Linear

Performs a linear transformation, also known as a fully connected layer, or dense layer.

Takes a configurable value <Units>, which specifies the number of output nodes. For each output node, computes a weighted sum of the inputs, and then adds a bias to that. This is performed over the last dimension. That is, given an input of shape (A, B, C), the output would be (A, B, <Units>), meaning that a total of A * B * <Units> weighted sums are computed.

The only supported input type is Float32, with at least one dimension. Given input of shape (*, N), where * means any number of dimensions including zero, outputs a value of shape (*, <Units>).

\(y = xA^T + \beta\)
Where \(A\) and \(\beta\)are learnable parameters. Given input of shape (*, C), the shape of \(A\) is (C, <Units>) and the shape of \(\beta\)is (<Units>).

By clicking the node the following parameters can be configured on the right panel:

• Units: The number of weighted sums to compute. This is the size of the last dimension in the output.
• Use bias: If false, the bias will not be applied.
• Number of connections: Increases the number of connections on this node. The first input will lead to the first output, the second input to the second output and so on. This is useful because even though the connections are separate they share the same learned parameters, i.e the learned mean and standard deviation.
• Weight initializer: Initial value for the weight.
• Bias initializer: Initial value for the bias.