Update app.py

app.py

@@ -178,9 +178,100 @@ class TextClassifier:
 
     def detailed_scan(self, text: str) -> Dict:
         """Original prediction method with modified window handling"""
-        # Clean up trailing whitespace
-        text = text.rstrip()
-        
+        if self.model is None or self.tokenizer is None:
+            self.load_model()
+
+        self.model.eval()
+        sentences = self.processor.split_into_sentences(text)
+        if not sentences:
+            return {}
+
+        # Create centered windows for each sentence
+        windows, window_sentence_indices = self.processor.create_centered_windows(sentences, WINDOW_SIZE)
+
+        # Track scores for each sentence
+        sentence_appearances = {i: 0 for i in range(len(sentences))}
+        sentence_scores = {i: {'human_prob': 0.0, 'ai_prob': 0.0} for i in range(len(sentences))}
+
+        # Process windows in batches
+        batch_size = 16
+        for i in range(0, len(windows), batch_size):
+            batch_windows = windows[i:i + batch_size]
+            batch_indices = window_sentence_indices[i:i + batch_size]
+
+            inputs = self.tokenizer(
+                batch_windows,
+                truncation=True,
+                padding=True,
+                max_length=MAX_LENGTH,
+                return_tensors="pt"
+            ).to(self.device)
+
+            with torch.no_grad():
+                outputs = self.model(**inputs)
+                probs = F.softmax(outputs.logits, dim=-1)
+
+                # Attribute predictions more carefully
+                for window_idx, indices in enumerate(batch_indices):
+                    center_idx = len(indices) // 2
+                    center_weight = 0.7  # Higher weight for center sentence
+                    edge_weight = 0.3 / (len(indices) - 1)  # Distribute remaining weight
+
+                    for pos, sent_idx in enumerate(indices):
+                        # Apply higher weight to center sentence
+                        weight = center_weight if pos == center_idx else edge_weight
+                        sentence_appearances[sent_idx] += weight
+                        sentence_scores[sent_idx]['human_prob'] += weight * probs[window_idx][1].item()
+                        sentence_scores[sent_idx]['ai_prob'] += weight * probs[window_idx][0].item()
+
+            del inputs, outputs, probs
+            if torch.cuda.is_available():
+                torch.cuda.empty_cache()
+
+        # Calculate final predictions
+        sentence_predictions = []
+        for i in range(len(sentences)):
+            if sentence_appearances[i] > 0:
+                human_prob = sentence_scores[i]['human_prob'] / sentence_appearances[i]
+                ai_prob = sentence_scores[i]['ai_prob'] / sentence_appearances[i]
+
+                # Only apply minimal smoothing at prediction boundaries
+                if i > 0 and i < len(sentences) - 1:
+                    prev_human = sentence_scores[i-1]['human_prob'] / sentence_appearances[i-1]
+                    prev_ai = sentence_scores[i-1]['ai_prob'] / sentence_appearances[i-1]
+                    next_human = sentence_scores[i+1]['human_prob'] / sentence_appearances[i+1]
+                    next_ai = sentence_scores[i+1]['ai_prob'] / sentence_appearances[i+1]
+
+                    # Check if we're at a prediction boundary
+                    current_pred = 'human' if human_prob > ai_prob else 'ai'
+                    prev_pred = 'human' if prev_human > prev_ai else 'ai'
+                    next_pred = 'human' if next_human > next_ai else 'ai'
+
+                    if current_pred != prev_pred or current_pred != next_pred:
+                        # Small adjustment at boundaries
+                        smooth_factor = 0.1
+                        human_prob = (human_prob * (1 - smooth_factor) + 
+                                    (prev_human + next_human) * smooth_factor / 2)
+                        ai_prob = (ai_prob * (1 - smooth_factor) + 
+                                (prev_ai + next_ai) * smooth_factor / 2)
+
+                sentence_predictions.append({
+                    'sentence': sentences[i],
+                    'human_prob': human_prob,
+                    'ai_prob': ai_prob,
+                    'prediction': 'human' if human_prob > ai_prob else 'ai',
+                    'confidence': max(human_prob, ai_prob)
+                })
+
+        return {
+            'sentence_predictions': sentence_predictions,
+            'highlighted_text': self.format_predictions_html(sentence_predictions),
+            'full_text': text,
+            'overall_prediction': self.aggregate_predictions(sentence_predictions)
+        }
+
+    def detailed_scan(self, text: str) -> Dict:
+        """Perform a detailed scan with improved sentence-level analysis."""
         if not text.strip():
             return {
                 'sentence_predictions': [],
@@ -192,25 +283,23 @@ class TextClassifier:
                     'num_sentences': 0
                 }
             }
-    
-        self.model.eval()
+
         sentences = self.processor.split_into_sentences(text)
         if not sentences:
             return {}
-    
+
         # Create centered windows for each sentence
         windows, window_sentence_indices = self.processor.create_centered_windows(sentences, WINDOW_SIZE)
-    
+
         # Track scores for each sentence
         sentence_appearances = {i: 0 for i in range(len(sentences))}
         sentence_scores = {i: {'human_prob': 0.0, 'ai_prob': 0.0} for i in range(len(sentences))}
-    
+
         # Process windows in batches
-        batch_size = 16
-        for i in range(0, len(windows), batch_size):
-            batch_windows = windows[i:i + batch_size]
-            batch_indices = window_sentence_indices[i:i + batch_size]
-    
+        for i in range(0, len(windows), BATCH_SIZE):
+            batch_windows = windows[i:i + BATCH_SIZE]
+            batch_indices = window_sentence_indices[i:i + BATCH_SIZE]
+
             inputs = self.tokenizer(
                 batch_windows,
                 truncation=True,
@@ -218,48 +307,48 @@ class TextClassifier:
                 max_length=MAX_LENGTH,
                 return_tensors="pt"
             ).to(self.device)
-    
+
             with torch.no_grad():
                 outputs = self.model(**inputs)
                 probs = F.softmax(outputs.logits, dim=-1)
-    
+
                 # Attribute predictions with weighted scoring
                 for window_idx, indices in enumerate(batch_indices):
                     center_idx = len(indices) // 2
                     center_weight = 0.7  # Higher weight for center sentence
                     edge_weight = 0.3 / (len(indices) - 1)  # Distribute remaining weight
-    
+
                     for pos, sent_idx in enumerate(indices):
                         # Apply higher weight to center sentence
                         weight = center_weight if pos == center_idx else edge_weight
                         sentence_appearances[sent_idx] += weight
                         sentence_scores[sent_idx]['human_prob'] += weight * probs[window_idx][1].item()
                         sentence_scores[sent_idx]['ai_prob'] += weight * probs[window_idx][0].item()
-    
+
             # Clean up memory
             del inputs, outputs, probs
             if torch.cuda.is_available():
                 torch.cuda.empty_cache()
-    
+
         # Calculate final predictions with boundary smoothing
         sentence_predictions = []
         for i in range(len(sentences)):
             if sentence_appearances[i] > 0:
                 human_prob = sentence_scores[i]['human_prob'] / sentence_appearances[i]
                 ai_prob = sentence_scores[i]['ai_prob'] / sentence_appearances[i]
-    
-                # Only apply minimal smoothing at prediction boundaries
+
+                # Apply minimal smoothing at prediction boundaries
                 if i > 0 and i < len(sentences) - 1:
                     prev_human = sentence_scores[i-1]['human_prob'] / sentence_appearances[i-1]
                     prev_ai = sentence_scores[i-1]['ai_prob'] / sentence_appearances[i-1]
                     next_human = sentence_scores[i+1]['human_prob'] / sentence_appearances[i+1]
                     next_ai = sentence_scores[i+1]['ai_prob'] / sentence_appearances[i+1]
-    
+
                     # Check if we're at a prediction boundary
                     current_pred = 'human' if human_prob > ai_prob else 'ai'
                     prev_pred = 'human' if prev_human > prev_ai else 'ai'
                     next_pred = 'human' if next_human > next_ai else 'ai'
-    
+
                     if current_pred != prev_pred or current_pred != next_pred:
                         # Small adjustment at boundaries
                         smooth_factor = 0.1
@@ -267,7 +356,7 @@ class TextClassifier:
                                     (prev_human + next_human) * smooth_factor / 2)
                         ai_prob = (ai_prob * (1 - smooth_factor) + 
                                 (prev_ai + next_ai) * smooth_factor / 2)
-    
+
                 sentence_predictions.append({
                     'sentence': sentences[i],
                     'human_prob': human_prob,
@@ -275,7 +364,7 @@ class TextClassifier:
                     'prediction': 'human' if human_prob > ai_prob else 'ai',
                     'confidence': max(human_prob, ai_prob)
                 })
-    
+
         return {
             'sentence_predictions': sentence_predictions,
             'highlighted_text': self.format_predictions_html(sentence_predictions),
@@ -283,7 +372,6 @@ class TextClassifier:
             'overall_prediction': self.aggregate_predictions(sentence_predictions)
         }
 
-
     def format_predictions_html(self, sentence_predictions: List[Dict]) -> str:
         """Format predictions as HTML with color-coding."""
         html_parts = []