Update app.py

app.py

@@ -163,16 +163,50 @@ def save_animation(anim, format="mp4"):
         print(f"{format.upper()} save failed:", e)
         return None
 
-# === Inference with fallback ===
+def find_all_matches(text, video_lookup):
+    words = text.strip().lower().split()
+    matched = []
+    i = 0
+    while i < len(words):
+        found = False
+        # Try trigram
+        if i + 2 < len(words):
+            phrase3 = " ".join(words[i:i+3])
+            if phrase3 in video_lookup:
+                matched.append((phrase3, video_lookup[phrase3]))
+                i += 3
+                found = True
+                continue
+        # Try bigram
+        if i + 1 < len(words):
+            phrase2 = " ".join(words[i:i+2])
+            if phrase2 in video_lookup:
+                matched.append((phrase2, video_lookup[phrase2]))
+                i += 2
+                found = True
+                continue
+        # Try unigram
+        word = words[i]
+        if word in video_lookup:
+            matched.append((word, video_lookup[word]))
+        else:
+            matched.append((word, None))  # not in lookup
+        i += 1
+    return matched
+
+
 def predict(text, threshold):
     if not text.strip():
         return None, "⚠️ Please enter valid text."
 
     try:
-        pose, conf, labels = concatenate_and_smooth_sequences(text, tokenizer, model, device, GLOBAL_MEAN_T, GLOBAL_STD_T)
-        if pose is None: 
+        pose, conf, labels = concatenate_and_smooth_sequences(
+            text, tokenizer, model, device, GLOBAL_MEAN_T, GLOBAL_STD_T
+        )
+
+        if pose is None:
             return None, "⚠️ No pose predicted."
-            
+
         anim = animate_pose(pose, pred_conf=conf, frame_labels=labels, conf_threshold=threshold)
         path = save_animation(anim, format="mp4")
         if not path:
@@ -184,31 +218,31 @@ def predict(text, threshold):
             msg = ""
 
         cleaned_text = text.strip().lower()
-        checks = []
+        matches = find_all_matches(cleaned_text, video_lookup)
+        checks = ["**Match Check (Phrase + Word Level):**"]
 
-        if cleaned_text in annotated_words:
-            checks.append("**Phrase Check:**")
-            checks.append(f"- “{cleaned_text}” ✅ in dataset")
-        else:
-            checks.append("**Word-by-Word Check:**")
-            for w in cleaned_text.split():
-                if w in annotated_words:
-                    checks.append(f"- “{w}” ✅ in dataset")
+        for phrase, url in matches:
+            # Dataset presence
+            if phrase in annotated_words:
+                line = f'- “{phrase}” ✅ in dataset'
+            else:
+                line = f'- “{phrase}” ⚠️ not in dataset — model will approximate'
+
+            # YouTube thumbnail
+            if url:
+                if "v=" in url:
+                    video_id = url.split("v=")[-1].split("&")[0]
+                    thumbnail_url = f"https://img.youtube.com/vi/{video_id}/0.jpg"
+                    line += f'\n  📺 [![Watch Video]({thumbnail_url})]({url})'
                 else:
-                    checks.append(f"- “{w}” ⚠️ not in dataset — will be generalized")
-
-        # === YouTube link lookup ===
-        video_url = get_youtube_link(cleaned_text)
-        if video_url:
-            video_id = video_url.split("v=")[-1]
-            thumbnail_url = f"https://img.youtube.com/vi/{video_id}/0.jpg"
-            checks.append(f"\n📺 [![Watch Reference Video]({thumbnail_url})]({video_url})")
-        else:
-            checks.append(f"\n⚠️ No YouTube video found for “{cleaned_text}” — the predicted pose is based on the model’s learned approximation.")
+                    line += f'\n  🔗 [Reference Video]({url})'
+            else:
+                line += f'\n  ⚠️ No YouTube video found for “{phrase}”'
 
+            checks.append(line)
 
         return path, msg + "\n" + "\n".join(checks)
-        
+
     except Exception as e:
         print("Error during prediction:", e)
         return None, f"❌ Runtime error: {str(e)}"