Add GRU/LSTM/Mamba/Transformer to the model list.

COMMANDS.md

@@ -2,10 +2,22 @@
 
 ## 训练
 
+### WINDOWS+GPU
+
 ```bash
 python src/train.py experiment=PPG_FieldStudy_Base callbacks=PPG_FieldStudy logger=tensorboard
+
 python src/train.py experiment=PPG_FieldStudy_Base callbacks=PPG_FieldStudy logger=tensorboard trainer=gpu
-python src/train.py experiment=PPG_FieldStudy_Base logger=tensorboard callbacks=PPG_FieldStudy trainer=gpu debug=fdr 
+
+python src/train.py experiment=PPG_FieldStudy_Base logger=tensorboard callbacks=PPG_FieldStudy trainer=gpu debug=fdr
+```
+
+### WINDOWS+CPU
+
+```powershell
+python src/train.py experiment=PPG_FieldStudy_Base callbacks=PPG_FieldStudy logger=tensorboard trainer.accelerator=cpu
+
+python src/train.py experiment=PPG_FieldStudy_LSTM callbacks=PPG_FieldStudy logger=tensorboard trainer.accelerator=cpu
 ```
 
 ## 实验结果查看
@@ -21,5 +33,3 @@ tensorboard --logdir=./logs/train/runs/
 ```bash
 python src/eval.py ckpt_path="D:\lightning-hydra-template\logs\train\runs\2024-10-06_11-35-20\checkpoints\epoch_005.ckpt"
 ```
-
-

README.md

@@ -9,3 +9,9 @@ Related models:
 3. Transformer
 4. Mamba
 5. KAN
+
+## Datasets
+
+### PPG_FieldStudy
+
+https://www.kaggle.com/datasets/dishantvishwakarma/ppg-dataset-shared/data

configs/experiment/PPG_FieldStudy_GRU.yaml

@@ -0,0 +1,46 @@
+# @package _global_
+
+# to execute this experiment run:
+# python train.py experiment=example
+
+defaults:
+  - override /data: PPG_FieldStudy
+  - override /model: PPGGRU
+  - override /callbacks: default
+  - override /trainer: default
+
+# all parameters below will be merged with parameters from default configurations set above
+# this allows you to overwrite only specified parameters
+
+tags: ["PPG_FieldStudy", "PPGGRU"]
+
+seed: 12345
+
+trainer:
+  min_epochs: 10
+  max_epochs: 50
+  gradient_clip_val: 0.5
+  accelerator: "gpu"
+  devices: 1
+  callbacks: [progress_bar]
+  enable_checkpointing: True
+
+model:
+  input_dim: 14
+  hidden_dim: 128
+  num_layers: 2
+  bidirectional: True
+  dropout: 0.3
+  lr: 1e-3
+  scheduler_step_size: 10
+  scheduler_gamma: 0.1
+
+data:
+  batch_size: 64
+
+logger:
+  wandb:
+    tags: ${tags}
+    group: "PPG_FieldStudy"
+  aim:
+    experiment: "PPG_FieldStudy"

configs/experiment/PPG_FieldStudy_LSTM.yaml

@@ -0,0 +1,46 @@
+# @package _global_
+
+# to execute this experiment run:
+# python train.py experiment=example
+
+defaults:
+  - override /data: PPG_FieldStudy
+  - override /model: PPGLSTM
+  - override /callbacks: default
+  - override /trainer: default
+
+# all parameters below will be merged with parameters from default configurations set above
+# this allows you to overwrite only specified parameters
+
+tags: ["PPG_FieldStudy", "PPGLSTM"]
+
+seed: 12345
+
+trainer:
+  min_epochs: 10
+  max_epochs: 50
+  gradient_clip_val: 0.5
+  accelerator: "gpu"
+  devices: 1
+  callbacks: [progress_bar]
+  enable_checkpointing: True
+
+model:
+  input_dim: 14
+  hidden_dim: 128
+  num_layers: 2
+  bidirectional: True
+  dropout: 0.3
+  lr: 1e-3
+  scheduler_step_size: 10
+  scheduler_gamma: 0.1
+
+data:
+  batch_size: 64
+
+logger:
+  wandb:
+    tags: ${tags}
+    group: "PPG_FieldStudy"
+  aim:
+    experiment: "PPG_FieldStudy"

configs/experiment/PPG_FieldStudy_Mamba.yaml

@@ -0,0 +1,47 @@
+# @package _global_
+
+# to execute this experiment run:
+# python train.py experiment=example
+
+defaults:
+  - override /data: PPG_FieldStudy
+  - override /model: PPGMamba
+  - override /callbacks: default
+  - override /trainer: default
+
+# all parameters below will be merged with parameters from default configurations set above
+# this allows you to overwrite only specified parameters
+
+tags: ["PPG_FieldStudy", "PPGMamba"]
+
+seed: 12345
+
+trainer:
+  min_epochs: 10
+  max_epochs: 50
+  gradient_clip_val: 0.5
+  accelerator: "gpu"
+  devices: 1
+  callbacks: [progress_bar]
+  enable_checkpointing: True
+
+model:
+  input_dim: 14
+  d_model: 128
+  d_state: 64
+  d_conv: 4
+  expand: 2
+  dropout: 0.3
+  lr: 1e-3
+  scheduler_step_size: 10
+  scheduler_gamma: 0.1
+
+data:
+  batch_size: 64
+
+logger:
+  wandb:
+    tags: ${tags}
+    group: "PPG_FieldStudy"
+  aim:
+    experiment: "PPG_FieldStudy"

configs/experiment/PPG_FieldStudy_Transformer.yaml

@@ -0,0 +1,47 @@
+# @package _global_
+
+# to execute this experiment run:
+# python train.py experiment=example
+
+defaults:
+  - override /data: PPG_FieldStudy
+  - override /model: PPGTransformer
+  - override /callbacks: default
+  - override /trainer: default
+
+# all parameters below will be merged with parameters from default configurations set above
+# this allows you to overwrite only specified parameters
+
+tags: ["PPG_FieldStudy", "PPGTransformer"]
+
+seed: 12345
+
+trainer:
+  min_epochs: 10
+  max_epochs: 50
+  gradient_clip_val: 0.5
+  accelerator: "gpu"
+  devices: 1
+  callbacks: [progress_bar]
+  enable_checkpointing: True
+
+model:
+  input_dim: 14
+  d_model: 128
+  nhead: 8
+  num_encoder_layers: 4
+  dim_feedforward: 256
+  dropout: 0.1
+  lr: 1e-3
+  scheduler_step_size: 10
+  scheduler_gamma: 0.1
+
+data:
+  batch_size: 64
+
+logger:
+  wandb:
+    tags: ${tags}
+    group: "PPG_FieldStudy"
+  aim:
+    experiment: "PPG_FieldStudy"

configs/model/PPGGRU.yaml

@@ -0,0 +1,10 @@
+_target_: src.models.PPGGRUModule.PPGGRUModule
+
+input_dim: 14
+hidden_dim: 128
+num_layers: 2
+bidirectional: True
+dropout: 0.3
+lr: 1e-3
+scheduler_step_size: 10
+scheduler_gamma: 0.1

configs/model/PPGLSTM.yaml

@@ -0,0 +1,10 @@
+_target_: src.models.PPGLSTMModule.PPGLSTMModule
+
+input_dim: 14
+hidden_dim: 128
+num_layers: 2
+bidirectional: True
+dropout: 0.3
+lr: 1e-3
+scheduler_step_size: 10
+scheduler_gamma: 0.1

configs/model/PPGMamba.yaml

@@ -0,0 +1,11 @@
+_target_: src.models.PPGMambaModule.PPGMambaModule
+
+input_dim: 14
+d_model: 128
+d_state: 64
+d_conv: 4
+expand: 2
+dropout: 0.3
+lr: 1e-3
+scheduler_step_size: 10
+scheduler_gamma: 0.1

configs/model/PPGTransformer.yaml

@@ -0,0 +1,11 @@
+_target_: src.models.PPGTransformerModule.PPGTransformerModule
+
+input_dim: 14
+d_model: 128
+nhead: 8
+num_encoder_layers: 4
+dim_feedforward: 256
+dropout: 0.1
+lr: 1e-3
+scheduler_step_size: 10
+scheduler_gamma: 0.1

src/data/PPGFieldStudyDatamodule.py

@@ -44,8 +44,8 @@ def __init__(
 
         # Validate labels length
         assert (
-            len(labels) == self.num_windows
-        ), f"Number of labels ({len(labels)}) does not match number of windows ({self.num_windows})"
+            len(labels) >= self.num_windows
+        ), f"Number of labels ({len(labels)}) is less than number of windows ({self.num_windows})."
 
         # Normalize signals using StandardScaler
         self.scalers = {}
@@ -61,7 +61,7 @@ def __init__(
                     self.scalers[location][sensor] = scaler
 
         self.data = data
-        self.labels = labels
+        self.labels = labels[:self.num_windows]  # Ensure labels match the number of windows
 
     def __len__(self):
         return self.num_windows
@@ -93,21 +93,21 @@ def __getitem__(self, idx):
         # Concatenate all sensor data along the feature dimension
         chest_features = np.concatenate(
             [features["chest"][sensor] for sensor in features["chest"]], axis=1
-        )  # Shape: [256, 8]
+        )  # Shape: [window_size, 8]
         wrist_features = np.concatenate(
             [features["wrist"][sensor] for sensor in features["wrist"]], axis=1
-        )  # Shape: [256, 6]
+        )  # Shape: [window_size, 6]
         combined_features = np.concatenate(
             [chest_features, wrist_features], axis=1
-        )  # Shape: [256, 14]
+        )  # Shape: [window_size, 14]
 
         if self.transform:
             combined_features = self.transform(combined_features)
 
-        # Aggregate window into single vector by computing the mean across the window_size dimension
-        aggregated_features = combined_features.mean(axis=0)  # Shape: [14]
+        # 保留整个窗口的特征，不进行汇聚
+        aggregated_features = combined_features  # Shape: [window_size, 14]
 
-        # Ensure labels are aligned with windows
+        # 获取对应的标签
         label = self.labels[idx]
 
         return torch.tensor(aggregated_features, dtype=torch.float32), torch.tensor(