espnet2.spk.loss.aamsoftmax_subcenter_intertopk.ArcMarginProduct_intertopk_subcenter

source

class espnet2.spk.loss.aamsoftmax_subcenter_intertopk.ArcMarginProduct_intertopk_subcenter(nout, nclasses, scale=32.0, margin=0.2, easy_margin=False, K=3, mp=0.06, k_top=5, do_lm=False)

Bases: AbsLoss

Implement of large margin arc distance with intertopk and subcenter.

This class implements the ArcMarginProduct with intertopk and subcenter techniques for improved speaker verification. It leverages concepts from the referenced papers to enhance the model’s performance.

References

MULTI-QUERY MULTI-HEAD ATTENTION POOLING AND INTER-TOPK PENALTY FOR SPEAKER VERIFICATION. https://arxiv.org/pdf/2110.05042.pdf Sub-center ArcFace: Boosting Face Recognition by Large-Scale Noisy Web Faces. https://ibug.doc.ic.ac.uk/media/uploads/documents/eccv_1445.pdf

• Parameters:
  • in_features (int) – Size of each input sample.
  • out_features (int) – Size of each output sample.
  • scale (float , optional) – Norm of input feature. Defaults to 32.0.
  • margin (float , optional) – Margin for cos(theta + margin). Defaults to 0.2.
  • easy_margin (bool , optional) – Use easy margin if True. Defaults to False.
  • K (int , optional) – Number of sub-centers. Defaults to 3.
  • mp (float , optional) – Margin penalty of hard samples. Defaults to 0.06.
  • k_top (int , optional) – Number of hard samples. Defaults to 5.
  • do_lm (bool , optional) – Whether to perform large margin finetune. Defaults to False.

in_features

Number of input features.

• Type: int

out_features

Number of output classes.

• Type: int

scale

Scaling factor for output.

• Type: float

margin

Margin value.

• Type: float

K

Number of sub-centers.

• Type: int

k_top

Number of top-k samples.

• Type: int

mp

Margin penalty for hard samples.

• Type: float

do_lm

Flag for large margin finetuning.

• Type: bool

weight

Weight parameter for the classifier.

• Type: torch.Tensor

easy_margin

Flag for easy margin setting.

• Type: bool

cos_m

Cosine of the margin.

• Type: float

sin_m

Sine of the margin.

• Type: float

th

Threshold for cosine value.

• Type: float

mm

Margin adjustment factor.

• Type: float

mm

m

Additional margin adjustment factor.

• Type: float

cos_m

p

Cosine for margin penalty.

• Type: float

sin_m

p

Sine for margin penalty.

• Type: float

ce

Cross-entropy loss function.

• Type: nn.CrossEntropyLoss

######### Examples

>>> arc_margin = ArcMarginProduct_intertopk_subcenter(
...     nout=512, nclasses=10, scale=32.0, margin=0.2, K=3, k_top=5
... )
>>> input_tensor = torch.randn(16, 512)  # Batch of 16 samples
>>> label_tensor = torch.randint(0, 10, (16, 1))  # Random labels
>>> loss = arc_margin.forward(input_tensor, label_tensor)

NOTE

Ensure that the input tensor is normalized before passing it to the forward method to achieve optimal results.

• Raises:ValueError – If the input and label sizes do not match.

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

forward(input, label)

Computes the forward pass for the ArcMarginProduct.

This method applies the ArcMarginProduct calculation to the input features and computes the loss based on the provided labels. It normalizes the input features and weight parameters, calculates the cosine and sine values for the margin, and applies the necessary transformations to generate the output.

• Parameters:
  • input (torch.Tensor) – The input features of shape (batch_size, in_features).
  • label (torch.Tensor) – The ground truth labels of shape (batch_size, num_classes) or (batch_size, 1).
• Returns: The computed loss value.
• Return type: torch.Tensor
• Raises:ValueError – If the input tensor does not match the expected dimensions.

######### Examples

>>> model = ArcMarginProduct_intertopk_subcenter(nout=512, nclasses=10)
>>> input = torch.randn(32, 512)  # Example input for batch size 32
>>> label = torch.randint(0, 10, (32,))  # Random labels for 10 classes
>>> loss = model(input, label)
>>> print(loss)

NOTE

Ensure that the input tensor is normalized before passing it to this method for optimal performance.

update(margin=0.2)

Implement of large margin arc distance with intertopk and subcenter.

This class implements the ArcMarginProduct with enhancements for speaker verification tasks. It includes intertopk and subcenter techniques to improve the robustness of the margin-based loss function.

Reference: : MULTI-QUERY MULTI-HEAD ATTENTION POOLING AND INTER-TOPK PENALTY FOR SPEAKER VERIFICATION. https://arxiv.org/pdf/2110.05042.pdf Sub-center ArcFace: Boosting Face Recognition by Large-Scale Noisy Web Faces. https://ibug.doc.ic.ac.uk/media/uploads/documents/eccv_1445.pdf

in_features

Size of each input sample.

• Type: int

out_features

Size of each output sample.

• Type: int

scale

Norm of input feature.

• Type: float

margin

Margin for the arc margin calculation.

• Type: float

K

Number of sub-centers.

• Type: int

k_top

Number of hard samples.

• Type: int

mp

Margin penalty of hard samples.

• Type: float

do_lm

Whether to perform large margin finetune.

• Type: bool

• Parameters:

  • nout (int) – Number of output features.
  • nclasses (int) – Number of classes for classification.
  • scale (float , optional) – Norm of input feature. Defaults to 32.0.
  • margin (float , optional) – Margin for the arc margin calculation. Defaults to 0.2.
  • easy_margin (bool , optional) – If True, use easy margin. Defaults to False.
  • K (int , optional) – Number of sub-centers. Defaults to 3.
  • mp (float , optional) – Margin penalty of hard samples. Defaults to 0.06.
  • k_top (int , optional) – Number of hard samples. Defaults to 5.
  • do_lm (bool , optional) – If True, performs large margin finetune. Defaults to False.

######### Examples

arc_margin = ArcMarginProduct_intertopk_subcenter(nout=512, nclasses=10) arc_margin.update(margin=0.3)