Model helpers

class softsensor.model.CNN(input_channels, window_size, filters, kernel_size, depth=1, pooling=None, pooling_size=2, activation='relu', bias=True, dropout=None)[source]

Convolutional Model

Parameters:

  • input_channels (int) – Number of input channels

  • window_size (int) – Size of the sliding window applied to the time series

  • filters (int or list of int) – Number of filters used in the convolution

  • kernel_size (int or list of int) – Width of the filter, (needs to be uneven)

  • depth (int,) – Depth of the Network, (how often is Convolution, Activation and Pooling repeated) The default is 1.

  • pooling (str, optional) – Pooling Variant to pool filtered time series. options are: ‘average’ and ‘max’ The default is None.

  • pooling_size (int, optional) – Kernel size of the pooling layer. The default is 2.

  • activation (str, optional) – Activation function to activate the feature space. The default is ‘relu’.

  • bias (bool, optional) – If True, bias weights are used. The default is True.

  • dropout (float [0,1], optional) – Adds dropout layers after each activation. The default is None.

Return type:

None.

Example

>>> import softsensor.model as model
>>> import torch
>>> model = model.CNN(input_channels=4, window_size=50, filters=8,
                      kernel_size=5, depth=2, pooling='average')
>>> print(model)
CNN(
  (ConvNet): Sequential(
    (0): conv_block(
      (Conv): Conv1d(4, 8, kernel_size=(5,), stride=(1,))
      (pool): AvgPool1d(kernel_size=(2,), stride=(2,), padding=(0,))
      (activation): ReLU()
    )
    (1): conv_block(
      (Conv): Conv1d(8, 8, kernel_size=(5,), stride=(1,))
      (pool): AvgPool1d(kernel_size=(2,), stride=(2,), padding=(0,))
      (activation): ReLU()
    )
  )
)
>>> input = torch.randn(32, 4, 50)
>>> output = model(input)
>>> print(output.size())
torch.Size([32, 8, 10])

To get the length of the output time series: >>> print(model.ts_length) 10

forward(inp)[source]

Forward function to probagate through the network

Parameters:

inp (torch.tensor dtype=torch.float) – Input tensor for forward propagation

Returns:

output – Output tensor

Return type:

torch.tensor dtype=torch.float()

class softsensor.model.Feed_ForwardNN(input_size, output_size, hidden_size=None, activation='relu', bias=True, dropout=None, concrete_dropout=False, bn=False)[source]

Deep Neural Network with Fully Connected Layers

Parameters:

  • input_size (int) – Size of input array.

  • output_size (int) – Size of output array.

  • hidden_size (list of int or None, optional) – List gives the size of hidden units. The default is None.

  • activation (str, optional) – Activation function to activate the feature space. The default is ‘relu’.

  • bias (bool, optional) – If True, bias weights are used. The default is True.

  • dropout (float [0,1], optional) – Adds dropout layers after each Linear Layer. The default is None.

  • concrete_dropout (bool, optional) – Whether to use normal or concrete dropout layers if dropout is not None. The default is False

Return type:

None.

Example

>>> import softsensor.model as model
>>> import torch
>>> model = model.Feed_ForwardNN(input_size=40, output_size=2,
                                 hidden_size=None, bias=True)
>>> print(model)
Feed_ForwardNN(
  (DNN): Sequential(
    (0): Linear(in_features=40, out_features=2, bias=True)
  )
)
>>> input = torch.randn(32, 40)
>>> output = model(input)
>>> print(output.size())
torch.Size([32, 2])
forward(inp)[source]

Forward function to probagate through the network

Parameters:

inp (torch.tensor dtype=torch.float) – Input tensor for forward propagation

Returns:

output – Output tensor

Return type:

torch.tensor dtype=torch.float()

class softsensor.model.Freq_Att_CNN(input_channels, filters, kernel_size, max_dilation=4, oscillations=4, depth=1, bypass=True, activation='relu', bias=True, dropout=None)[source]

Convolutional Model build on the parallel convolutions of the input with delated convolutions and subsequent convolution of the resulting parallel time series

Parameters:

  • input_channels (int) – Number of channels of the external excitation signal

  • filters (int or list of int) – Number of filters used in the convolution.

  • kernel_size (int or list of int) – Width of the filter

  • max_dilation (int) – maximum dilation in the parallel convolution. The derfault is 4

  • oscillation (int) – Number of full oscillations for each parallel convolution is processed. The default is 4

  • depth (int, optional) – Depth of the Convolutional Network. The depth is applied to the convolutions (how often is Convolution, Activation and Pooling repeated) The default is 1.

  • bypass (bool, optional) – if True bypass is applied for faster convergence. The default is True.

  • activation (str, optional) – Activation function to activate the feature space. The default is ‘relu’.

  • bias (bool, optional) – If True, bias weights are used. The default is True.

  • dropout (float [0,1], optional) – Adds dropout layers after each activation. The default is None.

Return type:

None.

Example

>>> import softsensor.model as model
>>> import torch
>>> model = model.Freq_Att_CNN(input_channels=2, filters=4, kernel_size=5,
                               max_dilation=2, oscillations=2)
>>> print(model)
Freq_Att_CNN(
  (stage1): _Parallel_Conv(
    (Convs): ModuleList(
      (0): _delated_conv(
        (ConvNet): Sequential(
          (0): Conv1d(2, 4, kernel_size=(5,), stride=(1,))
          (1): ReLU()
        )
      )
      (1): _delated_conv(
        (ConvNet): Sequential(
          (0): Conv1d(2, 4, kernel_size=(5,), stride=(2,), dilation=(2,))
          (1): ReLU()
        )
      )
    )
  )
  >>> ws = model.window_size
  >>> input = torch.randn(32, 2, ws)
  >>> output = model(input)
  >>> print(output.size())
  torch.Size([32, 6, 2])
forward(x)[source]

Forward function to probagate through the network

Parameters:

inp (torch.tensor dtype=torch.float) – Input tensor for forward propagation

Returns:

output – Output tensor

Return type:

torch.tensor dtype=torch.float()