nsl.lib.adv_regularizer

View source on GitHub

Calculates adversarial loss from generated adversarial samples.

nsl.lib.adv_regularizer(
    adv_neighbors,
    target_scores,
    model_fn,
    loss_fn
)

Args:

• adv_neighbors: dense float32 tensor, with two possible shapes: (a) [batch_size, feat_len] for pointwise samples, or (b) [batch_size, seq_len, feat_len] for sequence samples.
• target_scores: target tensor used to compute loss.
• model_fn: a method that has input tensor (same shape as adv_neighbors), is_train and reuse as inputs, and returns predicted logits.
• loss_fn: a loss function that has target and prediction as inputs, and returns a float32 scalar.

Returns:

• adv_loss: a float32 denoting the adversarial loss.