Update inference.py

inference.py

@@ -62,15 +62,14 @@ def load_model_and_scaler(args):
 def predict_future(args, model, scaler, device):
     df_input = pd.read_csv(args.predict_input_file)
     df_input['date'] = pd.to_datetime(df_input['date'])
-
     cols_to_scale = ['air_pres', 'wind_dir', 'wind_speed', 'air_temp', 'residual']
-    
     raw_input = df_input[cols_to_scale].tail(args.seq_len).values
     input_scaled = scaler.transform(raw_input)
     batch_x = torch.from_numpy(input_scaled).float().unsqueeze(0).to(device)
     
     df_stamp_enc = df_input.tail(args.seq_len)[['date']].reset_index(drop=True)
-    enc_mark = time_features(df_stamp_enc, timeenc=0, freq=args.freq)
+    # ⭐️ 1번 문제 해결: 불필요한 timeenc=0 인자 제거
+    enc_mark = time_features(df_stamp_enc, freq=args.freq)
     batch_x_mark = torch.from_numpy(enc_mark).float().unsqueeze(0).to(device)
     
     dec_inp_label = input_scaled[-args.label_len:]
@@ -81,7 +80,8 @@ def predict_future(args, model, scaler, device):
     last_date = df_stamp_enc['date'].iloc[-1]
     future_dates = pd.date_range(start=last_date, periods=args.pred_len + 1, freq='5T')[1:]
     df_stamp_dec = pd.DataFrame({'date': list(df_stamp_enc['date'].values[-args.label_len:]) + list(future_dates)})
-    dec_mark = time_features(df_stamp_dec, timeenc=0, freq=args.freq)
+    # ⭐️ 1번 문제 해결: 불필요한 timeenc=0 인자 제거
+    dec_mark = time_features(df_stamp_dec, freq=args.freq)
     batch_y_mark = torch.from_numpy(dec_mark).float().unsqueeze(0).to(device)
 
     with torch.no_grad():
@@ -89,7 +89,6 @@ def predict_future(args, model, scaler, device):
         
     prediction_scaled = outputs.detach().cpu().numpy()[0]
     
-    # ⭐️ 이 블록의 들여쓰기가 문제였습니다.
     if scaler.n_features_in_ > 1:
         padding = np.zeros((prediction_scaled.shape[0], scaler.n_features_in_ - args.c_out))
         prediction_padded = np.concatenate((padding, prediction_scaled), axis=1)
@@ -107,56 +106,47 @@ def evaluate_performance(args, model, scaler, device):
     raw_data = df_eval[cols_to_scale].values
     data_scaled = scaler.transform(raw_data)
     
-    df_stamp = time_features(df_eval[['date']], timeenc=0, freq=args.freq)
+    # ⭐️ 1번 문제 해결: 불필요한 timeenc=0 인자 제거
+    df_stamp = time_features(df_eval[['date']], freq=args.freq)
 
     preds_unscaled = []
     trues_unscaled = []
 
     num_samples = len(data_scaled) - args.seq_len - args.pred_len + 1
     for i in tqdm(range(num_samples), desc="Evaluating", file=sys.stderr):
+        # ... (이하 로직은 이전과 동일, 수정 없음) ...
         s_begin = i
         s_end = s_begin + args.seq_len
-        
         batch_x = data_scaled[s_begin:s_end]
         batch_x_mark = df_stamp[s_begin:s_end]
-        
         true_begin = s_end
         true_end = true_begin + args.pred_len
         true_scaled = data_scaled[true_begin:true_end]
-        
         dec_inp_label = batch_x[-args.label_len:]
         dec_inp_pred = np.zeros([args.pred_len, args.enc_in])
         batch_y = np.concatenate([dec_inp_label, dec_inp_pred], axis=0)
-        
         dec_mark_label = df_stamp[s_end-args.label_len:s_end]
         dec_mark_pred = df_stamp[true_begin:true_end]
         batch_y_mark = np.concatenate([dec_mark_label, dec_mark_pred], axis=0)
-
         batch_x = torch.from_numpy(batch_x).float().unsqueeze(0).to(device)
         batch_x_mark = torch.from_numpy(batch_x_mark).float().unsqueeze(0).to(device)
         batch_y = torch.from_numpy(batch_y).float().unsqueeze(0).to(device)
         batch_y_mark = torch.from_numpy(batch_y_mark).float().unsqueeze(0).to(device)
-        
         with torch.no_grad():
             outputs = model(batch_x, batch_x_mark, batch_y, batch_y_mark)
-        
         pred_scaled = outputs.detach().cpu().numpy()[0]
-        
         if scaler.n_features_in_ > 1:
             padding = np.zeros((pred_scaled.shape[0], scaler.n_features_in_ - args.c_out))
             pred_padded = np.concatenate((padding, pred_scaled), axis=1)
             pred_unscaled = scaler.inverse_transform(pred_padded)[:, -args.c_out:]
         else:
             pred_unscaled = scaler.inverse_transform(pred_scaled)
-        
         true_unscaled = scaler.inverse_transform(true_scaled)[:, -args.c_out:]
-
         preds_unscaled.append(pred_unscaled)
         trues_unscaled.append(true_unscaled)
 
     return np.array(preds_unscaled), np.array(trues_unscaled)
 
-
 if __name__ == '__main__':
     
     final_output = {}