pihnn.nn.ComplexLinear#
- class pihnn.nn.ComplexLinear(stack_features, in_features, out_features, n_domains=None, has_bias=True)#
Bases:
torch.nn.ModuleExtension of
torch.nn.Linearto complex values.So, a complex linear layer performs the operation
\[y = Wx + b,\]where the input \(x\) is a complex vector of dimension in_features, the output \(y\) and the bias \(b\) are complex vectors of dimension out_features, and the complex-valued weight tensor \(W\) has dimensions (in_features,out_features).
- Parameters:
in_features (int) – Number of input units at the current layer.
out_features (int) – Number of output units at the current layer.
has_bias (bool) – True if the current layer includes the bias term.
- init(scaling)#
Re-initialization of weights and bias.
Initialization is defined as a scaled complex-valued He initialization (Trabelsi [2018]):
\[\begin{split}\text{Re}(w),\text{Im}(w) &\sim \mathcal{N}\left(0,\frac{\texttt{scaling}}{2 \texttt{in_features}}\right), \\ bias &=0.\end{split}\]This allows us to easily include the initialization strategy from Calafà et al. [2024], Section 3.2.4.
- Parameters:
scaling (float) – Scaling in the He initialization.
- weight()#
- Returns:
weight (
torch.tensor) - Tensor with weights.
- bias()#
- Returns:
bias (
torch.tensor) - Bias vector.
- forward(input)#
Forward step \(y=Wx+b\).
- Parameters:
input (
torch.tensor) – The input vector \(x\).- Returns:
output (
torch.tensor) - The output vector \(y\).