Source code for torchflare.modules.am_softmax

"""Implements AM-softmax."""
import torch
import torch.nn as nn
import torch.nn.functional as F


[docs]class AMSoftmax(nn.Module): """Implementation of `Additive Margin Softmax <https://arxiv.org/abs/1801.05599>`_. Args: in_features: Size of the input features out_features: The size of output features(usually number of num_classes) s: The norm for input features. m: margin """ def __init__(self, in_features, out_features, m=0.35, s=32): """Class Constructor.""" super(AMSoftmax, self).__init__() self.in_features = in_features self.out_features = out_features self.m = m self.s = s self.eps = 1e-7 self.Weight = nn.Parameter(torch.FloatTensor(self.out_features, self.in_features)) nn.init.xavier_uniform_(self.Weight)
[docs] def forward(self, features: torch.Tensor, targets: torch.Tensor = None) -> torch.Tensor: """Forward Pass. Args: features: The input features of shape (BS x F) \ where BS is batch size and F is input feature dimension. targets: The targets with shape BS , where BS is batch size Returns: Logits with shape (BS x out_features) """ cos_theta = F.linear(F.normalize(features), F.normalize(self.Weight)) if targets is None: return cos_theta one_hot = torch.zeros_like(cos_theta) one_hot.scatter_(1, targets.view(-1, 1).long(), 1) logits = torch.where(one_hot.bool(), cos_theta - self.m, cos_theta) logits = torch.cos(logits) logits *= self.s return logits