.vscode/launch.json

@@ -2107,12 +2107,12 @@
             "args": "--config ./config/iTransformer/AirQuality.yaml"
         },
         {
-            "name": "iTransformer: SolarEnergy",
+            "name": "HI: PEMS-BAY",
             "type": "debugpy",
             "request": "launch",
             "program": "run.py",
             "console": "integratedTerminal",
-            "args": "--config ./config/iTransformer/SolarEnergy.yaml"
+            "args": "--config ./config/HI/PEMS-BAY.yaml"
         },
     ]
 }

config/HI/AirQuality.yaml

@@ -0,0 +1,48 @@
+basic:
+  dataset: AirQuality
+  device: cuda:0
+  mode: train
+  model: HI
+  seed: 2023
+
+data:
+  batch_size: 16
+  column_wise: false
+  days_per_week: 7
+  horizon: 24
+  input_dim: 6
+  lag: 24
+  normalizer: std
+  num_nodes: 35
+  steps_per_day: 24
+  test_ratio: 0.2
+  val_ratio: 0.2
+
+model:
+  input_len: 24
+  output_len: 24
+  reverse: False
+
+
+train:
+  batch_size: 16
+  debug: false
+  early_stop: true
+  early_stop_patience: 15
+  epochs: 1
+  grad_norm: false
+  log_step: 1000
+  loss_func: mae
+  lr_decay: true
+  lr_decay_rate: 0.3
+  lr_decay_step: 5,20,40,70
+  lr_init: 0.0001
+  mae_thresh: None
+  mape_thresh: 0.001
+  max_grad_norm: 5
+  output_dim: 35
+  optimizer: null
+  plot: false
+  real_value: true
+  scheduler: null
+  weight_decay: 0

config/HI/BJTaxi-Inflow.yaml

@@ -0,0 +1,48 @@
+basic:
+  dataset: BJTaxi-InFlow
+  device: cuda:0
+  mode: train
+  model: HI
+  seed: 2023
+
+data:
+  batch_size: 32
+  column_wise: false
+  days_per_week: 7
+  horizon: 24
+  input_dim: 1
+  lag: 24
+  normalizer: std
+  num_nodes: 1024
+  steps_per_day: 48
+  test_ratio: 0.2
+  val_ratio: 0.2
+
+model:
+  input_len: 24
+  output_len: 24
+  reverse: False
+
+
+train:
+  batch_size: 16
+  debug: false
+  early_stop: true
+  early_stop_patience: 15
+  epochs: 1
+  grad_norm: false
+  log_step: 1000
+  loss_func: mae
+  lr_decay: true
+  lr_decay_rate: 0.3
+  lr_decay_step: 5,20,40,70
+  lr_init: 0.0001
+  mae_thresh: None
+  mape_thresh: 0.001
+  max_grad_norm: 5
+  output_dim: 1024
+  optimizer: null
+  plot: false
+  real_value: true
+  scheduler: null
+  weight_decay: 0

config/HI/BJTaxi-Outflow.yaml

@@ -0,0 +1,48 @@
+basic:
+  dataset: BJTaxi-OutFlow
+  device: cuda:0
+  mode: train
+  model: HI
+  seed: 2023
+
+data:
+  batch_size: 32
+  column_wise: false
+  days_per_week: 7
+  horizon: 24
+  input_dim: 1
+  lag: 24
+  normalizer: std
+  num_nodes: 1024
+  steps_per_day: 48
+  test_ratio: 0.2
+  val_ratio: 0.2
+
+model:
+  input_len: 24
+  output_len: 24
+  reverse: False
+
+
+train:
+  batch_size: 16
+  debug: false
+  early_stop: true
+  early_stop_patience: 15
+  epochs: 1
+  grad_norm: false
+  log_step: 1000
+  loss_func: mae
+  lr_decay: true
+  lr_decay_rate: 0.3
+  lr_decay_step: 5,20,40,70
+  lr_init: 0.0001
+  mae_thresh: None
+  mape_thresh: 0.001
+  max_grad_norm: 5
+  output_dim: 1024
+  optimizer: null
+  plot: false
+  real_value: true
+  scheduler: null
+  weight_decay: 0

config/HI/METR-LA.yaml

@@ -0,0 +1,48 @@
+basic:
+  dataset: METR-LA
+  device: cuda:1
+  mode: train
+  model: HI
+  seed: 2023
+
+data:
+  batch_size: 16
+  column_wise: false
+  days_per_week: 7
+  horizon: 24
+  input_dim: 1
+  lag: 24
+  normalizer: std
+  num_nodes: 207
+  steps_per_day: 288
+  test_ratio: 0.2
+  val_ratio: 0.2
+
+model:
+  input_len: 24
+  output_len: 24
+  reverse: False
+
+
+train:
+  batch_size: 16
+  debug: false
+  early_stop: true
+  early_stop_patience: 15
+  epochs: 1
+  grad_norm: false
+  log_step: 1000
+  loss_func: mae
+  lr_decay: true
+  lr_decay_rate: 0.3
+  lr_decay_step: 5,20,40,70
+  lr_init: 0.0001
+  mae_thresh: None
+  mape_thresh: 0.001
+  max_grad_norm: 5
+  output_dim: 207
+  optimizer: null
+  plot: false
+  real_value: true
+  scheduler: null
+  weight_decay: 0

config/HI/NYCBike-Inflow.yaml

@@ -0,0 +1,48 @@
+basic:
+  dataset: NYCBike-InFlow
+  device: cuda:0
+  mode: train
+  model: HI
+  seed: 2023
+
+data:
+  batch_size: 32
+  column_wise: false
+  days_per_week: 7
+  horizon: 24
+  input_dim: 1
+  lag: 24
+  normalizer: std
+  num_nodes: 128
+  steps_per_day: 48
+  test_ratio: 0.2
+  val_ratio: 0.2
+
+model:
+  input_len: 24
+  output_len: 24
+  reverse: False
+
+
+train:
+  batch_size: 16
+  debug: false
+  early_stop: true
+  early_stop_patience: 15
+  epochs: 1
+  grad_norm: false
+  log_step: 1000
+  loss_func: mae
+  lr_decay: true
+  lr_decay_rate: 0.3
+  lr_decay_step: 5,20,40,70
+  lr_init: 0.0001
+  mae_thresh: None
+  mape_thresh: 0.001
+  max_grad_norm: 5
+  output_dim: 128
+  optimizer: null
+  plot: false
+  real_value: true
+  scheduler: null
+  weight_decay: 0

config/HI/NYCBike-Outflow.yaml

@@ -0,0 +1,48 @@
+basic:
+  dataset: NYCBike-OutFlow
+  device: cuda:0
+  mode: train
+  model: HI
+  seed: 2023
+
+data:
+  batch_size: 32
+  column_wise: false
+  days_per_week: 7
+  horizon: 24
+  input_dim: 1
+  lag: 24
+  normalizer: std
+  num_nodes: 128
+  steps_per_day: 48
+  test_ratio: 0.2
+  val_ratio: 0.2
+
+model:
+  input_len: 24
+  output_len: 24
+  reverse: False
+
+
+train:
+  batch_size: 16
+  debug: false
+  early_stop: true
+  early_stop_patience: 15
+  epochs: 1
+  grad_norm: false
+  log_step: 1000
+  loss_func: mae
+  lr_decay: true
+  lr_decay_rate: 0.3
+  lr_decay_step: 5,20,40,70
+  lr_init: 0.0001
+  mae_thresh: None
+  mape_thresh: 0.001
+  max_grad_norm: 5
+  output_dim: 128
+  optimizer: null
+  plot: false
+  real_value: true
+  scheduler: null
+  weight_decay: 0

config/HI/PEMS-BAY.yaml

@@ -0,0 +1,48 @@
+basic:
+  dataset: PEMS-BAY
+  device: cuda:0
+  mode: train
+  model: HI
+  seed: 2023
+
+data:
+  batch_size: 16
+  column_wise: false
+  days_per_week: 7
+  horizon: 24
+  input_dim: 1
+  lag: 24
+  normalizer: std
+  num_nodes: 325
+  steps_per_day: 288
+  test_ratio: 0.2
+  val_ratio: 0.2
+
+model:
+  input_len: 24
+  output_len: 24
+  reverse: False
+
+
+train:
+  batch_size: 16
+  debug: false
+  early_stop: true
+  early_stop_patience: 15
+  epochs: 1
+  grad_norm: false
+  log_step: 1000
+  loss_func: mae
+  lr_decay: true
+  lr_decay_rate: 0.3
+  lr_decay_step: 5,20,40,70
+  lr_init: 0.0001
+  mae_thresh: None
+  mape_thresh: 0.001
+  max_grad_norm: 5
+  output_dim: 325
+  optimizer: null
+  plot: false
+  real_value: true
+  scheduler: null
+  weight_decay: 0

config/HI/SolarEnergy.yaml

@@ -0,0 +1,48 @@
+basic:
+  dataset: SolarEnergy
+  device: cuda:0
+  mode: train
+  model: HI
+  seed: 2023
+
+data:
+  batch_size: 16
+  column_wise: false
+  days_per_week: 7
+  horizon: 24
+  input_dim: 6
+  lag: 24
+  normalizer: std
+  num_nodes: 137
+  steps_per_day: 24
+  test_ratio: 0.2
+  val_ratio: 0.2
+
+model:
+  input_len: 24
+  output_len: 24
+  reverse: False
+
+
+train:
+  batch_size: 16
+  debug: false
+  early_stop: true
+  early_stop_patience: 15
+  epochs: 1
+  grad_norm: false
+  log_step: 1000
+  loss_func: mae
+  lr_decay: 
+  lr_decay_rate: 
+  lr_decay_step: 
+  lr_init: 0.0001
+  mae_thresh: None
+  mape_thresh: 0.001
+  max_grad_norm: 5
+  output_dim: 137
+  optimizer: null
+  plot: false
+  real_value: true
+  scheduler: null
+  weight_decay: 0

dataloader/loader_selector.py

@@ -7,7 +7,11 @@ from dataloader.TSloader import get_dataloader as TS_loader
 
 
 def get_dataloader(config, normalizer, single):
+    TS_model = ["iTransformer", "HI"]
     model_name = config["basic"]["model"]
+    if model_name in TS_model:
+        return TS_loader(config, normalizer, single)
+    else :
         match model_name:
             case "STGNCDE":
                 return cde_loader(config, normalizer, single)
@@ -17,7 +21,5 @@ def get_dataloader(config, normalizer, single):
                 return DCRNN_loader(config, normalizer, single)
             case "EXP":
                 return EXP_loader(config, normalizer, single)
-        case "iTransformer":
-            return TS_loader(config, normalizer, single)
             case _:
                 return normal_loader(config, normalizer, single)

model/HI/HI.py

@@ -0,0 +1,45 @@
+from typing import List
+import torch
+from torch import nn
+
+
+class HI(nn.Module):
+    """
+    Paper: Historical Inertia: A Neglected but Powerful Baseline for Long Sequence Time-series Forecasting
+    Link: https://arxiv.org/abs/2103.16349
+    Official code: None
+    Venue: CIKM 2021
+    Task: Long-term Time Series Forecasting
+    """
+
+    def __init__(self, config):
+        """
+        Init HI.
+
+        Args:
+            config (HIConfig): model config.
+        """
+
+        super().__init__()
+        self.input_len    = config['input_len']
+        self.output_len   = config['output_len']
+        assert self.input_len >= self.output_len, "HI model requires input length > output length"
+        self.reverse         = config['reverse']
+        # self.fake_param      = nn.Linear(1, 1, bias=False) 
+
+    def forward(self, inputs: torch.Tensor) -> torch.Tensor:
+        """Forward function of HI.
+
+        Args:
+            inputs (torch.Tensor): shape = [B, L_in, N]
+
+        Returns:
+            torch.Tensor: model prediction [B, L_out, N].
+        """
+        # historical inertia
+        prediction = inputs[:, -self.output_len:, :]
+        # last point
+        # prediction = inputs[:, [-1], :].expand(-1, self.output_len, -1)
+        if self.reverse:
+            prediction = prediction.flip(dims=[1])
+        return prediction

model/model_selector.py

@@ -1,4 +1,5 @@
 from model.DDGCRN.DDGCRN import DDGCRN
+from model.HI import HI
 from model.TWDGCN.TWDGCN import TWDGCN
 from model.AGCRN.AGCRN import AGCRN
 from model.NLT.HierAttnLstm import HierAttnLstm
@@ -27,6 +28,7 @@ from model.ASTRA.astra import ASTRA as ASTRA
 from model.ASTRA.astrav2 import ASTRA as ASTRAv2
 from model.ASTRA.astrav3 import ASTRA as ASTRAv3
 from model.iTransformer.iTransformer import iTransformer
+from model.HI.HI import HI
 
 
 
@@ -92,3 +94,5 @@ def model_selector(config):
             return ASTRAv3(model_config)
         case "iTransformer":
             return iTransformer(model_config)
+        case "HI":
+            return HI(model_config)

utils/initializer.py

@@ -23,6 +23,9 @@ def init_model(args):
 
 
 def init_optimizer(model, args):
+    optimizer = None
+    lr_scheduler = None
+    
     optim = args.get("optimizer", "Adam")
     match optim :
         case "Adam":