Loss Functions¶

Loss modules in DeepFense are unified: they handle projection, loss calculation, and score generation.

Available Losses¶

1. AM-Softmax (AMSoftmax)¶

Additive Margin Softmax. Forces classes to be separated by a margin m on the hypersphere.

Configuration Signature:

loss:
  type: AMSoftmax
  args:
    embedding_dim: int
    n_classes: int
    m: float
    s: float
    weight: float

Parameters:

• embedding_dim - (int) Input feature dimension.
• n_classes - (int) Number of classes (usually 2).
• m - (float) Margin value (e.g., 0.3).
• s - (float) Scaling factor (e.g., 30.0).
• weight - (float) Loss weight in multi-loss setups.

Example:

loss:
  type: AMSoftmax
  args:
    embedding_dim: 192
    n_classes: 2
    m: 0.3
    s: 30.0

2. A-Softmax (ASoftmax)¶

Angular Softmax (SphereFace). Uses multiplicative angular margin.

Configuration Signature:

loss:
  type: ASoftmax
  args:
    embedding_dim: int
    n_classes: int
    m: int
    gamma: float
    lambda_min: float
    lambda_max: float

Parameters:

• embedding_dim - (int) Input dimension.
• n_classes - (int) Number of classes.
• m - (int) Integer margin (e.g., 4). Controls the angular margin size.
• gamma - (float) Focusing parameter.
• lambda_min/max - (float) Parameters for lambda annealing (stabilizes training).

Example:

loss:
  type: ASoftmax
  args:
    embedding_dim: 192
    m: 4

3. OC-Softmax (OCSoftmax)¶

One-Class Softmax. Designed specifically for spoofing. It compacts the "Bonafide" class into a tight cluster and pushes "Spoof" samples away.

Configuration Signature:

loss:
  type: OCSoftmax
  args:
    embedding_dim: int
    m_real: float
    m_fake: float
    alpha: float

Parameters:

• embedding_dim - (int) Input dimension.
• m_real - (float) Margin for real speech (default: 0.5).
• m_fake - (float) Margin for fake speech (default: 0.2).
• alpha - (float) Scaling factor (default: 20.0).

Example:

loss:
  type: OCSoftmax
  args:
    embedding_dim: 192
    m_real: 0.5
    m_fake: 0.2

4. Cross Entropy (CrossEntropy)¶

Standard classification loss with a linear projection layer.

Configuration Signature:

loss:
  type: CrossEntropy
  args:
    embedding_dim: int
    n_classes: int

Parameters:

• embedding_dim - (int) Input dimension (required to build the linear layer).
• n_classes - (int) Number of classes.

Example:

loss:
  type: CrossEntropy
  args:
    embedding_dim: 192
    n_classes: 2

Adding Custom Losses¶

See Extending DeepFense for a guide on creating custom loss modules.

Next Step: Augmentations →