training/loss/am_softmax.py

"""
 Copyright (c) 2018 Intel Corporation
 Licensed under the Apache License, Version 2.0 (the "License");
 you may not use this file except in compliance with the License.
 You may obtain a copy of the License at
      http://www.apache.org/licenses/LICENSE-2.0
 Unless required by applicable law or agreed to in writing, software
 distributed under the License is distributed on an "AS IS" BASIS,
 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License.
"""

import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn import Parameter
import torch as th

from .abstract_loss_func import AbstractLossClass
from metrics.registry import LOSSFUNC


#------------ AMSoftmax Loss ----------------------   

def focal_loss(input_values, gamma):
    """Computes the focal loss"""
    p = torch.exp(-input_values)
    loss = (1 - p) ** gamma * input_values
    return loss.mean()


@LOSSFUNC.register_module(module_name="am_softmax")
class AMSoftmaxLoss(AbstractLossClass):
    """Computes the AM-Softmax loss with cos or arc margin"""
    margin_types = ['cos', 'arc']

    def __init__(self, margin_type='cos', gamma=0., m=0.5, s=30, t=1.):
        super().__init__()
        assert margin_type in AMSoftmaxLoss.margin_types
        self.margin_type = margin_type
        assert gamma >= 0
        self.gamma = gamma
        assert m > 0
        self.m = m
        assert s > 0
        self.s = s
        self.cos_m = math.cos(m)
        self.sin_m = math.sin(m)
        self.th = math.cos(math.pi - m)
        assert t >= 1
        self.t = t

    def forward(self, cos_theta, target):
        if self.margin_type == 'cos':
            phi_theta = cos_theta - self.m
        else:
            sine = torch.sqrt(1.0 - torch.pow(cos_theta, 2))
            phi_theta = cos_theta * self.cos_m - sine * self.sin_m #cos(theta+m)
            phi_theta = torch.where(cos_theta > self.th, phi_theta, cos_theta - self.sin_m * self.m)

        index = torch.zeros_like(cos_theta, dtype=torch.uint8)
        index.scatter_(1, target.data.view(-1, 1), 1)
        output = torch.where(index, phi_theta, cos_theta)

        if self.gamma == 0 and self.t == 1.:
            return F.cross_entropy(self.s*output, target)

        if self.t > 1:
            h_theta = self.t - 1 + self.t*cos_theta
            support_vecs_mask = (1 - index) * \
                torch.lt(torch.masked_select(phi_theta, index).view(-1, 1).repeat(1, h_theta.shape[1]) - cos_theta, 0)
            output = torch.where(support_vecs_mask, h_theta, output)
            return F.cross_entropy(self.s*output, target)

        return focal_loss(F.cross_entropy(self.s*output, target, reduction='none'), self.gamma)


@LOSSFUNC.register_module(module_name="am_softmax_ohem")
class AMSoftmax_OHEM(AbstractLossClass):
    """Computes the AM-Softmax loss with cos or arc margin"""
    margin_types = ['cos', 'arc']

    def __init__(self, margin_type='cos', gamma=0., m=0.5, s=30, t=1., ratio=1.):
        super(self).__init__()
        assert margin_type in AMSoftmaxLoss.margin_types
        self.margin_type = margin_type
        assert gamma >= 0
        self.gamma = gamma
        assert m > 0
        self.m = m
        assert s > 0
        self.s = s
        self.cos_m = math.cos(m)
        self.sin_m = math.sin(m)
        self.th = math.cos(math.pi - m)
        assert t >= 1
        self.t = t
        self.ratio = ratio


    # ------- online hard example mining --------------------
    def get_subidx(self,x,y,ratio):                                                                                         
        num_inst = x.size(0)                                                       
        num_hns = int(ratio * num_inst)                                       
        x_ = x.clone()                                                             
        inst_losses = th.autograd.Variable(th.zeros(num_inst)).cuda()    

        for idx, label in enumerate(y.data):                                       
            inst_losses[idx] = -x_.data[idx, label]                                 

        _, idxs = inst_losses.topk(num_hns)                                        
        return idxs


    def forward(self, cos_theta, target):
        if self.margin_type == 'cos':
            phi_theta = cos_theta - self.m
        else:
            sine = torch.sqrt(1.0 - torch.pow(cos_theta, 2))
            phi_theta = cos_theta * self.cos_m - sine * self.sin_m #cos(theta+m)
            phi_theta = torch.where(cos_theta > self.th, phi_theta, cos_theta - self.sin_m * self.m)

        index = torch.zeros_like(cos_theta, dtype=torch.uint8)
        index.scatter_(1, target.data.view(-1, 1), 1)
        output = torch.where(index, phi_theta, cos_theta)

        out = F.log_softmax(output,dim=1)
        idxs = self.get_subidx(out,target,self.ratio) # select hard examples 

        output2 = output.index_select(0, idxs)                                             
        target2 = target.index_select(0, idxs)        

        if self.gamma == 0 and self.t == 1.:
            return F.cross_entropy(self.s*output2, target2)

        if self.t > 1:
            h_theta = self.t - 1 + self.t*cos_theta
            support_vecs_mask = (1 - index) * \
                torch.lt(torch.masked_select(phi_theta, index).view(-1, 1).repeat(1, h_theta.shape[1]) - cos_theta, 0)
            output2 = torch.where(support_vecs_mask, h_theta, output2)
            return F.cross_entropy(self.s*output2, target2)

        return focal_loss(F.cross_entropy(self.s*output2, target2, reduction='none'), self.gamma)