Source code for dtaianomaly.anomaly_detection.Transformer

from collections.abc import Callable
from typing import Literal

import torch

from dtaianomaly.anomaly_detection.BaseDetector import Supervision
from dtaianomaly.anomaly_detection.BaseNeuralDetector import (
    _ACTIVATION_FUNCTION_TYPE,
    _COMPILE_MODE_TYPE,
    _MODEL_PARAMETERS_TYPE,
    _OPTIMIZER_TYPE,
)
from dtaianomaly.anomaly_detection.BaseNeuralDetector_utils import (
    BaseNeuralForecastingDetector,
)


[docs] class Transformer(BaseNeuralForecastingDetector): """ Use a transformer to detect anomalies :cite:`vaswani2017attention`. A transformer anomaly detector first forecasts the time series, and then detects anomalies by measuring the deviation from the forecasted values to the actual observations. A transformer is a neural network consisting of only attention-layers: all you need is attention. The forecasting network therefore consists first of a transformer-encoder, which is connected to a linear layer to forecast the time series. The architecture of the transformer consists of 2 blocks: (1) a transformer -decoder consisting of one or more attention-layers, and (2) a single linear layer. Parameters ---------- window_size: int or str The window size to use for extracting sliding windows from the time series. This value will be passed to :py:meth:`~dtaianomaly.anomaly_detection.compute_window_size`. error_metric: {"mean-absolute-error", "mean-squared-error"}, default="mean-absolute-error" The error measure between the reconstructed window and the original window. num_heads: int, default=12 The number of heads in each attention layer. num_transformer_layers: int, default=1 The number of attention layers. dimension_feedforward: int, default=32, The dimension of the linear layer at the end of each attention layer. bias: bool, default=True Whether to use bias weights in each layer of the LSTM block. dropout_rate: float in interval [0, 1[, default=0.0 The dropout rate to put on each attention layer. activation_function: {"linear", "relu", "sigmoid", "tanh"} default="relu" The activation function to use at the end of each attention layer. stride: int, default=1 The stride, i.e., the step size for extracting sliding windows from the time series. standard_scaling: bool, default=True Whether to standard scale each window independently, before feeding it to the network. batch_size: int, default=32 The size of the batches to feed to the network. data_loader_kwargs: dictionary, default=None Additional kwargs to be passed to the data loader. For more information, see: https://docs.pytorch.org/docs/stable/data.html optimizer: {"adam", "sgd"} or callable default="adam" The optimizer to use for learning the weights. If "adam" is given, then the torch.optim.Adam optimizer will be used. If "sgd" is given, then the torch.optim.SGD optimizer will be used. Otherwise, a callable should be given, which takes as input the network parameters, and then creates an optimizer. learning_rate: float, default=1e-3 The learning rate to use for training the network. Has no effect if optimize is a callable. compile_model: bool, default=False Whether the network architecture should be compiled or not before training the weights. For more information, see: https://docs.pytorch.org/docs/stable/generated/torch.compile.html compile_mode: {"default", "reduce-overhead", "max-autotune", "max-autotune-no-cudagraphs"}, default="default" Method to compile the architecture. For more information, see: https://docs.pytorch.org/docs/stable/generated/torch.compile.html n_epochs: int, default=10 The number of epochs for which the neural network should be trained. loss_function: torch.nn.Module, default=torch.nn.MSELoss() The loss function to use for updating the weights. device: str, default="cpu" The device on which te neural network should be trained. For more information, see: https://docs.pytorch.org/docs/stable/tensor_attributes.html#torch-device seed: int, default=None The seed used for training the model. This seed will update the torch and numpy seed at the beginning of the fit method. Attributes ---------- window_size_: int The effectively used window size for this anomaly detector. optimizer_: torch.optim.Optimizer The optimizer used for learning the weights of the network. neural_network_: torch.nn.Module The PyTorch network architecture. Examples -------- >>> from dtaianomaly.anomaly_detection import Transformer >>> from dtaianomaly.data import demonstration_time_series >>> x, y = demonstration_time_series() >>> transformer = Transformer(10, seed=0).fit(x) >>> transformer.decision_function(x) # doctest: +ELLIPSIS, +NORMALIZE_WHITESPACE, +SKIP array([0.41845179, 0.41845179, 0.3603762 , ..., 0.46213843, 0.63743933, 0.08675425]...) See also -------- BaseNeuralForecastingDetector: Use a neural network to forecast the time series, and detect anomalies by measuring the difference with the actual observations. """ num_heads: int num_transformer_layers: int dimension_feedforward: int bias: bool dropout_rate: float activation_function: _ACTIVATION_FUNCTION_TYPE def __init__( self, window_size: str | int, error_metric: Literal[ "mean-absolute-error", "mean-squared-error" ] = "mean-absolute-error", forecast_length: int = 1, num_heads: int = 12, num_transformer_layers: int = 1, dimension_feedforward: int = 32, bias: bool = True, dropout_rate: float = 0.0, activation_function: _ACTIVATION_FUNCTION_TYPE = "relu", stride: int = 1, standard_scaling: bool = True, batch_size: int = 32, data_loader_kwargs: dict[str, any] = None, optimizer: ( _OPTIMIZER_TYPE | Callable[[_MODEL_PARAMETERS_TYPE], torch.optim.Optimizer] ) = "adam", learning_rate: float = 1e-3, compile_model: bool = False, compile_mode: _COMPILE_MODE_TYPE = "default", n_epochs: int = 10, loss_function: torch.nn.Module = torch.nn.MSELoss(), device: str = "cpu", seed: int = None, ): super().__init__( window_size=window_size, supervision=Supervision.SEMI_SUPERVISED, error_metric=error_metric, forecast_length=forecast_length, stride=stride, standard_scaling=standard_scaling, batch_size=batch_size, data_loader_kwargs=data_loader_kwargs, optimizer=optimizer, learning_rate=learning_rate, compile_model=compile_model, compile_mode=compile_mode, n_epochs=n_epochs, loss_function=loss_function, device=device, seed=seed, ) if not isinstance(num_heads, int) or isinstance(num_heads, bool): raise TypeError("`num_heads` should be integer") if num_heads < 1: raise ValueError("`num_heads` should be strictly positive") if not isinstance(num_transformer_layers, int) or isinstance( num_transformer_layers, bool ): raise TypeError("`num_transformer_layers` should be integer") if num_transformer_layers < 1: raise ValueError("`num_transformer_layers` should be strictly positive") if not isinstance(dimension_feedforward, int) or isinstance( dimension_feedforward, bool ): raise TypeError("`dimension_feedforward` should be integer") if dimension_feedforward < 1: raise ValueError("`dimension_feedforward` should be strictly positive") if not isinstance(bias, bool): raise TypeError("`bias` should be a bool") if not isinstance(dropout_rate, (float, int)) or isinstance(dropout_rate, bool): raise TypeError("`dropout_rate` should be a list of floats or a float") if not 0.0 <= dropout_rate < 1.0: raise ValueError(f"`dropout_rate` should be in interval [0, 1[.") if not isinstance(activation_function, str): raise TypeError("`activation_function` should be a string") if activation_function not in self._ACTIVATION_FUNCTIONS: raise ValueError( f"Unknown `activation_function` '{activation_function}'. Valid options are {list(self._ACTIVATION_FUNCTIONS.keys())}" ) self.num_heads = num_heads self.num_transformer_layers = num_transformer_layers self.dimension_feedforward = dimension_feedforward self.bias = bias self.dropout_rate = dropout_rate self.activation_function = activation_function def _build_architecture(self, n_attributes: int) -> torch.nn.Module: transformer = torch.nn.Sequential() transformer.add_module("flatten", torch.nn.Flatten()) d_model = n_attributes * self.window_size_ nhead = _adjust_nhead(d_model, self.num_heads) transformer.add_module( "transformer", torch.nn.TransformerEncoder( encoder_layer=torch.nn.TransformerEncoderLayer( d_model=d_model, nhead=nhead, dim_feedforward=self.dimension_feedforward, dropout=self.dropout_rate, activation=self._build_activation_function( self.activation_function ), batch_first=True, bias=self.bias, ), num_layers=self.num_transformer_layers, enable_nested_tensor=(nhead % 2) == 0, ), ) transformer.add_module( "linear", torch.nn.Linear( in_features=n_attributes * self.window_size_, out_features=n_attributes * self.forecast_length, ), ) return transformer
def _adjust_nhead(d_model, nhead) -> int: """ Computes a valid nhead for the given parameters, such that the constraint (d_model // nhead) * nhead == d_model is satisfied. This is done by finding the value closest to nhead which satisfies the constraint. This value can thus be larger or smaller. If the constraint is not satisfied by the given values, and there are two values equally far from nhead that satisfy the constraint, then the smaller value is returned (e.g., the constraint is not satisfied if d_model=100 and nhead=3, but it is satisfied for both 2 and 4, but 2 will be returned since it is lower). """ if d_model % nhead == 0: return nhead # Already valid # Search for closest valid nhead lower = nhead - 1 upper = nhead + 1 while lower > 1 and upper <= d_model: if d_model % lower == 0: return lower if d_model % upper == 0: return upper lower -= 1 upper += 1 return 1