app.py

import os
import cv2
import torch
import onnx
import numpy as np
import gradio as gr
from huggingface_hub import hf_hub_download
from gradio_webrtc import WebRTC
# --- YOLOv5n ONNX 추론 클래스 (간단히 구현) ---
import onnxruntime

class YOLOv5nONNX:
    def __init__(self, onnx_path):
        self.session = onnxruntime.InferenceSession(onnx_path, providers=['CPUExecutionProvider'])
        input_shape = self.session.get_inputs()[0].shape  # e.g. [1,3,640,640]
        self.input_height, self.input_width = input_shape[2], input_shape[3]
    
    def preprocess(self, image):
        # BGR to RGB, resize, normalize, transpose
        img = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
        img = cv2.resize(img, (self.input_width, self.input_height))
        img = img.astype(np.float32) / 255.0
        img = np.transpose(img, (2,0,1))  # HWC to CHW
        img = np.expand_dims(img, axis=0)
        return img
    
    def postprocess(self, outputs, conf_threshold=0.3):
        # 여기선 간단히 confidence 필터링 정도만, 보통은 NMS 필요
        preds = outputs[0]
        preds = preds[preds[:, 4] > conf_threshold]  # conf 필터링
        return preds
    
    def detect_objects(self, image, conf_threshold=0.3):
        input_tensor = self.preprocess(image)
        outputs = self.session.run(None, {'images': input_tensor})
        preds = self.postprocess(outputs, conf_threshold)

        # 간단한 박스 그리기 예시 (여기서는 bbox 좌표 가정 필요, 실제 yolov5 onnx 출력 형태에 맞게 수정 필요)
        for *box, conf, cls in preds:
            x1, y1, x2, y2 = map(int, box)
            cv2.rectangle(image, (x1,y1), (x2,y2), (0,255,0), 2)
            cv2.putText(image, f'{int(cls)}:{conf:.2f}', (x1,y1-10), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0,255,0), 2)
        return image

# --- PT → ONNX 변환 함수 ---
def convert_pt_to_onnx(pt_path, onnx_path, input_size=(640, 640)):
    model = torch.hub.load('ultralytics/yolov5', 'custom', path=pt_path)
    model.eval()
    dummy_input = torch.randn(1, 3, *input_size)
    torch.onnx.export(
        model.model,
        dummy_input,
        onnx_path,
        input_names=['images'],
        output_names=['output'],
        opset_version=12,
        dynamic_axes={'images': {0: 'batch'}, 'output': {0: 'batch'}},
    )
    onnx_model = onnx.load(onnx_path)
    onnx.checker.check_model(onnx_model)
    print(f"Converted {pt_path} to {onnx_path}")

# --- 모델 로딩 함수 ---
model = None
def load_model(weight_path):
    global model
    ext = os.path.splitext(weight_path)[1].lower()
    if ext == '.pt':
        onnx_path = weight_path.replace('.pt', '.onnx')
        if not os.path.exists(onnx_path):
            convert_pt_to_onnx(weight_path, onnx_path)
        model = YOLOv5nONNX(onnx_path)
    elif ext == '.onnx':
        model = YOLOv5nONNX(weight_path)
    else:
        raise ValueError("지원하지 않는 모델 파일 형식입니다. .pt 또는 .onnx만 가능")
    print(f"Model loaded from {weight_path}")

# --- Detection 함수 ---
def detection(image, weight_file, conf_threshold=0.3):
    global model
    if weight_file is not None:
        if model is None or weight_file.name != getattr(model, 'weight_path', None):
            # 모델 새로 로드
            load_model(weight_file.name)
            model.weight_path = weight_file.name

    if model is None:
        return image

    result_img = model.detect_objects(image, conf_threshold)
    return result_img

# --- Gradio UI ---
with gr.Blocks() as demo:
    gr.Markdown("# YOLOv5n ONNX Detection with Weight Upload (WebRTC)")

    weight_file = gr.File(label="Upload YOLOv5n weights (.pt or .onnx)", file_types=['.pt', '.onnx'])
    conf_threshold = gr.Slider(0.0, 1.0, value=0.3, step=0.05, label="Confidence Threshold")

    rtc_stream = WebRTC(label="WebRTC Webcam")

    output_image = gr.Image(label="Detection Output")

    rtc_stream.stream(
        fn=detection,
        inputs=[rtc_stream, weight_file, conf_threshold],
        outputs=output_image,
    )

if __name__ == "__main__":
    demo.launch()