Hello, I'd like to ask if I can use custom datasets to train and test the model?

Currently, I had different dataset with different image size (126x126), but still had the same characteristics as your sketch-photo dataset (ie. having set A and B, both in train and val set)

I once tricked the model by filling the val directory of sketch-photo dataset with my own dataset, and it worked. I just want to know is there a way to do this by just providing my dataset without reconfiguring the sketch-photo dataset directories

Thanks for providing the code. Just wondering, does the code support to resume training process? So I can continue the training process after a certain epoch without training from scratch again. Thanks.

In model.py, why is the training step ran twice for the generators?

  def run_optim(self,batch_A_imgs, batch_B_imgs,  counter, start_time):
        _, Adfake,Adreal,Bdfake,Bdreal, Ad, Bd = self.sess.run(
            [self.d_optim, self.Ad_loss_fake, self.Ad_loss_real, self.Bd_loss_fake, self.Bd_loss_real, self.Ad_loss, self.Bd_loss], 
            feed_dict = {self.real_A: batch_A_imgs, self.real_B: batch_B_imgs})
        _, Ag, Bg, Aloss, Bloss = self.sess.run(
            [self.g_optim, self.Ag_loss, self.Bg_loss, self.A_loss, self.B_loss], 
            feed_dict={ self.real_A: batch_A_imgs, self.real_B: batch_B_imgs})

        _, Ag, Bg, Aloss, Bloss = self.sess.run(
            [self.g_optim, self.Ag_loss, self.Bg_loss, self.A_loss, self.B_loss], 
            feed_dict={ self.real_A: batch_A_imgs, self.real_B: batch_B_imgs})

In the paper it says:

To optimize the DualGAN networks, we follow the training procedure proposed in WGAN [1]; see Alg. 1. We train the discriminators ncritic steps, then one step on generators

So, is the code doing the opposite (stepping generators twice and discriminator only once) ?

While the training is running (inside jenkins apparently) it keeps showing errors like:

c_allocator.cc:696] 18 Chunks of size 819200 totalling 14.06MiB
2017-08-27 20:07:53.768974: I C:\tf_jenkins\home\workspace\rel-win\M\windows-gpu\PY\35\tensorflow\core\common_runtime\b
c_allocator.cc:696] 9 Chunks of size 1638400 totalling 14.06MiB
2017-08-27 20:07:53.768996: I C:\tf_jenkins\home\workspace\rel-win\M\windows-gpu\PY\35\tensorflow\core\common_runtime\b
c_allocator.cc:696] 19 Chunks of size 3276800 totalling 59.38MiB
2017-08-27 20:07:53.769019: I C:\tf_jenkins\home\workspace\rel-win\M\windows-gpu\PY\35\tensorflow\core\common_runtime\b
c_allocator.cc:696] 9 Chunks of size 6553600 totalling 56.25MiB
2017-08-27 20:07:53.769041: I C:\tf_jenkins\home\workspace\rel-win\M\windows-gpu\PY\35\tensorflow\core\common_runtime\b
c_allocator.cc:696] 19 Chunks of size 13107200 totalling 237.50MiB
2017-08-27 20:07:53.769064: I C:\tf_jenkins\home\workspace\rel-win\M\windows-gpu\PY\35\tensorflow\core\common_runtime\b
c_allocator.cc:696] 52 Chunks of size 26214400 totalling 1.27GiB
2017-08-27 20:07:53.769157: I C:\tf_jenkins\home\workspace\rel-win\M\windows-gpu\PY\35\tensorflow\core\common_runtime\b
c_allocator.cc:696] 1 Chunks of size 46343680 totalling 44.20MiB
2017-08-27 20:07:53.769181: I C:\tf_jenkins\home\workspace\rel-win\M\windows-gpu\PY\35\tensorflow\core\common_runtime\b
c_allocator.cc:696] 27 Chunks of size 52428800 totalling 1.32GiB
2017-08-27 20:07:53.769202: I C:\tf_jenkins\home\workspace\rel-win\M\windows-gpu\PY\35\tensorflow\core\common_runtime\b
c_allocator.cc:700] Sum Total of in-use chunks: 3.00GiB
2017-08-27 20:07:53.769244: I C:\tf_jenkins\home\workspace\rel-win\M\windows-gpu\PY\35\tensorflow\core\common_runtime\b
c_allocator.cc:702] Stats:
Limit:                  3226265190
InUse:                  3226265088
MaxInUse:               3226265088
NumAllocs:                     808
MaxAllocSize:             52428800
...
 Resource exhausted: OOM when allocating tensor with shape[5,5,1,64]

I am not familiar to jenkins, is this normal? (I am on Windows 10, with 16GB RAM and only 30% is in use and the GPU is 1050TI 4GB)

LE: This doesn't seem to occur if I set --fcn_filter_dim 32

According to your paper, you use the loss format advocated by WGAN rather than the sigmoid cross-entropy loss used in the original GAN.

But in this repo, it looks no WGAN. That's why?

I run with the argument --use_labeled_data 'semi', and got an error : 'AttributeError: 'DualNet' object has no attribute 'C_d_vars' . I uncommented some code in the model.py, but still have problem running the code. Please give me some advice. Thanks.

Hi, This is great and very well engineered, well done. I do have a question about the identity loss which you see in a few papers on dual gans, loss( A2B, A ). I'm not sure I can see it there, is there a reason for this?

add_argument('--niter', dest='niter', type=int, default=30, help='# of iter at starting learning rate')

This argument is added but seems to never be used. Should it be removed or be implemented?

I have no ideas that why the epoch be set 45 in test stage.Please give me some advice，thanks! python main.py --phase test --dataset_name sketch-photo --image_size 256 --epoch 45 --lambda_A 20.0 --lambda_B 20.0 --A_channels 1 --B_channels 1

Hello, I'd like to know is it possible to train the Dual GAN without the discrinminator? which means I only need to minimize the "Reconstruction Error" during training?

Many thanks!

when I was trying to train a sketch-photo model, it caused the segmentation fault. And I changed to another datasets, the fault happened, too.

I used: python main.py --phase train --dataset_name sketch-photo --image_size 256 --epoch 45 --lambda_A 20.0 --lambda_B 20.0 --A_channels 1 --B_channels 1

Could you please help me?

I read your code about the design for loss and found your implementation is different from that proposed in the paper. So you use the traditional loss of GAN instead of the WGAN loss? Does it mean WGAN loss might not be a good choice in practice?