app.py

import spaces
import gradio as gr
import cv2
import tempfile
from PIL import Image, ImageDraw, ImageFont
from transformers import RTDetrForObjectDetection, RTDetrImageProcessor
import torch
import requests

image_processor = RTDetrImageProcessor.from_pretrained("PekingU/rtdetr_r50vd")
model = RTDetrForObjectDetection.from_pretrained("PekingU/rtdetr_r50vd", torch_dtype=torch.float16).to("cuda")
model = torch.compile(model, mode="reduce-overhead")

# Compile by running inference
url = 'http://images.cocodataset.org/val2017/000000039769.jpg' 
image = Image.open(requests.get(url, stream=True).raw)
inputs = image_processor(images=image, return_tensors="pt").to("cuda", torch.float16)
with torch.no_grad():
    outputs = model(**inputs)

def draw_bounding_boxes(image, results, model, threshold=0.3):
    draw = ImageDraw.Draw(image)
    for result in results:
        for score, label_id, box in zip(
            result["scores"], result["labels"], result["boxes"]
        ):
            if score > threshold:
                label = model.config.id2label[label_id.item()]
                box = [round(i) for i in box.tolist()]
                draw.rectangle(box, outline="red", width=3)
                draw.text((box[0], box[1]), f"{label}: {score:.2f}", fill="red")
    return image

import time

@spaces.GPU
def inference(image, conf_threshold):
    inputs = image_processor(images=image, return_tensors="pt")

    start = time.time()
    with torch.no_grad():
        outputs = model(**inputs)

    results = image_processor.post_process_object_detection(
        outputs, target_sizes=torch.tensor([image.size[::-1]]), threshold=conf_threshold
    )
    end = time.time()
    print("time: ", end - start)

    bbs = draw_bounding_boxes(image, results, model, threshold=conf_threshold)
    print("bbs: ", time.time() - end)
    return bbs


css = """.my-group {max-width: 600px !important; max-height: 600 !important;}
                      .my-column {display: flex !important; justify-content: center !important; align-items: center !important};"""

with gr.Blocks(css=css) as app:
    gr.HTML(
        """
    <h1 style='text-align: center'>
    Near Real-Time Webcam Stream with RT-DETR
    </h1>
    """
    )
    gr.HTML(
        """
        <h3 style='text-align: center'>
        <a href='https://arxiv.org/abs/2304.08069' target='_blank'>arXiv</a> | <a href='https://github.com/lyuwenyu/RT-DETR' target='_blank'>github</a>
        </h3>
        """
    )
    with gr.Column(elem_classes=["my-column"]):
        with gr.Group(elem_classes=["my-group"]):
            image = gr.Image(
                type="pil",
                label="Image",
                sources="webcam",
            )
            conf_threshold = gr.Slider(
                label="Confidence Threshold",
                minimum=0.0,
                maximum=1.0,
                step=0.05,
                value=0.85,
            )
            image.stream(
                fn=inference,
                inputs=[image, conf_threshold],
                outputs=[image],
                stream_every=0.1,
                time_limit=30,
            )
if __name__ == "__main__":
    app.launch()