View source on GitHub

Returns y = alpha * ln(1 + exp(x / alpha)) or min(y, clip).

This can be seen as a softplus applied to the scaled input, with the output appropriately scaled. As alpha tends to 0, scaled_softplus(x, alpha) tends to relu(x). The clipping is optional. As alpha->0, scaled_softplus(x, alpha) tends to relu(x), and scaled_softplus(x, alpha, clip=6) tends to relu6(x).

x A Tensor of inputs.
alpha A Tensor, indicating the amount of smoothness. The caller must ensure that alpha > 0.
clip (optional) A Tensor, the upper bound to clip the values.
name A name for the scope of the operations (optional).

A tensor of the size and type determined by broadcasting of the inputs.