main.py 文件源码

def forward(self, prob, target, reward):
        """
        Args:
            prob: (N, C), torch Variable 
            target : (N, ), torch Variable
            reward : (N, ), torch Variable
        """
        N = target.size(0)
        C = prob.size(1)
        one_hot = torch.zeros((N, C))
        if prob.is_cuda:
            one_hot = one_hot.cuda()
        one_hot.scatter_(1, target.data.view((-1,1)), 1)
        one_hot = one_hot.type(torch.ByteTensor)
        one_hot = Variable(one_hot)
        if prob.is_cuda:
            one_hot = one_hot.cuda()
        loss = torch.masked_select(prob, one_hot)
        loss = loss * reward
        loss =  -torch.sum(loss)
        return loss