espnet.nets.pytorch_backend.e2e_tts_tacotron2.GuidedAttentionLoss

source

class espnet.nets.pytorch_backend.e2e_tts_tacotron2.GuidedAttentionLoss(sigma=0.4, alpha=1.0, reset_always=True)

Bases: Module

Guided attention loss function module.

This module calculates the guided attention loss described in Efficiently Trainable Text-to-Speech System Based on Deep Convolutional Networks with Guided Attention, which forces the attention to be diagonal.

Initialize guided attention loss module.

• Parameters:
  • sigma (float , optional) – Standard deviation to control how close attention to a diagonal.
  • alpha (float , optional) – Scaling coefficient (lambda).
  • reset_always (bool , optional) – Whether to always reset masks.

forward(att_ws, ilens, olens)

Calculate forward propagation.

• Parameters:
  • att_ws (Tensor) – Batch of attention weights (B, T_max_out, T_max_in).
  • ilens (LongTensor) – Batch of input lengths (B,).
  • olens (LongTensor) – Batch of output lengths (B,).
• Returns: Guided attention loss value.
• Return type: Tensor