View source on GitHub
|
Calculates virtual adversarial loss for the given input.
nsl.lib.virtual_adv_regularizer(
input_layer, embedding_fn, virtual_adv_config, embedding=None
)
Virtual adversarial loss is defined as the distance between the embedding of the input and that of a slightly perturbed input. Optimizing this loss helps smooth models locally.
Reference paper: https://arxiv.org/pdf/1704.03976.pdf
Args | |
|---|---|
input_layer
|
a dense tensor for input features whose first dimension is the training batch size. |
embedding_fn
|
a unary function that computes the embedding for the given
input_layer input.
|
virtual_adv_config
|
an nsl.configs.VirtualAdvConfig object that specifies
parameters for generating adversarial examples and computing the
adversarial loss.
|
embedding
|
(optional) a dense tensor representing the embedding of
input_layer. If not provided, it will be calculated as
embedding_fn(input_layer).
|
Returns | |
|---|---|
virtual_adv_loss
|
a float32 denoting the virtural adversarial loss.
|