head

A module for network working as a regression or classification head.

FullyConnectedHead

Bases: Module

A fully connected (FC) network that can be used as a RUL regressor or a domain discriminator.

This network is a stack of fully connected layers with ReLU activation functions by default. The activation function can be customized through the act_func parameter. If the last layer of the network should not have an activation function, act_func_on_last_layer can be set to False.

The data flow is as follows: Inputs --> FC x n --> Outputs

The expected input shape is [batch_size, num_features].

Examples:

Default

>>> import torch
>>> from rul_adapt.model import FullyConnectedHead
>>> regressor = FullyConnectedHead(32, [16, 1])
>>> outputs = regressor(torch.randn(10, 32))
>>> outputs.shape
torch.Size([10, 1])
>>> type(outputs.grad_fn)
<class 'ReluBackward0'>

Custom activation function

>>> import torch
>>> from rul_adapt.model import FullyConnectedHead
>>> regressor = FullyConnectedHead(32, [16, 1], act_func=torch.nn.Sigmoid)
>>> outputs = regressor(torch.randn(10, 32))
>>> type(outputs.grad_fn)
<class 'SigmoidBackward0'>

Without activation function on last layer

>>> import torch
>>> from rul_adapt.model import FullyConnectedHead
>>> regressor = FullyConnectedHead(32, [16, 1], act_func_on_last_layer=False)
>>> outputs = regressor(torch.randn(10, 32))
>>> type(outputs.grad_fn)
<class 'AddmmBackward0'>

__init__(input_channels, units, dropout=0.0, act_func=nn.ReLU, act_func_on_last_layer=True)

Create a new fully connected head network.

The units are the number of output units for each FC layer. The number of output features is units[-1]. If dropout is used, it is applied in each layer, including input.

Parameters:

Name	Type	Description	Default
input_channels	int	The number of input channels.	required
units	List[int]	The number of output units for the FC layers.	required
dropout	float	The dropout probability before each layer. Set to zero to deactivate.	0.0
act_func	Type[Module]	The activation function for each layer.	ReLU
act_func_on_last_layer	bool	Whether to add the activation function to the last layer.	True