fastreid.layers.am_softmax.AMSoftmax Class Reference

Inheritance diagram for fastreid.layers.am_softmax.AMSoftmax:

Public Member Functions
	__init__ (self, cfg, in_feat, num_classes)
	forward (self, features, targets)
	extra_repr (self)

Public Attributes
	in_features
	s
	m
	weight
	in_feat

Protected Attributes
	_num_classes

Detailed Description

Implement of large margin cosine distance:
Args:
    in_feat: size of each input sample
    num_classes: size of each output sample

Definition at line 13 of file am_softmax.py.

Constructor & Destructor Documentation

__init__()

fastreid.layers.am_softmax.AMSoftmax.__init__	(		self,
			cfg,
			in_feat,
			num_classes )

Definition at line 20 of file am_softmax.py.

    def __init__(self, cfg, in_feat, num_classes):
        super().__init__()
        self.in_features = in_feat
        self._num_classes = num_classes
        self.s = cfg.MODEL.HEADS.SCALE
        self.m = cfg.MODEL.HEADS.MARGIN
        self.weight = Parameter(torch.Tensor(num_classes, in_feat))
        nn.init.xavier_uniform_(self.weight)
 

Member Function Documentation

extra_repr()

fastreid.layers.am_softmax.AMSoftmax.extra_repr

(

self

)

Definition at line 40 of file am_softmax.py.

    def extra_repr(self):
        return 'in_features={}, num_classes={}, scale={}, margin={}'.format(
            self.in_feat, self._num_classes, self.s, self.m
        )

forward()

fastreid.layers.am_softmax.AMSoftmax.forward	(		self,
			features,
			targets )

Definition at line 29 of file am_softmax.py.

    def forward(self, features, targets):
        # --------------------------- cos(theta) & phi(theta) ---------------------------
        cosine = F.linear(F.normalize(features), F.normalize(self.weight))
        phi = cosine - self.m
        # --------------------------- convert label to one-hot ---------------------------
        targets = F.one_hot(targets, num_classes=self._num_classes)
        output = (targets * phi) + ((1.0 - targets) * cosine)
        output *= self.s
 
        return output
 

Member Data Documentation

_num_classes

fastreid.layers.am_softmax.AMSoftmax._num_classes

protected

Definition at line 23 of file am_softmax.py.

in_feat

fastreid.layers.am_softmax.AMSoftmax.in_feat

Definition at line 42 of file am_softmax.py.

in_features

fastreid.layers.am_softmax.AMSoftmax.in_features

Definition at line 22 of file am_softmax.py.

m

fastreid.layers.am_softmax.AMSoftmax.m

Definition at line 25 of file am_softmax.py.

s

fastreid.layers.am_softmax.AMSoftmax.s

Definition at line 24 of file am_softmax.py.

weight

fastreid.layers.am_softmax.AMSoftmax.weight

Definition at line 26 of file am_softmax.py.

---

The documentation for this class was generated from the following file:

• smreid/fastreid/layers/am_softmax.py