_reader():
    return Reader(["en"])

@st.experimental_singleton
def create_processor():
    feature_extractor = LayoutLMv3FeatureExtractor(apply_ocr=False)
    tokenizer = LayoutLMv3TokenizerFast.from_pretrained(MICROSOFT_MODEL_NAME)
    return LayoutLMv3Processor(feature_extractor, tokenizer)

@st.experimental_singleton
def create_model():
    model = LayoutLMv3ForSequenceClassification.from_pretrained(MODEL_NAME)
    return model.eval().to(DEVICE)

def predict(
    image: Image, 
    reader: Reader, 
    processor: LayoutLMv3Processor, 
    model: LayoutLMv3ForSequenceClassification
):

    ocr_result = reader.readtext(image)

    width, height = image.size
    width_scale = 1000 / width
    height_scale = 1000 / height

    words = []
    boxes = []

    for bbox, word, confidence in ocr_result:
        words.append(word)
        boxes.append(creat_bounding_box(bbox, width_scale, height_scale))

    encoding = processor(
        image, 
        words, 
        boxes=boxes, 
        max_length=512, 
        padding="max_length", 
        truncation=True, 
        return_tensors="pt"
        )

    with torch.inference_mode():
        output = model(
            input_ids=encoding["input_ids"].to(DEVICE),
            attention_mask=encoding["attention_mask"].to(DEVICE),
            bbox=encoding["bbox"].to(DEVICE),
            pixel_values=encoding["pixel_values"].to(DEVICE),
        )

    logits = output.logits
    predicted_class = logits.argmax()
    probabilities = F.softmax(logits, dim=-1).flatten().tolist()

    return predicted_class.detach().item(), probabilities 


reader = create_ocr_reader()
processor = create_processor()
model = create_model()

uploaded_file = st.file_uploader("Upload Document Image", ["jpg", "png"])
if uploaded_file is not None:
    bytes_data = io.BytesIO(uploaded_file.getvalue())
    image = Image.open(bytes_data)
    st.image(image, "Your Document")

    predicted_class, probabilities = predict(image, reader, processor, model)
    predicted_label = model.config.id2label[predicted_class]
    
    st.markdown(f"Predicted document type: **{predicted_label}**")

    df_predictions = pd.DataFrame(
        {"Document": list(model.config.id2label.values()), "Confidence": probabilities}
    )

    fig = px.bar(df_predictions, x="Document", y="Confidence")
    st.plotly_chart(fig, use_container_width=True)