WGan-gp test in the Celeba dataset.

[zhangqianhui]

I test the wgan-gp in the celeba dataset. But the quality of the generative images is worse than the original dcgan. and i just change the below code in the basic of w-gan using dcgan generator and discirmator.

#gradient penalty
differences = self.fake_images - self.images
 alpha = tf.random_uniform(shape=[self.batch_size, 1], minval=0., maxval=1.)
 interpolates = self.images + (alpha*differences)
gradients = tf.gradients(self.critic(interpolates, True), [interpolates])[0]
 ##2 norm
 slopes = tf.sqrt(tf.reduce_sum(tf.square(gradients), reduction_indices=[1]))
 gradient_penalty = tf.reduce_mean((slopes - 1.)**2)

And the reason?

[igul222]

That code looks correct. It's hard to say without seeing the rest of the code, but if you point me to the repo I can try and debug.

[zhangqianhui]

This repo , https://github.com/zhangqianhui/wgan-gp-debug

Thank you !

[zhangqianhui]

The sample of gerneration in epoch 4 , iter= 12601

https://github.com/zhangqianhui/wgan-gp-debug/blob/master/sample/train_04_12601.png

in epoch 6, iter= 20401

https://github.com/zhangqianhui/wgan-gp-debug/blob/master/sample/train_06_20401.png

[martinarjovsky]

Could you share the learning curve? (I.e. negative of the critic's loss)

[zhangqianhui]

the curve: https://github.com/zhangqianhui/wgan-gp-debug/blob/master/sample/curve.png

[martinarjovsky]

That doesn't look good. @igul222 did you ever see something like that?

Could you share the full code?

Best :) Martin

[zhangqianhui]

@martinarjovsky Whose code ?

[martinarjovsky]

Yours!

[wchen342]

Don't know whether it is related but, in my experiments of wgan-gp the loss of G becomes negative, which is different from original wgan in which loss of G is generally positive. Is that normal?

[zhangqianhui]

https://github.com/zhangqianhui/wgan-gp-debug @martinarjovsky

[zhangqianhui]

@igul222 @martinarjovsky hello , have you found the reason about the face generation for lower quality?

[martinarjovsky]

Hi! I haven't looked at the code yet. Can you run ishaan's code (the one on this repo) and see if it gives the same results?

[zhangqianhui]

@martinarjovsky OK

[zhangqianhui]

@martinarjovsky But , his code have not trained in celeba dataset , so Which architecture i need to use ? Is it Ok to use gan_64x64.py and dcgan's architecure?

[martinarjovsky]

That should be fine.

[zhangqianhui]

I test in this project, and it can generate very realistic face image after training in the celeba data-set. But,I can't find reason that my code don't work the same as it.

[igul222]

Here are some differences I found between your implementation and ours which might be responsible:

• Gan.py#141: You train the critic for 100 iters every 500 steps. We don't do this, and it's probably responsible for the spikes in the loss curve. Try removing it.
• Gan.py#61: It looks like self.images and self.fake_images both have shape [self.batch_size, 64, 64, self.channel]. In this case, alpha should have shape [self.batch_size, 1, 1, 1], and also reduction_indices on line 67 should be [1,2,3].
• Because you're not using any normalization, you might want to check your weight initializations. We use the initialization in https://arxiv.org/abs/1502.01852 everywhere. Alternately, you might want to just use normalization: in our default LSUN implementation we use batch normalization in the generator and layer normalization in the critic.

Hope this helps!

[zhangqianhui]

@igul222 Thanks , I have solved this problem.!

[martinarjovsky]

Cool! What was the issue?

[zhangqianhui]

@martinarjovsky igul222: Gan.py#61: It looks like self.images and self.fake_images both have shape [self.batch_size, 64, 64, self.channel]. In this case, alpha should have shape [self.batch_size, 1, 1, 1], and also reduction_indices on line 67 should be [1,2,3].

[zhangqianhui]

and my nextbatch() also have som problem.

[zhangqianhui]

And I think layer normalization is very important.

Thanks! @igul222 @martinarjovsky

[timho102003]

Hi @zhangqianhui Im new to WGAN-GP Im wondering that if we define G_Loss as g_loss = -D(Fake_Image), we're expecting a loss to converge and minimize to a number in loss curve and a maximized loss curve while define G_Loss as g_loss = D(Fake_Image)? In C_Loss(Critic loss), if we define C_Loss as c_loss = D(Fake_Image)-D(Real_Image) does that mean we're expecting a loss to converge and minimize to a number and c_loss = D(Fake_Image)-D(Real_Image) is what you called "negative of critic loss"?

[zhangqianhui]

the critic loss should be negative, because the critic loss means the negative of the divergence between the real samples distribution and fake samples distribution. You should read the wgan-gp paper for more details. And g_loss= - c_loss = - D(fake_image) + D(real_image), but the gradient of D(real_images) will not affect g network, so, g_loss= -D(fake_image)

[zhangqianhui]

@timho102003

[zhangqianhui]

Doing Classification after using training wgan?

[timho102003]

Actually I've done the multi-task on Discriminator to not only determine the real/fake problem but also classify Identities and other informations from face which has already come up with a good performance on the protocol such as MPIE. So far I'm trying on merging the WGAN to the implementation. In my network, there's a average pooling(the output is Bx320x1x1)before a fc(in my previous version, fully connected layer do the multi-task work such as real/fake, Identity...). I take away the real/fake work from fully connected layer and use "view" function to reshape the dimension of average pooling(Bx320x1x1->BX320) and serve as the output of discriminator while calculating Wasserstein Loss. From my previous experience, I feel like the model is training on the right direction. However the negative critic loss didn't start from large negative number and finally reach 0 thru training.

Generated Image (20epoch): Number of critic=5 while training, LR=0.0001 https://imgur.com/a/QVizP Negative Critics Curve : - D(fake_image) + D(real_image): https://imgur.com/a/jLn3q D Wasserstein Loss: D(fake_image) - D(real_image) +Gradient Penalty https://imgur.com/a/G5Wvl G Wasserstein Loss: - D(fake_image) https://imgur.com/a/F535n

[y601757692l]

hi,now i am also trying to train on CelebA (cropped and resized to 64x64) in WGAN-GP mode . I just modify the DATA_DIR in gan_64x64.py. But there was a mistake like this： IOError: [Errno 2] No such file or directory: '/data-4T-B/yelu/data/dcgan-completion.tensorflow/aligned/img_align_celeba_png/train_64x64/train_64x64/0927649.png' Could you show me your code? thanks so much~~~