When use multiple optimizer, TypeError: unsupported operand type(s) for /: 'NoneType' and 'int'

[choieq]

I want build a Super-Resolution Network with multiple optimizer.

The code is below,

    def configure_optimizers(self):
        d_optimizer = torch.optim.Adam([{'params': self.parameters()}], lr=self.lr, betas=(0.5, 0.9))
        g_optimizer = torch.optim.Adam([{'params': self.parameters()}], lr=self.lr, betas=(0.5, 0.9))
        id_optimizer = torch.optim.Adam([{'params': self.parameters()}], lr=self.lr, betas=(0.5, 0.9))
        recon_optimizer = torch.optim.Adam([{'params': self.parameters()}], lr=self.lr, betas=(0.5, 0.9))
        # use multi optimizer
        return [g_optimizer, d_optimizer, id_optimizer, recon_optimizer]

def training_step(self, batch, batch_idx, optimizer_idx):
        print('optimizer_idx', optimizer_idx)
        # print('criterionG', next(self.criterionG.parameters()).requires_grad)
        # print('generator', next(self.generator.parameters()).requires_grad)

        lr_img, id_label, hr_img = batch

        fake_img = self(lr_img)
        d_fake = self.discriminator(fake_img)
        d_real = self.discriminator(hr_img)

        # train generator
        if optimizer_idx == 0:
            # log sampled images
            grid = torchvision.utils.make_grid(fake_img)
            self.logger.experiment.add_image('generated_images', grid, 0)
            g_loss = self.g_loss_function(d_fake, fake_img, hr_img)
            g_loss.requires_grad_(True)

            return {'loss': g_loss}

        # train discriminator
        elif optimizer_idx == 1:
            d_fake_loss = torch.mean(d_fake)
            d_real_loss = torch.mean(d_real)
            d_loss = (d_fake_loss + d_real_loss)/2
            tqdm_dict ={'d_loss': d_loss}
            self.log('d_loss', d_loss)

            return {'d_loss': d_loss}

        # fine-tuning arcface model
        elif optimizer_idx == 2:
            fake_img = self.conv1(fake_img)
            pred = self.recognition(fake_img)
            loss = self.loss_function(pred, id_label)
            self.log('id_loss', loss)
            tqdm_dict = {'id_loss': loss}
            output = OrderedDict({
                'id_loss': loss,
                'progress_bar': tqdm_dict,
                'log': tqdm_dict
            })
            return output

        # training reconstruction model
        elif optimizer_idx == 3:
            fake_lr = self.reconstruction(fake_img)
            loss = self.recon_loss_function(hr_img, fake_lr)
            self.log('recon_loss', loss)
            tqdm_dict = {'recon_loss': loss}
            output = OrderedDict({
                'recon_loss': loss,
                'pregress_bar': tqdm_dict,
                'log': tqdm_dict
            })
            return output

But, i got this error in 'if optimizer_idx == 0:'

    closure_loss = closure_loss / self.trainer.accumulate_grad_batches
TypeError: unsupported operand type(s) for /: 'NoneType' and 'int'

Can you give me a advice?
Thank you.

[akihironitta]

Hi @choieq, training_step needs to return one of:

• Tensor - The loss tensor
• dict - A dictionary. Can include any keys, but must include the key 'loss'
• None - Training will skip to the next batch

https://pytorch-lightning.readthedocs.io/en/latest/common/lightning_module.html#training-step

[awaelchli]

To expand on this answer, you can fix the error by adding a "loss" key to each of your output dicts.
Currently, only in your first if clause do your return a value for loss.

[choieq]

Thank you for your help, it works. :)