espnet2.spk.loss.aamsoftmax.AAMSoftmax

source

class espnet2.spk.loss.aamsoftmax.AAMSoftmax(nout: int, nclasses: int, margin: float = 0.3, scale: int = 15, easy_margin: bool = False, **kwargs)

Bases: AbsLoss

Additive angular margin softmax.

Reference: ArcFace: Additive Angular Margin Loss for Deep Face Recognition https://arxiv.org/pdf/1801.07698

• Parameters:
  • nout – Dimension of input features (embedding size)
  • nclasses – Number of output classes
  • margin – Angular margin for positive samples
  • scale – Feature scaling factor
  • easy_margin – Whether to use easy margin variant

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(input: Tensor, label: Tensor | None = None) → Tuple[Tensor | None, Tensor | None, Tensor]

Forward pass of AAMSoftmax loss.

• Parameters:
  • input – Input embeddings, shape (batch_size, embedding_dim)
  • label – Ground truth labels, shape (batch_size,)
• Returns: Cross-entropy loss with angular margins accuracy: Classification accuracy preds: Predicted class indices
• Return type: loss