app.py

app.py

@@ -1,20 +1,19 @@
-# app.py
 import tempfile
 import gradio as gr
 from main import load_artifacts_and_helpers, device
 
-# Load model + helpers once, using your original wrapper
+# Load model and helpers once when the Space starts
 final_pipeline, predict_fn, vocab, idx_to_answer, model, encode_fn = \
     load_artifacts_and_helpers(prefix="vqa_custom", map_location=device)
 
 def vqa_interface(image, question):
-    # image is a PIL image from Gradio, save to temp because your pipeline wants a path
+    # image is a PIL image from Gradio
     with tempfile.NamedTemporaryFile(suffix=".jpg") as f:
         image.save(f.name)
         answer = final_pipeline(
             f.name,
             question,
-            open_vqa_fn=None,    # we are NOT loading BLIP inside Space for now
+            open_vqa_fn=None,     # BLIP disabled in Space for speed
             translate_fn=None
         )
     return answer
@@ -27,7 +26,7 @@ demo = gr.Interface(
     ],
     outputs=gr.Textbox(label="Answer"),
     title="VQA-RAD Demo",
-    description="Custom VQA model (ResNet18 + LSTM) trained on VQA-RAD."
+    description="Custom ResNet18 + LSTM VQA model (top-50 answers)."
 )
 
 if __name__ == "__main__":