config/iTransformer/METR-LA.yaml

@@ -42,11 +42,11 @@ train:
   lr_decay: true
   lr_decay_rate: 0.3
   lr_decay_step: 5,20,40,70
-  lr_init: 0.003
+  lr_init: 0.0001
   mae_thresh: None
   mape_thresh: 0.001
   max_grad_norm: 5
-  output_dim: 1
+  output_dim: 207
   plot: false
   real_value: true
   weight_decay: 0

dataloader/TSloader.py

@@ -32,58 +32,24 @@ def get_dataloader(args, normalizer="std", single=True):
         scaler
     )
 
-
 def _prepare_data_with_windows(data, args, single):
     # Generate sliding windows for main data
     x = add_window_x(data, args["lag"], args["horizon"], single)
     y = add_window_y(data, args["lag"], args["horizon"], single)
-
-    # Generate time features
-    time_features = _generate_time_features(data.shape[0], args)
-    
-    # Add time features to x and y
-    x = _add_time_features(x, time_features, args["lag"], args["horizon"], single, add_window_x)
-    y = _add_time_features(y, time_features, args["lag"], args["horizon"], single, add_window_y)
-    
     return x, y
 
-
-def _generate_time_features(L, args):
-    # For time series data, we generate time features for each time step
-    # [L, 1] -> [L, T, 1] by repeating across time dimension
-    T = args.get("time_dim", 1)  # Get time dimension size if available
-    
-    time_in_day = [i % args["steps_per_day"] / args["steps_per_day"] for i in range(L)]
-    time_in_day = np.array(time_in_day)[:, None, None]  # [L, 1, 1]
-    time_in_day = np.tile(time_in_day, (1, T, 1))  # [L, T, 1]
-    
-    day_in_week = [(i // args["steps_per_day"]) % args["days_per_week"] for i in range(L)]
-    day_in_week = np.array(day_in_week)[:, None, None]  # [L, 1, 1]
-    day_in_week = np.tile(day_in_week, (1, T, 1))  # [L, T, 1]
-    
-    return time_in_day, day_in_week
-
-
-
-def _add_time_features(data, time_features, lag, horizon, single, window_fn):
-    time_in_day, day_in_week = time_features
-    time_day = window_fn(time_in_day, lag, horizon, single)
-    time_week = window_fn(day_in_week, lag, horizon, single)
-    return np.concatenate([data, time_day, time_week], axis=-1)
-
-
 def _normalize_data(train_data, val_data, test_data, args, normalizer):
-    scaler = normalize_dataset(train_data[..., : args["input_dim"]], normalizer, args["column_wise"])
+    scaler = normalize_dataset(train_data[..., : args["num_nodes"]], normalizer, args["column_wise"])
     
     for data in [train_data, val_data, test_data]:
-        data[..., : args["input_dim"]] = scaler.transform(data[..., : args["input_dim"]])
+        data[..., : args["num_nodes"]] = scaler.transform(data[..., : args["num_nodes"]])
     
     return scaler
 
 
 def _apply_existing_scaler(train_data, val_data, test_data, scaler, args):
     for data in [train_data, val_data, test_data]:
-        data[..., : args["input_dim"]] = scaler.transform(data[..., : args["input_dim"]])
+        data[..., : args["num_nodes"]] = scaler.transform(data[..., : args["num_nodes"]])
 
 
 def _create_dataloader(X_data, Y_data, batch_size, shuffle, drop_last):
@@ -105,15 +71,10 @@ def split_data_by_days(data, val_days, test_days, interval=30):
 def split_data_by_ratio(data, val_ratio, test_ratio):
     data_len = data.shape[0]
     test_data = data[-int(data_len * test_ratio) :]
-    val_data = data[
+    val_data = data[-int(data_len * (test_ratio + val_ratio)) : -int(data_len * test_ratio)]
-        -int(data_len * (test_ratio + val_ratio)) : -int(data_len * test_ratio)
-    ]
     train_data = data[: -int(data_len * (test_ratio + val_ratio))]
     return train_data, val_data, test_data
 
-
-
-
 def _generate_windows(data, window=3, horizon=1, offset=0):
     """
     Internal helper function to generate sliding windows.

model/iTransformer/iTransformer.py

@@ -38,6 +38,6 @@ class iTransformer(nn.Module):
         dec_out = self.projector(enc_out).permute(0, 2, 1)[:, :, :N]  # filter the covariates
         return dec_out, attns
 
-    def forward(self, x_enc, x_mark_enc):
+    def forward(self, x_enc, x_mark_enc=None):
         dec_out, attns = self.forecast(x_enc, x_mark_enc)
         return dec_out[:, -self.pred_len:, :]  # [B, T, C]