timesead.models.reconstruction.lstm_ae

Classes

`LSTMAEDecoder`	Base class for all neural network modules.
`LSTMAEDecoderSimple`	Reconstruct the time series using a LSTM decoder, starting with an initial hidden state from the encoder
`LSTMAEDecoderReverse`	Reconstruct the time series in the opposite direction, starting with an initial hidden state from the encoder
`LSTMAE`	Generic LSTMAE implementation
`LSTMAEMalhotra2016`	Implementation of Malhotra 2016 (https://arxiv.org/pdf/1607.00148.pdf, default parameters)
`LSTMAEMirza2018`	Mirza 2018 (http://repository.bilkent.edu.tr/bitstream/handle/11693/50234/Computer_network_intrusion_detection_using_sequential_LSTM_neural_networks_autoencoders.pdf?sequence=1)
`LSTMAEAnomalyDetector`	Base class for all neural network modules.

Functions

max_pool(→ torch.Tensor)

Module Contents

class timesead.models.reconstruction.lstm_ae.LSTMAEDecoder(enc_hidden_dimension: int, hidden_dimensions: List[int], output_dimension: int)

Bases: torch.nn.Module, abc.ABC

Base class for all neural network modules.

Your models should also subclass this class.

Modules can also contain other Modules, allowing them to be nested in a tree structure. You can assign the submodules as regular attributes:

import torch.nn as nn
import torch.nn.functional as F

class Model(nn.Module):
    def __init__(self) -> None:
        super().__init__()
        self.conv1 = nn.Conv2d(1, 20, 5)
        self.conv2 = nn.Conv2d(20, 20, 5)

    def forward(self, x):
        x = F.relu(self.conv1(x))
        return F.relu(self.conv2(x))

Submodules assigned in this way will be registered, and will also have their parameters converted when you call to(), etc.

Note

As per the example above, an __init__() call to the parent class must be made before assignment on the child.

Variables:

training (bool) – Boolean represents whether this module is in training or evaluation mode.

Parameters:

enc_hidden_dimension (int)
hidden_dimensions (List[int])
output_dimension (int)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

abstract forward(initial_hidden: torch.Tensor, seq_len: int, x: torch.Tensor = None) → torch.Tensor

Parameters:

initial_hidden (torch.Tensor)
seq_len (int)
x (torch.Tensor)

Return type:

torch.Tensor

class timesead.models.reconstruction.lstm_ae.LSTMAEDecoderSimple(enc_hidden_dimension: int, hidden_dimensions: List[int], output_dimension: int)

Bases: LSTMAEDecoder

Reconstruct the time series using a LSTM decoder, starting with an initial hidden state from the encoder that is used as input to every timestep of the decoder This corresponds to Mirza 2018

Initialize internal Module state, shared by both nn.Module and ScriptModule.

Parameters:

enc_hidden_dimension (int)
hidden_dimensions (List[int])
output_dimension (int)

lstm

forward(initial_hidden: List[torch.Tensor], seq_len: int, x: torch.Tensor = None) → torch.Tensor

Parameters:

initial_hidden (List[torch.Tensor]) – list (length hidden_layers) of tensors of shape (B, D)
seq_len (int) – int that determines the length of the produced sequence
x (torch.Tensor) – The ground truth sequence that should be reconstructed as a tensor of shape (T, B, D). This will be fed into the LSTM during training instead of the output from the previous step.

Returns:

Tensor of shape (T, B, D)

Return type:

torch.Tensor

class timesead.models.reconstruction.lstm_ae.LSTMAEDecoderReverse(enc_hidden_dimension: int, hidden_dimensions: List[int], output_dimension: int)

Bases: LSTMAEDecoder

Reconstruct the time series in the opposite direction, starting with an initial hidden state from the encoder This corresponds to Malhotra 2016

Initialize internal Module state, shared by both nn.Module and ScriptModule.

Parameters:

enc_hidden_dimension (int)
hidden_dimensions (List[int])
output_dimension (int)

lstm

linear

forward(initial_hidden: List[torch.Tensor], seq_len: int, x: torch.Tensor = None) → torch.Tensor

Parameters:

initial_hidden (List[torch.Tensor]) – tensor of shape (B, D)
seq_len (int) – int that determines the length of the produced sequence
x (torch.Tensor) – The ground truth sequence that should be reconstructed as a tensor of shape (T, B, D). This will be fed into the LSTM during training instead of the output from the previous step.

Returns:

Tensor of shape (T, B, D)

Return type:

torch.Tensor

timesead.models.reconstruction.lstm_ae.max_pool(x: torch.Tensor, dim: int = 0) → torch.Tensor

Parameters:

x (torch.Tensor)
dim (int)

Return type:

torch.Tensor

class timesead.models.reconstruction.lstm_ae.LSTMAE(input_dimension: int, hidden_dimensions=None, latent_pooling: str | Callable = 'last', decoder_class: Type[LSTMAEDecoder] = LSTMAEDecoderReverse, return_latent: bool = False)

Bases: timesead.models.common.AE, timesead.models.BaseModel

Generic LSTMAE implementation

Initialize internal Module state, shared by both nn.Module and ScriptModule.

Parameters:

input_dimension (int)
latent_pooling (Union[str, Callable])
decoder_class (Type[LSTMAEDecoder])
return_latent (bool)

encode(x: torch.Tensor) → List[torch.Tensor]

Parameters:: x (torch.Tensor)
Return type:: List[torch.Tensor]

forward(inputs: Tuple[torch.Tensor]) → torch.Tensor | Tuple[torch.Tensor, torch.Tensor]

Parameters:: inputs (Tuple[torch.Tensor]) – Tuple with a single tensor of shape (T, B, D)
Returns:: tensor of shape (T, B, D)
Return type:: Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]

class timesead.models.reconstruction.lstm_ae.LSTMAEMalhotra2016(input_dimension: int, hidden_dimensions=None)

Bases: LSTMAE

Implementation of Malhotra 2016 (https://arxiv.org/pdf/1607.00148.pdf, default parameters)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

Parameters:: input_dimension (int)

class timesead.models.reconstruction.lstm_ae.LSTMAEMirza2018(input_dimension: int, hidden_dimensions: List[int] = [64], latent_pooling: str = 'mean')

Bases: LSTMAE

Mirza 2018 (http://repository.bilkent.edu.tr/bitstream/handle/11693/50234/Computer_network_intrusion_detection_using_sequential_LSTM_neural_networks_autoencoders.pdf?sequence=1)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

Parameters:

input_dimension (int)
hidden_dimensions (List[int])
latent_pooling (str)

sigmoid

forward(inputs: Tuple[torch.Tensor]) → torch.Tensor | Tuple[torch.Tensor, torch.Tensor]

Parameters:: inputs (Tuple[torch.Tensor]) – Tuple with a single tensor of shape (T, B, D)
Returns:: tensor of shape (T, B, D)
Return type:: Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]

class timesead.models.reconstruction.lstm_ae.LSTMAEAnomalyDetector(model: LSTMAE)

Bases: timesead.models.common.AnomalyDetector

Base class for all neural network modules.

Your models should also subclass this class.

Modules can also contain other Modules, allowing them to be nested in a tree structure. You can assign the submodules as regular attributes:

import torch.nn as nn
import torch.nn.functional as F

class Model(nn.Module):
    def __init__(self) -> None:
        super().__init__()
        self.conv1 = nn.Conv2d(1, 20, 5)
        self.conv2 = nn.Conv2d(20, 20, 5)

    def forward(self, x):
        x = F.relu(self.conv1(x))
        return F.relu(self.conv2(x))

Submodules assigned in this way will be registered, and will also have their parameters converted when you call to(), etc.

Note

As per the example above, an __init__() call to the parent class must be made before assignment on the child.

Variables:: training (bool) – Boolean represents whether this module is in training or evaluation mode.
Parameters:: model (LSTMAE)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

model

fit(dataset: torch.utils.data.DataLoader) → None

Fit this anomaly detector on a dataset. Note that we assume only normal data here.

Parameters:: dataset (torch.utils.data.DataLoader) – A dataset
Return type:: None

compute_online_anomaly_score(inputs: Tuple[torch.Tensor, Ellipsis]) → torch.Tensor

Compute the online anomaly score for a batch of inputs. The output tensor must have the same shape as the output of format_targets when called with the corresponding targets for this batch. This method expects a window (or a batch of windows) as its input and should return a score for the last point in the window.

Parameters:: inputs (Tuple[torch.Tensor, Ellipsis]) – tuple of input tensors
Returns:: Tensor of shape (B,) that contains the anomaly scores for this batch
Return type:: torch.Tensor

abstract compute_offline_anomaly_score(inputs: Tuple[torch.Tensor, Ellipsis]) → torch.Tensor

Compute the offline anomaly score for a batch of inputs. The output tensor must have the same shape as the output of format_targets when called with the corresponding targets for this batch. This method expects a window (or a batch of windows) as its input and should return a score for the last point in the window.

Parameters:: inputs (Tuple[torch.Tensor, Ellipsis]) – tuple of input tensors
Returns:: Tensor of shape (N,) that contains the anomaly scores for this batch
Return type:: torch.Tensor

format_online_targets(targets: Tuple[torch.Tensor, Ellipsis]) → torch.Tensor

Format the labels for a batch of targets. The output tensor must have the same shape as the output of compute_online_anomaly_score when called with the corresponding inputs for this batch.

Parameters:: targets (Tuple[torch.Tensor, Ellipsis]) – tuple of target tensors
Returns:: Tensor of shape (B,) that contains the ground truth labels for this batch
Return type:: torch.Tensor