cnn

A module of feature extractors based on convolutional neural networks.

CnnExtractor

Bases: Module

A Convolutional Neural Network (CNN) based network that extracts a feature vector from same-length time windows.

This feature extractor consists of multiple CNN layers and an optional fully connected (FC) layer. Each CNN layer can be configured with a number of filters and a kernel size. Additionally, batch normalization, same-padding and dropout can be applied. The fully connected layer can have a separate dropout probability.

Both CNN and FC layers use ReLU activation functions by default. Custom activation functions can be set for each layer type.

The data flow is as follows: Input --> CNN x n --> [FC] --> Output

The expected input shape is [batch_size, num_features, window_size]. The output of this network is always flattened to [batch_size, num_extracted_features].

Examples:

Without FC
>>> import torch
>>> from rul_adapt.model import CnnExtractor
>>> cnn = CnnExtractor(14,units=[16, 1],seq_len=30)
>>> cnn(torch.randn(10, 14, 30)).shape
torch.Size([10, 26])

With FC
>>> import torch
>>> from rul_adapt.model import CnnExtractor
>>> cnn = CnnExtractor(14,units=[16, 1],seq_len=30,fc_units=16)
>>> cnn(torch.randn(10, 14, 30)).shape
torch.Size([10, 16])

__init__(input_channels, units, seq_len, kernel_size=3, dilation=1, stride=1, padding=False, fc_units=None, dropout=0.0, fc_dropout=0.0, batch_norm=False, act_func=nn.ReLU, fc_act_func=nn.ReLU)

Create a new CNN-based feature extractor.

The units are the number of output filters for each CNN layer. The seq_len is needed to calculate the input units for the FC layer. The kernel size of each CNN layer can be set by passing a list to kernel_size. If an integer is passed, each layer has the same kernel size. If padding is true, same-padding is applied before each CNN layer, which keeps the window_size the same. If batch_norm is set, batch normalization is applied for each CNN layer. If fc_units is set, a fully connected layer is appended.

Dropout can be applied to each CNN layer by setting conv_dropout to a number greater than zero. The same is valid for the fully connected layer and fc_dropout. Dropout will never be applied to the input layer.

The whole network uses ReLU activation functions. This can be customized by setting either conv_act_func or fc_act_func.

Parameters:

Name	Type	Description	Default
input_channels	int	The number of input channels.	required
units	List[int]	The list of output filters for the CNN layers.	required
seq_len	int	The window_size of the input data.	required
kernel_size	Union[int, List[int]]	The kernel size for the CNN layers. Passing an integer uses the same kernel size for each layer.	3
dilation	int	The dilation for the CNN layers.	1
stride	int	The stride for the CNN layers.	1
padding	bool	Whether to apply same-padding before each CNN layer.	False
fc_units	Optional[int]	Number of output units for the fully connected layer.	None
dropout	float	The dropout probability for the CNN layers.	0.0
fc_dropout	float	The dropout probability for the fully connected layer.	0.0
batch_norm	bool	Whether to use batch normalization on the CNN layers.	False
act_func	Type[Module]	The activation function for the CNN layers.	ReLU
fc_act_func	Type[Module]	The activation function for the fully connected layer.	ReLU