Add NeuTraL-AD model, loss, detector, transform and experiments

README.md

@@ -22,6 +22,12 @@ The paper can be found [here](https://openreview.net/forum?id=iMmsCI0JsS).
 
 For installation and usage guides please refer to the [documentation](https://timesead.readthedocs.io/en/latest).
 
+## Third-party models
+
+TimeSeAD includes the NeuTraL-AD (Neural Transformation Learning for Anomaly Detection) implementation, which is
+licensed under the AGPL-3.0 license by Robert Bosch GmbH. See `timesead/models/other/neutral_ad.py` for the license
+header and make sure the AGPL requirements are compatible with your intended use.
+
 ## Citation and Contact
 
 If you use our work, please consider citing the paper

experiment_configs/recon/other/train_neutral_ad_recon.yml

@@ -0,0 +1,46 @@
+params:
+  training_experiment: other.train_neutral_ad
+  validation_metric: best_ts_f1_score
+  evaluation_metrics:
+    - best_ts_f1_score
+    - ts_auprc
+    - best_ts_f1_score_classic
+    - ts_auprc_unweighted
+    - best_f1_score
+    - auprc
+dataset:
+  name: SMDMiniDataset
+training_param_updates:
+  dataset:
+    name: SMDMiniDataset
+  training:
+    epochs: 25
+    batch_size: 32
+    drop_last: True
+    loss:
+      args:
+        temperature: 0.1
+        use_euclidean: False
+training_param_grid:
+  model_params:
+    num_trans:
+      - 4
+    trans_type:
+      - residual
+    enc_hdim:
+      - 32
+    enc_nlayers:
+      - 4
+    trans_nlayers:
+      - 4
+    latent_dim:
+      - 32
+    batch_norm:
+      - False
+    enc_bias:
+      - False
+  training:
+    optimizer:
+      args:
+        lr:
+          - 1.0e-3

experiment_configs/smd/other/train_neutral_ad_on_smd.yml

@@ -0,0 +1,46 @@
+params:
+  training_experiment: other.train_neutral_ad
+  validation_metric: best_ts_f1_score
+  evaluation_metrics:
+    - best_ts_f1_score
+    - ts_auprc
+    - best_ts_f1_score_classic
+    - ts_auprc_unweighted
+    - best_f1_score
+    - auprc
+dataset:
+  name: SMDDataset
+training_param_updates:
+  dataset:
+    name: SMDDataset
+  training:
+    epochs: 100
+    batch_size: 32
+    drop_last: True
+    loss:
+      args:
+        temperature: 0.1
+        use_euclidean: False
+training_param_grid:
+  model_params:
+    num_trans:
+      - 4
+    trans_type:
+      - residual
+    enc_hdim:
+      - 32
+    enc_nlayers:
+      - 4
+    trans_nlayers:
+      - 4
+    latent_dim:
+      - 32
+    batch_norm:
+      - False
+    enc_bias:
+      - False
+  training:
+    optimizer:
+      args:
+        lr:
+          - 1.0e-3

timesead/data/transforms/__init__.py

@@ -10,7 +10,7 @@
     TranslateXTransform,
 )
 from .target_transforms import ReconstructionTargetTransform, OneVsRestTargetTransform, PredictionTargetTransform, \
-    OverlapPredictionTargetTransform
+    OverlapPredictionTargetTransform, WindowLabelFilterTransform
 from .window_transform import WindowTransform
 
 from .dataset_source import DatasetSource, make_dataset_split

timesead/data/transforms/target_transforms.py

@@ -149,3 +149,29 @@ def seq_len(self) -> Union[int, List[int]]:
             return parent_seq_len - self.offset
 
         return [slen - self.offset for slen in parent_seq_len]
+
+
+class WindowLabelFilterTransform(Transform):
+    """
+    Filters windows based on their label values.
+
+    By default, only windows whose labels are entirely normal (all zeros) are kept.
+    """
+    def __init__(self, parent: Transform, label_index: int = 0, normal_value: int = 0, keep_normal: bool = True):
+        super().__init__(parent)
+        self.label_index = label_index
+        self.normal_value = normal_value
+        self.keep_normal = keep_normal
+        self.indices = []
+        for idx in range(len(parent)):
+            _, targets = parent.get_datapoint(idx)
+            labels = targets[label_index]
+            is_normal = torch.all(labels == normal_value).item()
+            if is_normal == keep_normal:
+                self.indices.append(idx)
+
+    def _get_datapoint_impl(self, item: int) -> Tuple[Tuple[torch.Tensor, ...], Tuple[torch.Tensor, ...]]:
+        return self.parent.get_datapoint(self.indices[item])
+
+    def __len__(self) -> Optional[int]:
+        return len(self.indices)

timesead/models/other/__init__.py

@@ -1,4 +1,5 @@
 from .lstm_ae_ocsvm import LSTMAEOCSVMAnomalyDetector
 from .mtad_gat import MTAD_GATLoss, MTAD_GATAnomalyDetector, MTAD_GAT
 from .ncad import NCAD, NCADAnomalyDetector, NCADTrainer
-from .thoc import THOC, THOCAnomalyDetector, THOCLoss, THOCTrainer
+from .neutral_ad import NeutralAD, NeutralADAnomalyDetector, NeutralADLoss
+from .thoc import THOC, THOCAnomalyDetector, THOCLoss, THOCTrainer