Softmax GAN example does not produce good looking digits

[mtrencseni]

🐛 Bug

I think the softmax GAN example is buggy, it doesn't produce good digits after 100-200 epochs.

This: https://github.com/PyTorchLightning/pytorch-lightning/blob/master/pl_examples/domain_templates/generative_adversarial_net.py

To Reproduce

https://colab.research.google.com/drive/1T6TpBvtFt14UrvCwDgP3eIAzaz-Af_e9#scrollTo=8Dq7kWkVF31y

Expected behavior

Good looking digits are produced.

[carmocca]

Hi! Thanks for reporting this issue.

I know what's the problem. Will try to fix it as soon as possible.

[mtrencseni]

@carmocca in the example, where should I call optimizer.zero_grad() to make it work?

[carmocca]

Before you compute the loss. I've updated your colab link.

However, we will be rolling out a solution in #6147

[mtrencseni]

I can't find zero_grad in there. Should it be in training_step()?
Btw. I don't think you can edit my colab file, no?
Sorry, I'm not an expert at PL (yet), I don't know how to retrieve the optimizer object in these callbacks.

[carmocca]

This is the updated training_step

    def training_step(self, batch, batch_idx, optimizer_idx):
        imgs, _ = batch
        # sample noise
        z = torch.randn(imgs.shape[0], self.hparams.latent_dim)
        z = z.type_as(imgs)

        # CHANGE 
        g_opt, d_opt = self.optimizers()

        # train generator
        if optimizer_idx == 0:
            # generate images
            self.generated_imgs = self(z)
            # log sampled images
            sample_imgs = self.generated_imgs[:6]
            grid = torchvision.utils.make_grid(sample_imgs)
            self.logger.experiment.add_image('generated_images', grid, 0)
            # ground truth result (ie: all fake)
            # put on GPU because we created this tensor inside training_loop
            valid = torch.ones(imgs.size(0), 1)
            valid = valid.type_as(imgs)

            # CHANGE 
            g_opt.zero_grad()

            # adversarial loss is binary cross-entropy
            g_loss = self.adversarial_loss(self.discriminator(self(z)), valid)
            tqdm_dict = {'g_loss': g_loss}
            output = OrderedDict({
                'loss': g_loss,
                'progress_bar': tqdm_dict,
                'log': tqdm_dict
            })
            return output
        # train discriminator
        if optimizer_idx == 1:
            # Measure discriminator's ability to classify real from generated samples
            # how well can it label as real?
            valid = torch.ones(imgs.size(0), 1)
            valid = valid.type_as(imgs)
            real_loss = self.adversarial_loss(self.discriminator(imgs), valid)
            # how well can it label as fake?
            fake = torch.zeros(imgs.size(0), 1)
            fake = fake.type_as(imgs)
            fake_loss = self.adversarial_loss(
                self.discriminator(self(z).detach()), fake)

            # CHANGE 
            d_opt.zero_grad()

            # discriminator loss is the average of these
            d_loss = (real_loss + fake_loss) / 2
            tqdm_dict = {'d_loss': d_loss}
            output = OrderedDict({
                'loss': d_loss,
                'progress_bar': tqdm_dict,
                'log': tqdm_dict
            })
            return output

[mtrencseni]

Awesome, thanks!

[mtrencseni]

Btw. do you know how many epochs are really needed for good looking digits?

[carmocca]

Not really, sorry! Feel free to open a PR adding a comment about it if you find out 😉

[mtrencseni]

Doesn't work, this is what I get after 100 epochs:

[akihironitta]

@mtrencseni I ran the original example code (not your code on Google Colab), but, for an unknown reason, I didn't get a similar result to yours...

The generated image below is at epoch 15 with batch_size=64.

[akihironitta]

@mtrencseni Could you try the example (https://github.com/PyTorchLightning/pytorch-lightning/blob/master/pl_examples/domain_templates/generative_adversarial_net.py) again? If the problem still persists, please feel free to reopen this issue.