feat: default disabling early-stopping mechanism during model training;

pypots/base.py

@@ -203,6 +203,7 @@ class BaseNNModel(BaseModel):
  patience : int,
  Number of epochs the training procedure will keep if loss doesn't decrease.
  Once exceeding the number, the training will stop.
+ Must be smaller than or equal to the value of `epoches`.
 
  learning_rate : float,
  The learning rate of the optimizer.
@@ -252,6 +253,13 @@ def __init__(
  ):
  super().__init__(device, tb_file_saving_path)
 
+ if patience is None:
+ patience = -1 # early stopping on patience won't work if it is set as < 0
+ else:
+ assert (
+ patience <= epochs
+ ), f"patience must be smaller than epoches which is {epochs}, but got patience={patience}"
+
  # training hype-parameters
  self.batch_size = batch_size
  self.epochs = epochs

pypots/classification/brits.py

@@ -153,7 +153,7 @@ def __init__(
  reconstruction_weight: float = 1,
  batch_size: int = 32,
  epochs: int = 100,
- patience: int = 10,
+ patience: int = None,
  learning_rate: float = 1e-3,
  weight_decay: float = 1e-5,
  num_workers: int = 0,

pypots/classification/grud.py

@@ -135,7 +135,7 @@ def __init__(
  n_classes: int,
  batch_size: int = 32,
  epochs: int = 100,
- patience: int = 10,
+ patience: int = None,
  learning_rate: float = 1e-3,
  weight_decay: float = 1e-5,
  num_workers: int = 0,

pypots/classification/raindrop.py

@@ -637,7 +637,7 @@ def __init__(
  static,
  batch_size=32,
  epochs=100,
- patience=10,
+ patience: int = None,
  learning_rate=1e-3,
  weight_decay=1e-5,
  num_workers: int = 0,

pypots/clustering/crli.py

@@ -342,7 +342,7 @@ def __init__(
  D_steps: int = 1,
  batch_size: int = 32,
  epochs: int = 100,
- patience: int = 10,
+ patience: int = None,
  learning_rate: float = 1e-3,
  weight_decay: float = 1e-5,
  num_workers: int = 0,

pypots/clustering/vader.py

@@ -387,7 +387,7 @@ def __init__(
  batch_size: int = 32,
  epochs: int = 100,
  pretrain_epochs: int = 10,
- patience: int = 10,
+ patience: int = None,
  learning_rate: float = 1e-3,
  weight_decay: float = 1e-5,
  num_workers: int = 0,

pypots/data/generating.py

@@ -268,8 +268,14 @@ def gene_incomplete_random_walk_dataset(
  return data
 
 
-def gene_physionet2012():
- """Generate PhysioNet2012."""
+def gene_physionet2012(artificially_missing: bool = True):
+ """Generate a full-prepared PhysioNet-2012 dataset for model testing.
+
+ Parameters
+ ----------
+ artificially_missing : bool, default = True,
+ Whether to artificially mask out 10% observed values and hold out for imputation performance evaluation.
+ """
  # generate samples
  df = load_specific_dataset("physionet_2012")
  X = df["X"]
@@ -288,11 +294,13 @@ def gene_physionet2012():
  val_set.to_numpy(),
  test_set.to_numpy(),
  )
+
  # normalization
  scaler = StandardScaler()
  train_X = scaler.fit_transform(train_X)
  val_X = scaler.transform(val_X)
  test_X = scaler.transform(test_X)
+
  # reshape into time series samples
  train_X = train_X.reshape(len(train_set_ids), 48, -1)
  val_X = val_X.reshape(len(val_set_ids), 48, -1)
@@ -303,16 +311,6 @@ def gene_physionet2012():
  test_y = y[y.index.isin(test_set_ids)]
  train_y, val_y, test_y = train_y.to_numpy(), val_y.to_numpy(), test_y.to_numpy()
 
- # mask values in the validation set as ground truth
- val_X_intact, val_X, val_X_missing_mask, val_X_indicating_mask = mcar(val_X, 0.1)
- val_X = masked_fill(val_X, 1 - val_X_missing_mask, torch.nan)
-
- # mask values in the test set as ground truth
- test_X_intact, test_X, test_X_missing_mask, test_X_indicating_mask = mcar(
- test_X, 0.1
- )
- test_X = masked_fill(test_X, 1 - test_X_missing_mask, torch.nan)
-
  data = {
  "n_classes": 2,
  "n_steps": 48,
@@ -321,11 +319,26 @@ def gene_physionet2012():
  "train_y": train_y.flatten(),
  "val_X": val_X,
  "val_y": val_y.flatten(),
- "val_X_intact": val_X_intact,
- "val_X_indicating_mask": val_X_indicating_mask,
  "test_X": test_X,
  "test_y": test_y.flatten(),
- "test_X_intact": test_X_intact,
- "test_X_indicating_mask": test_X_indicating_mask,
  }
+
+ if artificially_missing:
+ # mask values in the validation set as ground truth
+ val_X_intact, val_X, val_X_missing_mask, val_X_indicating_mask = mcar(
+ val_X, 0.1
+ )
+ val_X = masked_fill(val_X, 1 - val_X_missing_mask, torch.nan)
+
+ # mask values in the test set as ground truth
+ test_X_intact, test_X, test_X_missing_mask, test_X_indicating_mask = mcar(
+ test_X, 0.1
+ )
+ test_X = masked_fill(test_X, 1 - test_X_missing_mask, torch.nan)
+
+ data["test_X_intact"] = test_X_intact
+ data["test_X_indicating_mask"] = test_X_indicating_mask
+ data["val_X_intact"] = val_X_intact
+ data["val_X_indicating_mask"] = val_X_indicating_mask
+
  return data

pypots/imputation/brits.py

@@ -517,7 +517,7 @@ def __init__(
  rnn_hidden_size: int,
  batch_size: int = 32,
  epochs: int = 100,
- patience: int = 10,
+ patience: int = None,
  learning_rate: float = 1e-3,
  weight_decay: float = 1e-5,
  num_workers: int = 0,

pypots/imputation/saits.py

@@ -183,7 +183,7 @@ def __init__(
  MIT_weight: int = 1,
  batch_size: int = 32,
  epochs: int = 100,
- patience: int = 10,
+ patience: int = None,
  learning_rate: float = 1e-3,
  weight_decay: float = 1e-5,
  num_workers: int = 0,

pypots/imputation/transformer.py

@@ -295,7 +295,7 @@ def __init__(
  MIT_weight: int = 1,
  batch_size: int = 32,
  epochs: int = 100,
- patience: int = 10,
+ patience: int = None,
  learning_rate: float = 1e-3,
  weight_decay: float = 1e-5,
  num_workers: int = 0,