Update app.py

app.py

@@ -1,32 +1,30 @@
-import os
-import traceback
-
+import os, sys, types, traceback
 import gradio as gr
 import torch
 from huggingface_hub import snapshot_download
-from fastai.vision.all import *  # importante: registra clases/transforms de fastai
 
-# --- Parche defensivo por si el pickle trae refs de Pillow ---
+# Importa fastai (con Plum 2) -> fasttransform ya no revienta
+from fastai.vision.all import PILImage
+
+# --------- Alias para pickles antiguos que esperan plum.function ---------
 try:
-    from PIL import Image
-    if not hasattr(Image, "Resampling") and hasattr(Image, "NEAREST"):
-        class _Resampling:
-            NEAREST = Image.NEAREST
-            BILINEAR = Image.BILINEAR
-            BICUBIC = Image.BICUBIC
-            LANCZOS = getattr(Image, "LANCZOS", Image.BICUBIC)
-        Image.Resampling = _Resampling
-    if not hasattr(Image, "ANTIALIAS") and hasattr(Image, "Resampling"):
-        Image.ANTIALIAS = Image.Resampling.LANCZOS
+    import plum
+    from plum import Function  # en Plum 2 existe aquí
+    if "plum.function" not in sys.modules:
+        m = types.ModuleType("plum.function")
+        m.Function = Function
+        sys.modules["plum.function"] = m
 except Exception:
-    print("Warning: Pillow patch failed:\n", traceback.format_exc())
+    print("Error preparando alias plum.function:\n", traceback.format_exc())
+    raise
 
+# --------- Loader robusto ---------
 def load_fastai_from_hub(repo_id: str, filename: str = "model.pkl"):
     repo_dir = snapshot_download(repo_id)
     pkl_path = os.path.join(repo_dir, filename)
     print("Loading:", pkl_path, "exists:", os.path.exists(pkl_path), "size:", os.path.getsize(pkl_path))
 
-    # CLAVE en torch 2.6: para cargar un Learner completo necesitas weights_only=False
+    # PyTorch >=2.6: hay que forzar weights_only=False para cargar objetos completos
     learn = torch.load(pkl_path, map_location="cpu", weights_only=False)
     learn.dls.cpu()
     return learn
@@ -40,13 +38,9 @@ def predict(img):
     pred, pred_idx, probs = learner.predict(img)
     return {labels[i]: float(probs[i]) for i in range(len(labels))}
 
-# evita romper si los ejemplos no existen en el repo del Space
-example_files = [f for f in ["NormalT.jpeg", "AnormalT.jpeg"] if os.path.exists(f)]
-
 demo = gr.Interface(
     fn=predict,
     inputs=gr.Image(type="pil"),
     outputs=gr.Label(num_top_classes=3),
-    examples=example_files if example_files else None,
 )
 demo.launch()