Layers

class softsensor.layers.ConcreteDropout(layer, init_min=0.1, init_max=0.1)

Dropout layer that uses continuous relaxation of dropout’s discrete masks This allows for automatic tuning of the dropout probability during training, resulting in a more robust method

See “Concrete Dropout” [Gal et al. 2017 https://arxiv.org/pdf/1705.07832.pdf]

Parameters:

• layer (Preceding layer that the weight dropout is applied to)

• weight_regularizer (Penalty for large weights that considers dropout probability)

• dropout_regularizer (Penalty for small dropout rate (entropy of dropout))

• init_min (Minimum of dropout distribution)

• init_max (Maximum of dropout distribution)

Return type:

None

Note

See “Concrete Dropout” [Gal et al. 2017 https://arxiv.org/pdf/1705.07832.pdf]

forward(x)

Forward function to apply concrete dropout to the outputs of self.layer

Parameters:

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

Returns:

out – Output tensor

Return type:

torch.tensor dtype=torch.float()