app.py

import os

import gradio as gr
import matplotlib.pyplot as plt
import numpy as np
import spaces
import torch
from model import define_model
from PIL import Image
from utils import plot_sample, postprocess_preds_page_element, reformat_for_plotting

MODEL_NAME="nvidia/nemoretriever-page-elements-v3"

model = define_model("page_element_v3")


@spaces.GPU
def inference(image):
    image = np.array(image)
    with torch.inference_mode():
        x = model.preprocess(image)
        preds = model(x, image.shape)[0]

    boxes, labels, scores = postprocess_preds_page_element(
        preds, model.thresholds_per_class, model.labels
    )
    boxes_plot, confs = reformat_for_plotting(
        boxes, labels, scores, image.shape, model.num_classes
    )
    plt.figure(figsize=(15, 10))
    plot_sample(image, boxes_plot, confs, labels=model.labels)
    plt.savefig("output.png", bbox_inches='tight', dpi=150)
    return Image.open("output.png").convert("RGB")


def gradio_reset():
    return gr.update(value=None), gr.update(value=None)


with gr.Blocks() as demo:
    with gr.Row():
        with gr.Column():
            input_img = gr.Image(label=" ", interactive=True)
            with gr.Row():
                clear = gr.Button(value="Clear")
                predict = gr.Button(value="Detect", interactive=True, variant="primary")

        with gr.Column():
            output_img = gr.Image(label=" ", interactive=False)

    clear.click(gradio_reset, inputs=None, outputs=[input_img, output_img])
    predict.click(
        inference,
        inputs=[input_img],
        outputs=[output_img],
    )

demo.launch()