Upload 3 files

CovNextV2Base_best_model.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b3d10b865b83d0fcda631e31e1aac7b2b51f43dc139674706611bd5c1b68afd8
+size 413251664

app.py

@@ -0,0 +1,523 @@
+import os
+import numpy as np
+import torch
+import torch.nn as nn
+import gradio as gr
+import time
+import timm
+from torchvision.ops import nms, box_iou
+import torch.nn.functional as F
+from torchvision import transforms
+from PIL import Image, ImageDraw, ImageFont, ImageFilter
+from breed_health_info import breed_health_info
+from breed_noise_info import breed_noise_info
+from dog_database import get_dog_description
+from scoring_calculation_system import UserPreferences
+from recommendation_html_format import format_recommendation_html, get_breed_recommendations
+from history_manager import UserHistoryManager
+from search_history import create_history_tab, create_history_component
+from styles import get_css_styles
+from breed_detection import create_detection_tab
+from breed_comparison import create_comparison_tab
+from breed_recommendation import create_recommendation_tab
+from html_templates import (
+    format_description_html,
+    format_single_dog_result,
+    format_multiple_breeds_result,
+    format_error_message,
+    format_warning_html,
+    format_multi_dog_container,
+    format_breed_details_html,
+    get_color_scheme,
+    get_akc_breeds_link
+)
+from model_architecture import BaseModel, dog_breeds
+from urllib.parse import quote
+from ultralytics import YOLO
+import asyncio
+import traceback
+
+history_manager = UserHistoryManager()
+
+class ModelManager:
+    """
+    Singleton class for managing model instances and device allocation
+    specifically designed for Hugging Face Spaces deployment.
+    """
+    _instance = None
+    _initialized = False
+    _yolo_model = None
+    _breed_model = None
+    _device = None
+
+    def __new__(cls):
+        if cls._instance is None:
+            cls._instance = super().__new__(cls)
+        return cls._instance
+
+    def __init__(self):
+        if not ModelManager._initialized:
+            self._device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+            ModelManager._initialized = True
+    
+    @property
+    def device(self):
+        if self._device is None:
+            self._device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+        return self._device
+
+    @property
+    def yolo_model(self):
+        if self._yolo_model is None:
+            self._yolo_model = YOLO('yolov8x.pt')
+        return self._yolo_model
+
+    @property
+    def breed_model(self):
+        if self._breed_model is None:
+            self._breed_model = BaseModel(
+                num_classes=len(dog_breeds), 
+                device=self.device
+            ).to(self.device)
+            
+            checkpoint = torch.load(
+                'ConvNextV2Base_best_model.pth',
+                map_location=self.device
+            )
+            self._breed_model.load_state_dict(checkpoint['base_model'], strict=False)
+            self._breed_model.eval()
+        return self._breed_model
+
+# Initialize model manager
+model_manager = ModelManager()
+
+def preprocess_image(image):
+    """Preprocesses images for model input"""
+    if isinstance(image, np.ndarray):
+        image = Image.fromarray(image)
+
+    transform = transforms.Compose([
+        transforms.Resize((224, 224)),
+        transforms.ToTensor(),
+        transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
+    ])
+
+    return transform(image).unsqueeze(0)
+
+@spaces.GPU
+def predict_single_dog(image):
+    """Predicts dog breed for a single image"""
+    image_tensor = preprocess_image(image).to(model_manager.device)
+    
+    with torch.no_grad():
+        logits = model_manager.breed_model(image_tensor)[0]
+        probs = F.softmax(logits, dim=1)
+        
+        top5_prob, top5_idx = torch.topk(probs, k=5)
+        breeds = [dog_breeds[idx.item()] for idx in top5_idx[0]]
+        probabilities = [prob.item() for prob in top5_prob[0]]
+        
+        sum_probs = sum(probabilities[:3])
+        relative_probs = [f"{(prob/sum_probs * 100):.2f}%" for prob in probabilities[:3]]
+        
+        return probabilities[0], breeds[:3], relative_probs
+
+def enhanced_preprocess(image, is_standing=False, has_overlap=False):
+    """
+    Enhanced image preprocessing function with special handling for different poses
+    and overlapping cases.
+    """
+    target_size = 224
+    w, h = image.size
+    
+    if is_standing:
+        if h > w * 1.5:
+            new_h = target_size
+            new_w = min(target_size, int(w * (target_size / h)))
+            new_w = max(new_w, int(target_size * 0.6))
+    elif has_overlap:
+        scale = min(target_size/w, target_size/h) * 0.95
+        new_w = int(w * scale)
+        new_h = int(h * scale)
+    else:
+        scale = min(target_size/w, target_size/h)
+        new_w = int(w * scale)
+        new_h = int(h * scale)
+    
+    resized = image.resize((new_w, new_h), Image.Resampling.LANCZOS)
+    final_image = Image.new('RGB', (target_size, target_size), (240, 240, 240))
+    paste_x = (target_size - new_w) // 2
+    paste_y = (target_size - new_h) // 2
+    final_image.paste(resized, (paste_x, paste_y))
+    
+    return final_image
+
+@spaces.GPU
+def detect_multiple_dogs(image, conf_threshold=0.3, iou_threshold=0.3):
+    """
+    Enhanced multiple dog detection with improved bounding box handling and
+    intelligent boundary adjustments.
+    """
+    results = model_manager.yolo_model(image, conf=conf_threshold, iou=iou_threshold)[0]
+    img_width, img_height = image.size
+    detected_boxes = []
+    
+    # Phase 1: Initial detection and processing
+    for box in results.boxes:
+        if box.cls.item() == 16:  # Dog class
+            xyxy = box.xyxy[0].tolist()
+            confidence = box.conf.item()
+            x1, y1, x2, y2 = map(int, xyxy)
+            w = x2 - x1
+            h = y2 - y1
+            
+            detected_boxes.append({
+                'coords': [x1, y1, x2, y2],
+                'width': w,
+                'height': h,
+                'center_x': (x1 + x2) / 2,
+                'center_y': (y1 + y2) / 2,
+                'area': w * h,
+                'confidence': confidence,
+                'aspect_ratio': w / h if h != 0 else 1
+            })
+    
+    if not detected_boxes:
+        return [(image, 1.0, [0, 0, img_width, img_height], False)]
+    
+    # Phase 2: Analysis of detection relationships
+    avg_height = sum(box['height'] for box in detected_boxes) / len(detected_boxes)
+    avg_width = sum(box['width'] for box in detected_boxes) / len(detected_boxes)
+    avg_area = sum(box['area'] for box in detected_boxes) / len(detected_boxes)
+    
+    def calculate_iou(box1, box2):
+        x1 = max(box1['coords'][0], box2['coords'][0])
+        y1 = max(box1['coords'][1], box2['coords'][1])
+        x2 = min(box1['coords'][2], box2['coords'][2])
+        y2 = min(box1['coords'][3], box2['coords'][3])
+        
+        if x2 <= x1 or y2 <= y1:
+            return 0.0
+            
+        intersection = (x2 - x1) * (y2 - y1)
+        area1 = box1['area']
+        area2 = box2['area']
+        return intersection / (area1 + area2 - intersection)
+    
+    # Phase 3: Processing each detection
+    processed_boxes = []
+    overlap_threshold = 0.2
+    
+    for i, box_info in enumerate(detected_boxes):
+        x1, y1, x2, y2 = box_info['coords']
+        w = box_info['width']
+        h = box_info['height']
+        center_x = box_info['center_x']
+        center_y = box_info['center_y']
+        
+        # Check for overlaps
+        has_overlap = False
+        for j, other_box in enumerate(detected_boxes):
+            if i != j and calculate_iou(box_info, other_box) > overlap_threshold:
+                has_overlap = True
+                break
+        
+        # Adjust expansion strategy
+        base_expansion = 0.03
+        max_expansion = 0.05
+        
+        is_standing = h > 1.5 * w
+        is_sitting = 0.8 <= h/w <= 1.2
+        is_abnormal_size = (h * w) > (avg_area * 1.5) or (h * w) < (avg_area * 0.5)
+        
+        if has_overlap:
+            h_expansion = w_expansion = base_expansion * 0.8
+        else:
+            if is_standing:
+                h_expansion = min(base_expansion * 1.2, max_expansion)
+                w_expansion = base_expansion
+            elif is_sitting:
+                h_expansion = w_expansion = base_expansion
+            else:
+                h_expansion = w_expansion = base_expansion * 0.9
+        
+        # Position compensation
+        if center_x < img_width * 0.2 or center_x > img_width * 0.8:
+            w_expansion *= 0.9
+        
+        if is_abnormal_size:
+            h_expansion *= 0.8
+            w_expansion *= 0.8
+        
+        # Calculate final bounding box
+        expansion_w = w * w_expansion
+        expansion_h = h * h_expansion
+        
+        new_x1 = max(0, center_x - (w + expansion_w)/2)
+        new_y1 = max(0, center_y - (h + expansion_h)/2)
+        new_x2 = min(img_width, center_x + (w + expansion_w)/2)
+        new_y2 = min(img_height, center_y + (h + expansion_h)/2)
+        
+        # Crop and process image
+        cropped_image = image.crop((int(new_x1), int(new_y1), 
+                                  int(new_x2), int(new_y2)))
+        
+        processed_image = enhanced_preprocess(
+            cropped_image,
+            is_standing=is_standing,
+            has_overlap=has_overlap
+        )
+        
+        processed_boxes.append((
+            processed_image, 
+            box_info['confidence'],
+            [new_x1, new_y1, new_x2, new_y2], 
+            True
+        ))
+    
+    return processed_boxes
+
+@spaces.GPU
+def predict(image):
+    """
+    Main prediction function that handles both single and multiple dog detection.
+    Args:
+        image: PIL Image or numpy array
+    Returns:
+        tuple: (html_output, annotated_image, initial_state)
+    """
+    if image is None:
+        return format_hint_html("Please upload an image to start."), None, None
+
+    try:
+        if isinstance(image, np.ndarray):
+            image = Image.fromarray(image)
+
+        # 檢測圖片中的物體
+        dogs = detect_multiple_dogs(image)
+        color_scheme = get_color_scheme(len(dogs) == 1)
+
+        # 準備標註
+        annotated_image = image.copy()
+        draw = ImageDraw.Draw(annotated_image)
+
+        try:
+            font = ImageFont.truetype("arial.ttf", 24)
+        except:
+            font = ImageFont.load_default()
+
+        dogs_info = ""
+
+        # 處理每個檢測到的物體
+        for i, (cropped_image, detection_confidence, box, is_dog) in enumerate(dogs):
+            print(f"Predict processing - Object {i+1}:")
+            print(f"    Is dog: {is_dog}")
+            print(f"    Detection confidence: {detection_confidence:.4f}")
+
+            # 如果是狗且進行品種預測，在這裡也加入打印語句
+            if is_dog:
+                top1_prob, topk_breeds, relative_probs = predict_single_dog(cropped_image)
+                print(f"    Breed prediction - Top probability: {top1_prob:.4f}")
+                print(f"    Top breeds: {topk_breeds[:3]}")
+            color = color_scheme if len(dogs) == 1 else color_scheme[i % len(color_scheme)]
+
+            # 繪製框和標籤
+            draw.rectangle(box, outline=color, width=4)
+            label = f"Dog {i+1}" if is_dog else f"Object {i+1}"
+            label_bbox = draw.textbbox((0, 0), label, font=font)
+            label_width = label_bbox[2] - label_bbox[0]
+            label_height = label_bbox[3] - label_bbox[1]
+
+            # 繪製標籤背景和文字
+            label_x = box[0] + 5
+            label_y = box[1] + 5
+            draw.rectangle(
+                [label_x - 2, label_y - 2, label_x + label_width + 4, label_y + label_height + 4],
+                fill='white',
+                outline=color,
+                width=2
+            )
+            draw.text((label_x, label_y), label, fill=color, font=font)
+
+            try:
+                # 首先檢查是否為狗
+                if not is_dog:
+                    dogs_info += format_not_dog_message(color, i+1)
+                    continue
+
+                # 如果是狗，進行品種預測
+                top1_prob, topk_breeds, relative_probs = predict_single_dog(cropped_image)
+                combined_confidence = detection_confidence * top1_prob
+
+                # 根據信心度決定輸出格式
+                if combined_confidence < 0.15:
+                    dogs_info += format_unknown_breed_message(color, i+1)
+                elif top1_prob >= 0.4:
+                    breed = topk_breeds[0]
+                    description = get_dog_description(breed)
+                    if description is None:
+                        description = {
+                            "Name": breed,
+                            "Size": "Unknown",
+                            "Exercise Needs": "Unknown",
+                            "Grooming Needs": "Unknown",
+                            "Care Level": "Unknown",
+                            "Good with Children": "Unknown",
+                            "Description": f"Identified as {breed.replace('_', ' ')}"
+                        }
+                    dogs_info += format_single_dog_result(breed, description, color)
+                else:
+                    dogs_info += format_multiple_breeds_result(
+                        topk_breeds,
+                        relative_probs,
+                        color,
+                        i+1,
+                        lambda breed: get_dog_description(breed) or {
+                            "Name": breed,
+                            "Size": "Unknown",
+                            "Exercise Needs": "Unknown",
+                            "Grooming Needs": "Unknown",
+                            "Care Level": "Unknown",
+                            "Good with Children": "Unknown",
+                            "Description": f"Identified as {breed.replace('_', ' ')}"
+                        }
+                    )
+            except Exception as e:
+                print(f"Error formatting results for dog {i+1}: {str(e)}")
+                dogs_info += format_unknown_breed_message(color, i+1)
+
+        # 包裝最終的HTML輸出
+        html_output = format_multi_dog_container(dogs_info)
+
+        # 準備初始狀態
+        initial_state = {
+            "dogs_info": dogs_info,
+            "image": annotated_image,
+            "is_multi_dog": len(dogs) > 1,
+            "html_output": html_output
+        }
+
+        return html_output, annotated_image, initial_state
+
+    except Exception as e:
+        error_msg = f"An error occurred: {str(e)}\n\nTraceback:\n{traceback.format_exc()}"
+        print(error_msg)
+        return format_hint_html(error_msg), None, None
+
+
+def show_details_html(choice, previous_output, initial_state):
+    """
+    Generate detailed HTML view for a selected breed.
+
+    Args:
+        choice: str, Selected breed option
+        previous_output: str, Previous HTML output
+        initial_state: dict, Current state information
+
+    Returns:
+        tuple: (html_output, gradio_update, updated_state)
+    """
+    if not choice:
+        return previous_output, gr.update(visible=True), initial_state
+
+    try:
+        breed = choice.split("More about ")[-1]
+        description = get_dog_description(breed)
+        html_output = format_breed_details_html(description, breed)
+
+        # Update state
+        initial_state["current_description"] = html_output
+        initial_state["original_buttons"] = initial_state.get("buttons", [])
+
+        return html_output, gr.update(visible=True), initial_state
+
+    except Exception as e:
+        error_msg = f"An error occurred while showing details: {e}"
+        print(error_msg)
+        return format_hint_html(error_msg), gr.update(visible=True), initial_state
+
+def main():
+    with gr.Blocks(css=get_css_styles()) as iface:
+        # Header HTML
+
+        gr.HTML("""
+        <header style='text-align: center; padding: 20px; margin-bottom: 20px;'>
+            <h1 style='font-size: 2.5em; margin-bottom: 10px; color: #2D3748;'>
+                🐾 PawMatch AI
+            </h1>
+            <h2 style='font-size: 1.2em; font-weight: normal; color: #4A5568; margin-top: 5px;'>
+                Your Smart Dog Breed Guide
+            </h2>
+            <div style='width: 50px; height: 3px; background: linear-gradient(90deg, #4299e1, #48bb78); margin: 15px auto;'></div>
+            <p style='color: #718096; font-size: 0.9em;'>
+                Powered by AI • Breed Recognition • Smart Matching • Companion Guide
+            </p>
+        </header>
+        """)
+
+        # 先創建歷史組件實例（但不創建標籤頁）
+        history_component = create_history_component()
+
+        with gr.Tabs():
+            # 1. 品種檢測標籤頁
+            example_images = [
+                'Border_Collie.jpg',
+                'Golden_Retriever.jpeg',
+                'Saint_Bernard.jpeg',
+                'Samoyed.jpeg',
+                'French_Bulldog.jpeg'
+            ]
+            detection_components = create_detection_tab(predict, example_images)
+
+            # 2. 品種比較標籤頁
+            comparison_components = create_comparison_tab(
+                dog_breeds=dog_breeds,
+                get_dog_description=get_dog_description,
+                breed_health_info=breed_health_info,
+                breed_noise_info=breed_noise_info
+            )
+
+            # 3. 品種推薦標籤頁
+            recommendation_components = create_recommendation_tab(
+                UserPreferences=UserPreferences,
+                get_breed_recommendations=get_breed_recommendations,
+                format_recommendation_html=format_recommendation_html,
+                history_component=history_component
+            )
+
+
+            # 4. 最後創建歷史記錄標籤頁
+            create_history_tab(history_component)
+
+        # Footer
+        gr.HTML('''
+            <div style="
+                display: flex;
+                align-items: center;
+                justify-content: center;
+                gap: 20px;
+                padding: 20px 0;
+            ">
+                <p style="
+                    font-family: 'Arial', sans-serif;
+                    font-size: 14px;
+                    font-weight: 500;
+                    letter-spacing: 2px;
+                    background: linear-gradient(90deg, #555, #007ACC);
+                    -webkit-background-clip: text;
+                    -webkit-text-fill-color: transparent;
+                    margin: 0;
+                    text-transform: uppercase;
+                    display: inline-block;
+                ">EXPLORE THE CODE →</p>
+                <a href="https://github.com/Eric-Chung-0511/Learning-Record/tree/main/Data%20Science%20Projects/PawMatchAI" style="text-decoration: none;">
+                    <img src="https://img.shields.io/badge/GitHub-PawMatch_AI-007ACC?logo=github&style=for-the-badge">
+                </a>
+            </div>
+        ''')
+
+    return iface
+
+if __name__ == "__main__":
+    iface = main()
+    iface.launch()

model_architecture.py

@@ -0,0 +1,315 @@
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import timm
+import numpy as np
+
+dog_breeds = ["Afghan_Hound", "African_Hunting_Dog", "Airedale", "American_Staffordshire_Terrier",
+              "Appenzeller", "Australian_Terrier", "Bedlington_Terrier", "Bernese_Mountain_Dog", "Bichon_Frise",
+              "Blenheim_Spaniel", "Border_Collie", "Border_Terrier", "Boston_Bull", "Bouvier_Des_Flandres",
+              "Brabancon_Griffon", "Brittany_Spaniel", "Cardigan", "Chesapeake_Bay_Retriever",
+              "Chihuahua", "Dachshund", "Dandie_Dinmont", "Doberman", "English_Foxhound", "English_Setter",
+              "English_Springer", "EntleBucher", "Eskimo_Dog", "French_Bulldog", "German_Shepherd",
+              "German_Short-Haired_Pointer", "Gordon_Setter", "Great_Dane", "Great_Pyrenees",
+              "Greater_Swiss_Mountain_Dog","Havanese", "Ibizan_Hound", "Irish_Setter", "Irish_Terrier",
+              "Irish_Water_Spaniel", "Irish_Wolfhound", "Italian_Greyhound", "Japanese_Spaniel",
+              "Kerry_Blue_Terrier", "Labrador_Retriever", "Lakeland_Terrier", "Leonberg", "Lhasa",
+              "Maltese_Dog", "Mexican_Hairless", "Newfoundland", "Norfolk_Terrier", "Norwegian_Elkhound",
+              "Norwich_Terrier", "Old_English_Sheepdog", "Pekinese", "Pembroke", "Pomeranian",
+              "Rhodesian_Ridgeback", "Rottweiler", "Saint_Bernard", "Saluki", "Samoyed",
+              "Scotch_Terrier", "Scottish_Deerhound", "Sealyham_Terrier", "Shetland_Sheepdog", "Shiba_Inu",
+              "Shih-Tzu", "Siberian_Husky", "Staffordshire_Bullterrier", "Sussex_Spaniel",
+              "Tibetan_Mastiff", "Tibetan_Terrier", "Walker_Hound", "Weimaraner",
+              "Welsh_Springer_Spaniel", "West_Highland_White_Terrier", "Yorkshire_Terrier",
+              "Affenpinscher", "Basenji", "Basset", "Beagle", "Black-and-Tan_Coonhound", "Bloodhound",
+              "Bluetick", "Borzoi", "Boxer", "Briard", "Bull_Mastiff", "Cairn", "Chow", "Clumber",
+              "Cocker_Spaniel", "Collie", "Curly-Coated_Retriever", "Dhole", "Dingo",
+              "Flat-Coated_Retriever", "Giant_Schnauzer", "Golden_Retriever", "Groenendael", "Keeshond",
+              "Kelpie", "Komondor", "Kuvasz", "Malamute", "Malinois", "Miniature_Pinscher",
+              "Miniature_Poodle", "Miniature_Schnauzer", "Otterhound", "Papillon", "Pug", "Redbone",
+              "Schipperke", "Silky_Terrier", "Soft-Coated_Wheaten_Terrier", "Standard_Poodle",
+              "Standard_Schnauzer", "Toy_Poodle", "Toy_Terrier", "Vizsla", "Whippet",
+              "Wire-Haired_Fox_Terrier"]
+
+
+class MorphologicalFeatureExtractor(nn.Module):
+
+    def __init__(self, in_features):
+        super().__init__()
+
+        # 基礎特徵維度設置
+        self.reduced_dim = in_features // 4
+        self.spatial_size = max(7, int(np.sqrt(self.reduced_dim // 64)))
+
+        # 1. 特徵空間轉換器：將一維特徵轉換為二維空間表示
+        self.dimension_transformer = nn.Sequential(
+            nn.Linear(in_features, self.spatial_size * self.spatial_size * 64),
+            nn.LayerNorm(self.spatial_size * self.spatial_size * 64),
+            nn.ReLU()
+        )
+
+        # 2. 形態特徵分析器：分析具體的形態特徵
+        self.morphological_analyzers = nn.ModuleDict({
+            # 體型分析器：分析整體比例和大小
+            'body_proportion': nn.Sequential(
+                # 使用大卷積核捕捉整體體型特徵
+                nn.Conv2d(64, 128, kernel_size=7, padding=3),
+                nn.BatchNorm2d(128),
+                nn.ReLU(),
+                # 使用較小的卷積核精煉特徵
+                nn.Conv2d(128, 128, kernel_size=3, padding=1),
+                nn.BatchNorm2d(128),
+                nn.ReLU()
+            ),
+
+            # 頭部特徵分析器：關注耳朵、臉部等
+            'head_features': nn.Sequential(
+                # 中等大小的卷積核，適合分析頭部結構
+                nn.Conv2d(64, 128, kernel_size=5, padding=2),
+                nn.BatchNorm2d(128),
+                nn.ReLU(),
+                # 小卷積核捕捉細節
+                nn.Conv2d(128, 128, kernel_size=3, padding=1),
+                nn.BatchNorm2d(128),
+                nn.ReLU()
+            ),
+
+            # 尾部特徵分析器
+            'tail_features': nn.Sequential(
+                nn.Conv2d(64, 128, kernel_size=5, padding=2),
+                nn.BatchNorm2d(128),
+                nn.ReLU(),
+                nn.Conv2d(128, 128, kernel_size=3, padding=1),
+                nn.BatchNorm2d(128),
+                nn.ReLU()
+            ),
+
+            # 毛髮特徵分析器：分析毛髮長度、質地等
+            'fur_features': nn.Sequential(
+                # 使用多個小卷積核捕捉毛髮紋理
+                nn.Conv2d(64, 128, kernel_size=3, padding=1),
+                nn.BatchNorm2d(128),
+                nn.ReLU(),
+                nn.Conv2d(128, 128, kernel_size=3, padding=1),
+                nn.BatchNorm2d(128),
+                nn.ReLU()
+            ),
+
+            # 顏色特徵分析器：分析顏色分佈
+            'color_pattern': nn.Sequential(
+                # 第一層：捕捉基本顏色分布
+                nn.Conv2d(64, 128, kernel_size=3, padding=1),
+                nn.BatchNorm2d(128),
+                nn.ReLU(),
+
+                # 第二層：分析顏色模式和花紋
+                nn.Conv2d(128, 128, kernel_size=3, padding=1),
+                nn.BatchNorm2d(128),
+                nn.ReLU(),
+
+                # 第三層：整合顏色信息
+                nn.Conv2d(128, 128, kernel_size=1),
+                nn.BatchNorm2d(128),
+                nn.ReLU()
+            )
+        })
+
+        # 3. 特徵注意力機制：動態關注不同特徵
+        self.feature_attention = nn.MultiheadAttention(
+            embed_dim=128,
+            num_heads=8,
+            dropout=0.1,
+            batch_first=True
+        )
+
+        # 4. 特徵關係分析器：分析不同特徵之間的關係
+        self.relation_analyzer = nn.Sequential(
+            nn.Linear(128 * 5, 256),  # 4個特徵分析器的輸出
+            nn.LayerNorm(256),
+            nn.ReLU(),
+            nn.Linear(256, 128),
+            nn.LayerNorm(128),
+            nn.ReLU()
+        )
+
+        # 5. 特徵整合器：將所有特徵智能地組合在一起
+        self.feature_integrator = nn.Sequential(
+            nn.Linear(128 * 6, in_features),  # 5個原始特徵 + 1個關係特徵
+            nn.LayerNorm(in_features),
+            nn.ReLU()
+        )
+
+    def forward(self, x):
+        batch_size = x.size(0)
+
+        # 1. 將特徵轉換為空間形式
+        spatial_features = self.dimension_transformer(x).view(
+            batch_size, 64, self.spatial_size, self.spatial_size
+        )
+
+        # 2. 分析各種形態特徵
+        morphological_features = {}
+        for name, analyzer in self.morphological_analyzers.items():
+            # 提取特定形態特徵
+            features = analyzer(spatial_features)
+            # 使用自適應池化統一特徵大小
+            pooled_features = F.adaptive_avg_pool2d(features, (1, 1))
+            # 重塑特徵為向量形式
+            morphological_features[name] = pooled_features.view(batch_size, -1)
+
+        # 3. 特徵注意力處理
+        # 將所有特徵堆疊成序列
+        stacked_features = torch.stack(list(morphological_features.values()), dim=1)
+        # 應用注意力機制
+        attended_features, _ = self.feature_attention(
+            stacked_features, stacked_features, stacked_features
+        )
+
+        # 4. 分析特徵之間的關係
+        # 將所有特徵連接起來
+        combined_features = torch.cat(list(morphological_features.values()), dim=1)
+        # 提取特徵間的關係
+        relation_features = self.relation_analyzer(combined_features)
+
+        # 5. 特徵整合
+        # 將原始特徵和關係特徵結合
+        final_features = torch.cat([
+            *morphological_features.values(),
+            relation_features
+        ], dim=1)
+
+        # 6. 最終整合
+        integrated_features = self.feature_integrator(final_features)
+
+        # 添加殘差連接
+        return integrated_features + x
+
+
+class MultiHeadAttention(nn.Module):
+
+    def __init__(self, in_dim, num_heads=8):
+        """
+        Initializes the MultiHeadAttention module.
+        Args:
+            in_dim (int): Dimension of the input features.
+            num_heads (int): Number of attention heads. Defaults to 8.
+        """
+        super().__init__()
+        self.num_heads = num_heads
+        self.head_dim = max(1, in_dim // num_heads)  
+        self.scaled_dim = self.head_dim * num_heads  
+        self.fc_in = nn.Linear(in_dim, self.scaled_dim)  
+        self.query = nn.Linear(self.scaled_dim, self.scaled_dim)  # Query projection
+        self.key = nn.Linear(self.scaled_dim, self.scaled_dim)  # Key projection
+        self.value = nn.Linear(self.scaled_dim, self.scaled_dim)  # Value projection
+        self.fc_out = nn.Linear(self.scaled_dim, in_dim)  # Linear layer to project output back to in_dim
+
+    def forward(self, x):
+        """
+        Forward pass for multi-head attention mechanism.
+        Args:
+            x (Tensor): Input tensor of shape (batch_size, input_dim).
+            x 是 (N,D), N:批次大小, D:輸入特徵維度
+        Returns:
+            Tensor: Output tensor after applying attention mechanism.
+        """
+        N = x.shape[0]  # Batch size
+        x = self.fc_in(x)  # Project input to scaled_dim
+        q = self.query(x).view(N, self.num_heads, self.head_dim)  # Compute queries
+        k = self.key(x).view(N, self.num_heads, self.head_dim)  # Compute keys
+        v = self.value(x).view(N, self.num_heads, self.head_dim)  # Compute values
+
+        # Calculate attention scores
+        energy = torch.einsum("nqd,nkd->nqk", [q, k])  
+        attention = F.softmax(energy / (self.head_dim ** 0.5), dim=2)  # Apply softmax with scaling
+
+        # Compute weighted sum of values based on attention scores
+        out = torch.einsum("nqk,nvd->nqd", [attention, v]) 
+        out = out.reshape(N, self.scaled_dim)  # Concatenate all heads
+        out = self.fc_out(out)  # Project back to original input dimension
+        return out
+
+
+class BaseModel(nn.Module):
+
+    def __init__(self, num_classes, device='cuda' if torch.cuda.is_available() else 'cpu'):
+        super().__init__()
+        self.device = device
+
+        # 1. Initialize backbone
+        self.backbone = timm.create_model(
+                'convnextv2_base',
+                pretrained=True,
+                num_classes=0
+        )
+
+        # 2. 使用測試數據來確定實際的特徵維度
+        with torch.no_grad():  
+            dummy_input = torch.randn(1, 3, 224, 224)
+            features = self.backbone(dummy_input)
+
+            if len(features.shape) > 2:
+                features = features.mean([-2, -1])
+
+            self.feature_dim = features.shape[1]
+
+        print(f"Feature Dimension from V2 backbone: {self.feature_dim}")
+
+        # 3. Setup multi-head attention layer
+        self.num_heads = max(1, min(8, self.feature_dim // 64))
+        self.attention = MultiHeadAttention(self.feature_dim, num_heads=self.num_heads)
+
+        # 4. Setup classifier
+        self.classifier = nn.Sequential(
+            nn.LayerNorm(self.feature_dim),
+            nn.Dropout(0.3),
+            nn.Linear(self.feature_dim, num_classes)
+        )
+
+        self.morphological_extractor = MorphologicalFeatureExtractor(
+            in_features=self.feature_dim
+        )
+
+        self.feature_fusion = nn.Sequential(
+            nn.Linear(self.feature_dim * 3, self.feature_dim),  
+            nn.LayerNorm(self.feature_dim),
+            nn.ReLU(),
+            nn.Linear(self.feature_dim, self.feature_dim),
+            nn.LayerNorm(self.feature_dim),
+            nn.ReLU()
+        )
+
+    def forward(self, x):
+        """
+        Forward propagation process, combining V2's FCCA and multi-head attention mechanism
+        Args:
+            x (Tensor): Input image tensor of shape [batch_size, channels, height, width]
+        Returns:
+            Tuple[Tensor, Tensor]: Classification logits and attention features
+        """
+        x = x.to(self.device)
+
+        # 1. Extract base features
+        features = self.backbone(x)
+        if len(features.shape) > 2:
+            features = features.mean([-2, -1])
+
+        # 2. Extract morphological features (including all detail features)
+        morphological_features = self.morphological_extractor(features)
+
+        # 3. Feature fusion (note dimension alignment with new fusion layer)
+        combined_features = torch.cat([
+            features,  # Original features
+            morphological_features,  # Morphological features
+            features * morphological_features  # Feature interaction information
+        ], dim=1)
+        fused_features = self.feature_fusion(combined_features)
+
+        # 4. Apply attention mechanism
+        attended_features = self.attention(fused_features)
+
+        # 5. Final classifier
+        logits = self.classifier(attended_features)
+
+        return logits, attended_features