fix: 兼容模型加载失败场景，手动构造 image-classification pipeline

- 增加模型加载的异常处理，当 pipeline 直接加载失败时回退为
    手动加载 AutoImageProcessor 和 AutoModelForImageClassification
  - 添加 torchvision 依赖

app.py

@@ -12,7 +12,7 @@ from PIL import Image
 # 国内 HuggingFace 镜像加速
 os.environ.setdefault("HF_ENDPOINT", "https://hf-mirror.com")
 
-from transformers import pipeline  # noqa: E402
+from transformers import AutoImageProcessor, AutoModelForImageClassification, pipeline  # noqa: E402
 
 # ─── Disease Label Mapping ──────────────────────────────────────────────────
 # 模型输出 LABEL_0 ~ LABEL_6，映射为实际病害名称
@@ -90,10 +90,22 @@ html, body, [class*="css"] {
 @st.cache_resource
 def load_model():
     """加载 HuggingFace 模型（首次运行自动下载，约 343MB）"""
-    classifier = pipeline(
-        "image-classification",
-        model="Dmitry43243242/banana-disease-leaf-model",
-    )
+    model_name = "Dmitry43243242/banana-disease-leaf-model"
+    try:
+        classifier = pipeline(
+            "image-classification",
+            model=model_name,
+        )
+    except ValueError:
+        # 模型的 preprocessor_config.json 中 image_processor_type 可能是旧版名称
+        # 手动加载并构造 pipeline
+        processor = AutoImageProcessor.from_pretrained(
+            model_name, trust_remote_code=False
+        )
+        model = AutoModelForImageClassification.from_pretrained(model_name)
+        classifier = pipeline(
+            "image-classification", model=model, image_processor=processor
+        )
     return classifier