feat: show preprocessing vs inference timing in output

app.py

@@ -5,6 +5,7 @@ AD-Copilot Demo: Comparison-Aware Anomaly Detection with Vision-Language Model
 import json
 import os
 import re
+import time
 import traceback
 import spaces
 import gradio as gr
@@ -115,28 +116,26 @@ def predict(
         return "Please upload at least one image.", None
 
     try:
+        t_start = time.time()
         max_new_tokens = int(max_new_tokens)
 
         # Build message content based on available images
         content = []
 
         if has_ref and has_test:
-            # Paired comparison mode
             ref = reference_image.copy()
             tst = test_image.copy()
             ref.thumbnail((512, 512), Image.Resampling.LANCZOS)
             tst.thumbnail((512, 512), Image.Resampling.LANCZOS)
             content.append({"type": "image", "image": ref})
             content.append({"type": "image", "image": tst})
-            vis_source = tst  # visualize on test image
+            vis_source = tst
         elif has_test:
-            # Single image mode (only test image)
             tst = test_image.copy()
             tst.thumbnail((512, 512), Image.Resampling.LANCZOS)
             content.append({"type": "image", "image": tst})
             vis_source = tst
         else:
-            # Single image mode (only reference image)
             ref = reference_image.copy()
             ref.thumbnail((512, 512), Image.Resampling.LANCZOS)
             content.append({"type": "image", "image": ref})
@@ -158,9 +157,14 @@ def predict(
             return_tensors="pt",
         ).to(model.device)
 
+        t_preprocess = time.time()
+
         generated_ids = model.generate(
             **inputs, max_new_tokens=max_new_tokens, do_sample=False
         )
+
+        t_generate = time.time()
+
         generated_ids_trimmed = [
             out[len(inp) :] for inp, out in zip(inputs.input_ids, generated_ids)
         ]
@@ -176,6 +180,11 @@ def predict(
         if bboxes:
             vis_image = draw_bboxes(vis_source, bboxes)
 
+        # Append timing info
+        prep_time = t_preprocess - t_start
+        gen_time = t_generate - t_preprocess
+        output += f"\n\n---\nPreprocessing: {prep_time:.1f}s | Inference: {gen_time:.1f}s"
+
         return output, vis_image
     except Exception as e:
         tb = traceback.format_exc()