about frcnn loss

[cl886699]

hey its a nice work , here i have a quesion in your essays the total loss added frcnn loss , but in this code ,i did not see

` if step % self.D_update_ratio == 0 and step > self.D_init_iters: if self.cri_pix: #pixel loss l_g_pix = self.l_pix_w self.cri_pix(self.fake_H, self.var_H) l_g_total += l_g_pix if self.cri_fea: # feature loss real_fea = self.netF(self.var_H).detach() #don't want to backpropagate this, need proper explanation fake_fea = self.netF(self.fake_H) #In netF normalize=False, check it l_g_fea = self.l_fea_w self.cri_fea(fake_fea, real_fea) l_g_total += l_g_fea

        pred_g_fake = self.netD(self.fake_H)
        if self.configT['gan_type'] == 'gan':
            l_g_gan = self.l_gan_w * self.cri_gan(pred_g_fake, True)
        elif self.configT['gan_type'] == 'ragan':
            pred_d_real = self.netD(self.var_ref).detach()
            l_g_gan = self.l_gan_w * (
            self.cri_gan(pred_d_real - torch.mean(pred_g_fake), False) +
            self.cri_gan(pred_g_fake - torch.mean(pred_d_real), True)) / 2
        l_g_total += l_g_gan
        #EESN calculate loss
        self.lap_HR = kornia.laplacian(self.var_H, 3)
        if self.cri_charbonnier: # charbonnier pixel loss HR and SR
            l_e_charbonnier = 5 * (self.cri_charbonnier(self.final_SR, self.var_H)
                                    + self.cri_charbonnier(self.x_learned_lap_fake, self.lap_HR))#change the weight to empirically
        l_g_total += l_e_charbonnier
        #### did not see the frcnn loss
        l_g_total.backward(retain_graph=True)
        self.optimizer_G.step()

`

[Jakaria08]

Faster RCNN loss was backpropagated separately like GAN loss, check the later lines. Therefore, the total loss is shown by addition.