lime

app.py

@@ -13,7 +13,7 @@ from pycaret.classification import *
 import warnings
 warnings.filterwarnings("ignore", category=FutureWarning, module="torch.storage")
 
-from util import load_data
+from util import load_data_and_prepare
 import view
 
 
@@ -33,17 +33,6 @@ def parse_args(args):
     return parser.parse_args(args)
 
 
-def load_data_and_prepare(data_dir, excel_file, mode, scale, smote):
-    # Load train, validation, and test data
-    train_df,val_df = load_data(data_dir, excel_file, mode, scale, smote)
-    
-    train_df.drop(columns=['patient_id','target'],inplace = True)
-    val_df.drop(columns=['patient_id','target'],inplace = True)
-    
-    train = pd.concat([train_df,val_df],axis=0)
-    
-    return train
-
 
 # Inference function
 def classify(tabular_data):
@@ -82,6 +71,64 @@ def classify(tabular_data):
     except Exception as e:
         return f"An error occurred during classification: {str(e)}"
 
+# Inference function
+def predict_proba_fn(instance):
+    """
+    PyCaret의 predict_model을 활용한 확률 예측 함수.
+    """
+    # 2D 형태로 변환
+    if instance.ndim == 1:
+        instance = instance.reshape(1, -1)
+    
+    # DataFrame으로 변환
+    instance_df = pd.DataFrame(instance, columns=train.columns)
+    
+    # predict_model을 통해 예측 수행
+    predictions = predict_model(model, data=instance_df)
+    
+    # prediction_label이 1이면 prediction_score, 0이면 1-prediction_score
+    predictions['class_1_prob'] = np.where(predictions['prediction_label'] == 1, 
+                                            predictions['prediction_score'], 
+                                            0)
+    
+    predictions['class_0_prob'] = np.where(predictions['prediction_label'] == 0, 
+                                            predictions['prediction_score'], 
+                                            0)
+    
+    # class_0_prob와 class_1_prob 반환
+    return predictions[['class_0_prob', 'class_1_prob']].values
+
+
+def explain_with_lime(tabular_data):
+    instance = np.array(tabular_data[0],dtype='float')
+    
+    # Create an explainer instance for classification
+    explainer = LimeTabularExplainer(
+        training_data=train.values,  # Use your training data
+        feature_names=tabular_header,
+        class_names=['intermediate', 'High'],  # Replace with actual class names
+        mode='classification'
+    )
+
+    # LIME expects a 2D numpy array or DataFrame for input, and we need to provide the correct number of features
+    explanation = explainer.explain_instance(
+        instance,  # Single instance (first row of the tabular data)
+        predict_proba_fn,  # The prediction function
+        num_features=len(tabular_header)  # Number of features to display in the explanation
+    )
+    
+    # Plot LIME explanation
+    fig = explanation.as_pyplot_figure()
+    fig.set_size_inches(25, 8) 
+    buf = io.BytesIO()
+    fig.savefig(buf, format='png')
+    buf.seek(0)
+    encoded_image = base64.b64encode(buf.read()).decode('utf-8')
+    buf.close()
+    plt.close(fig)
+    
+    return f"<img src='data:image/png;base64,{encoded_image}'/>"
+
 if __name__ == '__main__':
     args = parse_args(sys.argv[1:])
     train = load_data_and_prepare(args.data_dir, args.excel_file, args.mode, args.scale, args.smote)
@@ -119,7 +166,7 @@ if __name__ == '__main__':
         lime_output = gr.HTML(label="LIME Explanation")
         gr.Examples(examples=examples, inputs=[tabular_input, info])
         btn.click(fn=classify, inputs=tabular_input, outputs=result_output)
-        # btn.click(fn=explain_with_lime, inputs=tabular_input, outputs=lime_output)  # Add LIME button    
+        btn.click(fn=explain_with_lime, inputs=tabular_input, outputs=lime_output)  # Add LIME button    
     
         # Clear functionality: resets inputs and outputs
         def clear_fields():