app.py

#!/usr/bin/env python

from __future__ import annotations

import pathlib

import gradio as gr
import numpy as np
import PIL.Image
import torch
from sahi.prediction import ObjectPrediction
from sahi.utils.cv import visualize_object_predictions
from transformers import AutoImageProcessor, DetaForObjectDetection
from ultralytics import YOLO

DESCRIPTION = '# Compare DETA and YOLOv8'

device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')

MODEL_ID = 'jozhang97/deta-swin-large'
image_processor = AutoImageProcessor.from_pretrained(MODEL_ID)
model_deta = DetaForObjectDetection.from_pretrained(MODEL_ID)
model_deta.to(device)

model_yolo = YOLO('yolov8x.pt')


@torch.inference_mode()
def run_deta(image_path: str, threshold: float) -> np.ndarray:
    image = PIL.Image.open(image_path)
    inputs = image_processor(images=image, return_tensors='pt').to(device)
    outputs = model_deta(**inputs)
    target_sizes = torch.tensor([image.size[::-1]])
    results = image_processor.post_process_object_detection(
        outputs, threshold=threshold, target_sizes=target_sizes)[0]

    boxes = results['boxes'].cpu().numpy()
    scores = results['scores'].cpu().numpy()
    cat_ids = results['labels'].cpu().numpy().tolist()
    preds = []
    for box, score, cat_id in zip(boxes, scores, cat_ids):
        box = np.round(box).astype(int)
        cat_label = model_deta.config.id2label[cat_id]
        pred = ObjectPrediction(bbox=box,
                                category_id=cat_id,
                                category_name=cat_label,
                                score=score)
        preds.append(pred)
    res = visualize_object_predictions(np.asarray(image), preds)['image']
    return res


def run_yolov8(image_path: str, threshold: float) -> np.ndarray:
    image = PIL.Image.open(image_path)
    results = model_yolo(image, imgsz=640, conf=threshold)
    boxes = results[0].boxes.cpu().numpy().data
    preds = []
    for box in boxes:
        score = box[4]
        cat_id = int(box[5])
        box = np.round(box[:4]).astype(int)
        cat_label = model_yolo.model.names[cat_id]
        pred = ObjectPrediction(bbox=box,
                                category_id=cat_id,
                                category_name=cat_label,
                                score=score)
        preds.append(pred)
    res = visualize_object_predictions(np.asarray(image), preds)['image']
    return res


def run(image_path: str, threshold: float) -> tuple[np.ndarray, np.ndarray]:
    return run_deta(image_path, threshold), run_yolov8(image_path, threshold)


with gr.Blocks(css='style.css') as demo:
    gr.Markdown(DESCRIPTION)
    with gr.Row():
        with gr.Column():
            image = gr.Image(label='Input image', type='filepath')
            threshold = gr.Slider(label='Score threshold',
                                  minimum=0,
                                  maximum=1,
                                  value=0.5,
                                  step=0.01)
            run_button = gr.Button('Run')
        with gr.Column():
            result_deta = gr.Image(label='Result (DETA)', type='numpy')
            result_yolo = gr.Image(label='Result (YOLOv8)', type='numpy')

    with gr.Row():
        paths = sorted(pathlib.Path('images').glob('*.jpg'))
        gr.Examples(examples=[[path.as_posix(), 0.5] for path in paths],
                    inputs=[
                        image,
                        threshold,
                    ],
                    outputs=[
                        result_deta,
                        result_yolo,
                    ],
                    fn=run,
                    cache_examples=True)

    run_button.click(fn=run,
                     inputs=[
                         image,
                         threshold,
                     ],
                     outputs=[
                         result_deta,
                         result_yolo,
                     ])

demo.queue().launch()