app.py

import gradio as gr
import torch
import numpy as np
from PIL import Image
from huggingface_hub import hf_hub_download
import torchvision.transforms as transforms

# Загрузка модели при старте
print("Loading model...")
model_path = hf_hub_download(
    repo_id="cwlachap/hrnet-cephalometric-landmark-detection",
    filename="best_model.pth"
)

# Здесь нужна архитектура HRNet - упрощенная версия для демо
class SimpleHRNet(torch.nn.Module):
    def __init__(self):
        super().__init__()
        # Simplified model architecture
        self.features = torch.nn.Sequential(
            torch.nn.Conv2d(3, 64, 3, padding=1),
            torch.nn.ReLU(),
            torch.nn.AdaptiveAvgPool2d((1, 1))
        )
        self.landmarks = torch.nn.Linear(64, 19 * 2)  # 19 landmarks, x and y
    
    def forward(self, x):
        x = self.features(x)
        x = x.view(x.size(0), -1)
        x = self.landmarks(x)
        return x.view(-1, 19, 2)

try:
    model = SimpleHRNet()
    checkpoint = torch.load(model_path, map_location='cpu')
    if 'model_state_dict' in checkpoint:
        model.load_state_dict(checkpoint['model_state_dict'], strict=False)
    model.eval()
    print("Model loaded successfully!")
except Exception as e:
    print(f"Error loading model: {e}")
    model = None

# Названия landmarks
LANDMARK_NAMES = [
    "Sella", "Nasion", "Orbitale", "Porion", "Point A", "Point B",
    "Pogonion", "Menton", "Gnathion", "Gonion", "Lower Incisor Tip",
    "Upper Incisor Tip", "Upper Lip", "Lower Lip", "Subnasale",
    "Soft Tissue Pogonion", "Posterior Nasal Spine", "Anterior Nasal Spine",
    "Articulare"
]

def analyze_image(image):
    if model is None:
        return {"error": "Model not loaded"}, None
    
    try:
        # Преобразование изображения
        transform = transforms.Compose([
            transforms.Resize((768, 768)),
            transforms.ToTensor(),
            transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
        ])
        
        img = Image.fromarray(image).convert('RGB')
        img_tensor = transform(img).unsqueeze(0)
        
        # Inference
        with torch.no_grad():
            landmarks = model(img_tensor)
        
        # Преобразование координат обратно к размеру изображения
        h, w = image.shape[:2]
        landmarks = landmarks.squeeze().cpu().numpy()
        landmarks[:, 0] = landmarks[:, 0] * w / 768
        landmarks[:, 1] = landmarks[:, 1] * h / 768
        
        # Формирование результата
        result = []
        for i, (x, y) in enumerate(landmarks):
            result.append({
                "landmark": LANDMARK_NAMES[i],
                "x": float(x),
                "y": float(y)
            })
        
        return result, image
    
    except Exception as e:
        return {"error": str(e)}, None

def process_for_api(image):
    result, _ = analyze_image(image)
    return result

# Gradio интерфейс
with gr.Blocks() as demo:
    gr.Markdown("# Cephalometric Landmark Detection")
    gr.Markdown("Upload a lateral cephalometric radiograph to detect 19 anatomical landmarks")
    
    with gr.Row():
        input_image = gr.Image(label="Upload X-ray")
        output_json = gr.JSON(label="Detected Landmarks")
    
    analyze_btn = gr.Button("Analyze")
    analyze_btn.click(fn=analyze_image, inputs=input_image, outputs=[output_json, gr.Image(visible=False)])

# API endpoint
api_demo = gr.Interface(
    fn=process_for_api,
    inputs=gr.Image(),
    outputs=gr.JSON(),
    api_name="predict"
)

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