ArcFaceClsHead

class mmpretrain.models.heads.ArcFaceClsHead(num_classes, in_channels, num_subcenters=1, scale=64.0, margins=0.5, easy_margin=False, loss={'loss_weight': 1.0, 'type': 'CrossEntropyLoss'}, init_cfg=None)

ArcFace classifier head.

A PyTorch implementation of paper ArcFace: Additive Angular Margin Loss for Deep Face Recognition and Sub-center ArcFace: Boosting Face Recognition by Large-Scale Noisy Web Faces

Example

To use ArcFace in config files.

1. use vanilla ArcFace

mode = dict(
    backbone = xxx,
    neck = xxxx,
    head=dict(
        type='ArcFaceClsHead',
        num_classes=5000,
        in_channels=1024,
        loss = dict(type='CrossEntropyLoss', loss_weight=1.0),
        init_cfg=None),
)

2. use SubCenterArcFace with 3 sub-centers

mode = dict(
    backbone = xxx,
    neck = xxxx,
    head=dict(
        type='ArcFaceClsHead',
        num_classes=5000,
        in_channels=1024,
        num_subcenters=3,
        loss = dict(type='CrossEntropyLoss', loss_weight=1.0),
        init_cfg=None),
)

3. use SubCenterArcFace With CountPowerAdaptiveMargins

mode = dict(
    backbone = xxx,
    neck = xxxx,
    head=dict(
        type='ArcFaceClsHead',
        num_classes=5000,
        in_channels=1024,
        num_subcenters=3,
        loss = dict(type='CrossEntropyLoss', loss_weight=1.0),
        init_cfg=None),
)

custom_hooks = [dict(type='SetAdaptiveMarginsHook')]

Parameters:

• num_classes (int) – Number of categories excluding the background category.

• in_channels (int) – Number of channels in the input feature map.

• num_subcenters (int) – Number of subcenters. Defaults to 1.

• scale (float) – Scale factor of output logit. Defaults to 64.0.

• margins (float) –

  The penalty margin. Could be the fllowing formats:

  • float: The margin, would be same for all the categories.

  • Sequence[float]: The category-based margins list.

  • str: A ‘.txt’ file path which contains a list. Each line represents the margin of a category, and the number in the i-th row indicates the margin of the i-th class.

  Defaults to 0.5.

• easy_margin (bool) – Avoid theta + m >= PI. Defaults to False.

• loss (dict) – Config of classification loss. Defaults to dict(type='CrossEntropyLoss', loss_weight=1.0).

• init_cfg (dict, optional) – the config to control the initialization. Defaults to None.

forward(feats, target=None)

The forward process.

loss(feats, data_samples, **kwargs)

Calculate losses from the classification score.

Parameters:

• feats (tuple[Tensor]) – The features extracted from the backbone. Multiple stage inputs are acceptable but only the last stage will be used to classify. The shape of every item should be (num_samples, num_classes).

• data_samples (List[DataSample]) – The annotation data of every samples.

• **kwargs – Other keyword arguments to forward the loss module.

Returns:

a dictionary of loss components

Return type:

dict[str, Tensor]

pre_logits(feats)

The process before the final classification head.

The input feats is a tuple of tensor, and each tensor is the feature of a backbone stage. In ArcFaceHead, we just obtain the feature of the last stage.

set_margins(margins)

set margins of arcface head.

Parameters:

margins (Union[Sequence[float], float]) – The marigins.