Update src/webApp.py

src/webApp.py

@@ -139,11 +139,19 @@ def get_word_importance(review, tokenizer, model, predicted_label):
         max_length=128,
         return_attention_mask=True
     )
+
     with torch.no_grad():
-        outputs = model(**inputs, output_attentions=True)
+        outputs = model(
+            **inputs,
+            output_attentions=True  # explicitly request attentions
+        )
+
+    # Check attentions were returned
+    if not outputs.attentions:
+        return []
 
-    tokens      = tokenizer.convert_ids_to_tokens(inputs["input_ids"][0])
-    attentions  = torch.stack(outputs.attentions)
+    tokens        = tokenizer.convert_ids_to_tokens(inputs["input_ids"][0])
+    attentions    = torch.stack(list(outputs.attentions))
     cls_attention = attentions.mean(dim=0)[0].mean(dim=0)[0]
 
     word_importance = []
@@ -157,35 +165,6 @@ def get_word_importance(review, tokenizer, model, predicted_label):
     return word_importance[:15]
 
 
-def word_importance_chart(word_importance, predicted_label):
-    if not word_importance:
-        return None
-
-    words, scores = zip(*word_importance)
-    colors        = ["#e63946" if s > 0 else "#2a9d8f" for s in scores]
-    sentiment_word = "Positive" if predicted_label == "LABEL_1" else "Negative"
-
-    fig = go.Figure(go.Bar(
-        x=list(scores),
-        y=list(words),
-        orientation="h",
-        marker_color=colors,
-        text=[f"{v:+.4f}" for v in scores],
-        textposition="outside"
-    ))
-    fig.update_layout(
-        title=f"Word Importance for {sentiment_word} Sentiment",
-        xaxis_title=f"Importance Score (positive = pushes toward {sentiment_word})",
-        height=450,
-        margin=dict(t=50, b=40, l=10, r=80),
-        plot_bgcolor="rgba(0,0,0,0)",
-        paper_bgcolor="rgba(0,0,0,0)",
-        font=dict(color="white"),
-        yaxis=dict(autorange="reversed")
-    )
-    return fig
-
-
 def shap_bar_chart(shap_values, feature_names):
     """
     Renders a horizontal bar chart of SHAP values for the suspicious review model.