mask_mae

.vscode/launch.json

@@ -5,6 +5,14 @@
     "version": "0.2.0",
     
     "configurations": [
+        {
+            "name": "DDGCRN: METR-LA",
+            "type": "debugpy",
+            "request": "launch",
+            "program": "run.py",
+            "console": "integratedTerminal",
+            "args": "--config ./config/DDGCRN/METR-LA.yaml"
+        },
         // STID 模型组
         {
             "name": "STID: PEMS-BAY",

config/DDGCRN/METR-LA.yaml

@@ -0,0 +1,59 @@
+basic:
+  dataset: "METR-LA"
+  mode: "train"
+  device: "cuda:0"
+  model: "DDGCRN"
+  seed: 2023
+
+data:
+  add_day_in_week: true
+  add_time_in_day: true
+  column_wise: false
+  days_per_week: 7
+  default_graph: true
+  horizon: 24
+  lag: 24
+  normalizer: std
+  num_nodes: 207
+  steps_per_day: 288
+  test_ratio: 0.2
+  tod: false
+  val_ratio: 0.2
+  sample: 1
+  input_dim: 1
+  batch_size: 64
+
+model:
+  input_dim: 1
+  output_dim: 1
+  embed_dim: 12
+  rnn_units: 64
+  num_layers: 1
+  cheb_order: 2
+  use_day: True
+  use_week: False
+  num_nodes: 207
+  horizon: 24
+
+
+train:
+  loss_func: mask_mae
+  seed: 10
+  batch_size: 64
+  epochs: 100
+  lr_init: 0.003
+  weight_decay: 0
+  lr_decay: False
+  lr_decay_rate: 0.3
+  lr_decay_step: "5,20,40,70"
+  early_stop: True
+  early_stop_patience: 15
+  grad_norm: False
+  max_grad_norm: 5
+  debug: false
+  mae_thresh: 0.0
+  mape_thresh: 0.0
+  log_step: 10000
+  plot: False
+  output_dim: 1
+

trainer/Trainer.py

@@ -203,7 +203,7 @@ class Trainer:
                 self.stats.record_step_time(step_time, mode)
 
                 # 累积损失和预测结果
-                total_loss += loss.item()
+                total_loss += d_loss.item()
                 y_pred.append(d_output.detach().cpu())
                 y_true.append(d_label.detach().cpu())
 

utils/initializer.py

@@ -84,9 +84,10 @@ def init_device(args):
 def init_loss(args, scaler):
     device = args["basic"]["device"]
     args = args["train"]
+    mask_v = args['mae_thresh']
     match args["loss_func"]:
         case "mask_mae":
-            return masked_mae_loss(scaler, mask_value=None)
+            return masked_mae_loss(scaler, mask_value=mask_v)
         case "mae":
             return torch.nn.L1Loss().to(device)
         case "mse":

utils/loss_function.py

@@ -1,30 +1,16 @@
+import torch
+
+
 def masked_mae_loss(scaler, mask_value):
     def loss(preds, labels):
-        # 仅对预测反归一化；标签在数据管道中保持原始量纲
-        if scaler:
-            preds = scaler.inverse_transform(preds)
+        # # 仅对预测反归一化；标签在数据管道中保持原始量纲
+        # if scaler:
+        #     preds = scaler.inverse_transform(preds)
         return mae_torch(pred=preds, true=labels, mask_value=mask_value)
 
     return loss
 
 
-def get_loss_function(args, scaler):
-    if args["loss_func"] == "mask_mae":
-        # Return callable loss (no .to for function closures); disable masking by default
-        return masked_mae_loss(scaler, mask_value=None)
-    elif args["loss_func"] == "mae":
-        return torch.nn.L1Loss().to(args["device"])
-    elif args["loss_func"] == "mse":
-        return torch.nn.MSELoss().to(args["device"])
-    elif args["loss_func"] == "Huber":
-        return torch.nn.HuberLoss().to(args["device"])
-    else:
-        raise ValueError("Unsupported loss function: {}".format(args.loss_func))
-
-
-import torch
-
-
 def mae_torch(pred, true, mask_value=None):
     if mask_value is not None:
         mask = torch.gt(true, mask_value)