Updated "non-understood" threshold.

app.py

@@ -21,7 +21,7 @@ def predict(img) -> Tuple[Dict, float]:
         y_logits = MODEL(X)
         y_prob = torch.softmax(y_logits, dim=1)
     # Prediction threshold for non understood images.
-    if y_prob.max() < .10:
+    if y_prob.max() < .15:
         raise gr.Error("This image might not be of food.")
     # Float casting due to Gradio's assumption that numpy objects
     # should be iterated. Running `tolist()` prior to this