Skip learn.predict and run manual inference

app.py

@@ -31,9 +31,20 @@ import pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)
 #/export
 from fastai.vision.all import *
 import gradio as gr
+import torch
 
 learn = load_learner('model-2.pkl')
 
+# Use valid batch to get input size and device for manual inference (avoids learn.predict
+# passing (Image, dict) into transforms and triggering TypeError: 'PILImage' + 'dict').
+try:
+    _inf_batch = next(iter(learn.dls.valid))
+    _INFERENCE_DEVICE = _inf_batch[0].device
+    _INFERENCE_SIZE = _inf_batch[0].shape[-1]
+except Exception:
+    _INFERENCE_DEVICE = torch.device("cpu")
+    _INFERENCE_SIZE = 224
+
 #Pydantic Warnings
 # UnsupportedFieldAttributeWarning: The 'repr' attribute with value False ...
 # UnsupportedFieldAttributeWarning: The 'frozen' attribute with value True ...
@@ -89,12 +100,23 @@ learn = load_learner('model-2.pkl')
 #     img = PILImage.create(img)
 #     pred, pred_idx, probs = learn.predict(img)
 
+def _preprocess_for_learner(pil_img):
+    """Turn PIL image into a batch tensor matching the learner's expected size and normalization."""
+    from torchvision import transforms as T
+    # Match typical fastai/ImageNet preprocessing
+    transform = T.Compose([
+        T.Resize((_INFERENCE_SIZE, _INFERENCE_SIZE)),
+        T.ToTensor(),
+        T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
+    ])
+    x = transform(pil_img).unsqueeze(0).to(_INFERENCE_DEVICE)
+    return x
+
+
 def classify_flower(img):
     if img is None:
         return None
-    # Gradio can pass PIL Image, numpy array, or dict (e.g. {"path": "..."}).
-    # Fastai's transforms expect its own Image type; raw PIL/dict causes:
-    # TypeError: unsupported operand type(s) for +: 'Image' and 'dict'
+    # Normalize Gradio input to PIL (can be dict with "path" or "image", or numpy, or PIL).
     if isinstance(img, dict):
         path = img.get("path")
         if isinstance(path, str):
@@ -104,8 +126,20 @@ def classify_flower(img):
             img = img["image"]
         else:
             return None
-    img = PILImage.create(img)
-    pred, pred_idx, probs = learn.predict(img)
+    from PIL import Image as PILImageModule
+    if not isinstance(img, PILImageModule.Image):
+        img = np.asarray(img)
+        if img.ndim == 2:
+            img = np.stack([img] * 3, axis=-1)
+        img = PILImageModule.fromarray(img).convert("RGB")
+    else:
+        img = img.convert("RGB")
+    # Bypass learn.predict() to avoid (PILImage, dict) in the transform pipeline.
+    x = _preprocess_for_learner(img)
+    learn.model.eval()
+    with torch.no_grad():
+        logits = learn.model(x)
+    probs = torch.softmax(logits, dim=1)[0].cpu().numpy()
     return {
         learn.dls.vocab[i]: float(probs[i])
         for i in range(len(probs))