hi

app.py

@@ -5,23 +5,34 @@ from torchvision import transforms, models
 from PIL import Image
 import numpy as np
 
-# === 使用 pytorch-grad-cam ===
+# === 使用 grad-cam (套件名稱已修正) ===
 from pytorch_grad_cam import GradCAM
 from pytorch_grad_cam.utils.model_targets import ClassifierOutputTarget
 from pytorch_grad_cam.utils.image import show_cam_on_image
 
+# 設定設備
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 
 # ==================== 1. 載入肺炎模型 (ResNet50) ====================
-model_pneumonia = models.resnet50(pretrained=False)
+# 使用 weights=None 取代已棄用的 pretrained=False
+model_pneumonia = models.resnet50(weights=None)
 num_ftrs = model_pneumonia.fc.in_features
-model_pneumonia.fc = nn.Linear(num_ftrs, 2)  # NORMAL vs PNEUMONIA
+model_pneumonia.fc = nn.Linear(num_ftrs, 2)  # 兩類：NORMAL vs PNEUMONIA
 
 model_path_pneumonia = "best_pneumonia_model.pth"
-model_pneumonia.load_state_dict(torch.load(model_path_pneumonia, map_location=device))
+
+# 載入模型權重
+try:
+    # 建議使用 weights_only=False 以確保舊版模型權重能正確讀取
+    model_pneumonia.load_state_dict(torch.load(model_path_pneumonia, map_location=device, weights_only=False))
+    print(f"成功載入模型權重：{model_path_pneumonia}")
+except Exception as e:
+    print(f"模型載入失敗，請確認檔案是否存在：{e}")
+
 model_pneumonia.to(device)
 model_pneumonia.eval()
 
+# 設定 Grad-CAM 目標層 (ResNet50 的最後一層捲積層)
 target_layers = [model_pneumonia.layer4[-1]]
 
 # ==================== 2. 影像前處理 ====================
@@ -37,9 +48,11 @@ def predict_pneumonia(image):
     if image is None:
         return "請上傳影像", None, "無影像"
     
+    # 轉換圖片格式
     pil_img = Image.fromarray(image).convert("RGB")
     input_tensor = img_transform(pil_img).unsqueeze(0).to(device)
     
+    # 模型推論
     with torch.no_grad():
         output = model_pneumonia(input_tensor)
         probs = torch.softmax(output, dim=1).cpu().numpy()[0]
@@ -49,41 +62,44 @@ def predict_pneumonia(image):
     label = "PNEUMONIA" if pred_class == 1 else "NORMAL"
     result_text = f"肺炎檢測結果：{label} (信心度：{confidence:.1%})"
     
-    # Grad-CAM
+    # --- Grad-CAM 視覺化 ---
+    # 注意：GradCAM 在計算時需要梯度，所以不放在 no_grad 中
     cam = GradCAM(model=model_pneumonia, target_layers=target_layers)
     targets = [ClassifierOutputTarget(pred_class)]
+    
+    # 產生熱圖
     grayscale_cam = cam(input_tensor=input_tensor, targets=targets)[0]
     
-    visualization = show_cam_on_image(np.array(pil_img.resize((224,224))) / 255.0,
-                                      grayscale_cam,
-                                      use_rgb=True)
-    visualization = (visualization * 255).astype(np.uint8)
+    # 將原始圖片縮放到 224x224 以符合熱圖尺寸
+    input_float_img = np.array(pil_img.resize((224, 224))).astype(np.float32) / 255.0
+    visualization = show_cam_on_image(input_float_img, grayscale_cam, use_rgb=True)
     
-    # 風險評估
+    # --- 風險評估建議 ---
     if label == "PNEUMONIA" and confidence > 0.7:
-        risk_assessment = "🚨 **高風險**：高度懷疑肺炎，建議立即就醫並進行抗生素治療。"
+        risk_assessment = "🚨 **高風險**：高度懷疑肺炎，建議立即就醫並進行進一步檢查。"
     elif label == "PNEUMONIA":
-        risk_assessment = "⚠️ **中等風險**：疑似肺炎，建議盡快就醫評估。"
+        risk_assessment = "⚠️ **中等風險**：疑似肺炎徵象，建議儘快諮詢醫療專業人員。"
     else:
-        risk_assessment = "✅ **正常**：目前未檢測到肺炎徵象，但仍需注意症狀變化。"
+        risk_assessment = "✅ **正常**：目前影像未顯示明顯肺炎特徵，若有呼吸道症狀仍請留意。"
     
     return result_text, visualization, risk_assessment
 
 # ==================== 4. Gradio 介面 ====================
-with gr.Blocks(title="肺炎檢測系統") as demo:
-    gr.Markdown("# 🏥 肺炎檢測系統")
-    gr.Markdown("基於深度學習的胸部X光影像肺炎檢測系統")
+with gr.Blocks(title="肺炎檢測 AI 輔助系統") as demo:
+    gr.Markdown("# 🏥 肺炎檢測 AI 輔助系統")
+    gr.Markdown("這是一個基於深度學習的胸部 X 光影像分析工具。請上傳一張 X 光片，模型將分析是否存在肺炎徵兆，並顯示模型關注的區域。")
     
     with gr.Row():
         with gr.Column(scale=1):
-            img_input = gr.Image(label="上傳胸部X光影像", type="numpy")
-            btn = gr.Button("開始檢測", variant="primary")
+            img_input = gr.Image(label="1. 上傳胸部 X 光影像", type="numpy")
+            btn = gr.Button("🔍 開始檢測", variant="primary")
         
         with gr.Column(scale=1):
-            pneumonia_out = gr.Textbox(label="肺炎檢測結果")
-            cam_out = gr.Image(label="模型關注區域熱圖（紅色區域為模型���點關注）")
-            risk_out = gr.Markdown(label="風險評估建議")
+            pneumonia_out = gr.Textbox(label="檢測結果")
+            cam_out = gr.Image(label="模型關注區域（紅色代表重點關注部位）")
+            risk_out = gr.Markdown(label="專業建議與風險評估")
     
+    # 設定按鈕邏輯
     btn.click(
         fn=predict_pneumonia,
         inputs=[img_input],
@@ -91,8 +107,9 @@ with gr.Blocks(title="肺炎檢測系統") as demo:
     )
     
     gr.Markdown("---")
-    gr.Markdown("⚠️ 本系統僅供研究與教育展示使用，非正式醫療診斷工具。實際診斷請以醫師判斷為準。")
+    gr.Markdown("⚠️ **免責聲明**：本工具僅作為研究及技術展示用途，**不可**取代專業醫師的醫療診斷。如果您感到身體不適，請務必尋求正式醫療協助。")
 
-# ==================== 5. 啟動 ====================
+# ==================== 5. 啟動服務 ====================
 if __name__ == "__main__":
-    demo.launch(share=True)
+    # 部署在 Hugging Face 時 share=True 是可選的，建議維持預設
+    demo.launch()

requirements.txt

@@ -3,6 +3,5 @@ torch
 torchvision
 pillow
 numpy
-pytorch-grad-cam
-opencv-python
-matplotlib
+grad-cam
+opencv-python-headless