There is no 'train_cogan_usps2mnist.py'
But, in CoGAN-pytorch readme, you said that 'python train_cogan_usps2mnist.py' ..

why?

I run the mnist example without any problem.
But when I try to reimplement the example on the RGBD datset in your paper, I can't genarate the right objects. I exactly followed your parameters in your supplementary materials, but I can't get it right.
Could your please update an example for the RGBD dataset, or simply provide the .train.ptt file for me to verify my errors? That will help me a lot.
Thank you.

Hello,

I have an issue running the code for Pytorch. I am trying to run the simple example:
cd cogan_pytorch/src;
python train_cogan_mnistedge.py --config ../exps/mnistedge_cogan.yaml;

However, I get the following output and error:
python train_cogan_mnistedge.py --config ../exps/mnistedge_cogan.yaml
self.scale=2.0
self.log='../outputs/mnistedge_cogan/mnistedge_cogan.log'
self.snapshot_prefix='../outputs/mnistedge_cogan/mnistedge_cogan'
self.max_iter=5000
self.batch_size=64
self.bias=0.5
self.latent_dims=100
self.snapshot_iter=500
self.display=10
<net_config.NetConfig object at 0x7fd9cd2def90>
Traceback (most recent call last):
File "train_cogan_mnistedge.py", line 95, in
main(sys.argv)
File "train_cogan_mnistedge.py", line 35, in main
os.remove(config.log)
OSError: [Errno 21] Is a directory: '../outputs/mnistedge_cogan/mnistedge_cogan.log'

Any help would be greatly appreciated! Thanks!

The Network archtecture of Discriminator is strange in "MNIST_EDGE" task.
Why the number of ouput channel is 2?
When I train the train_cogan_mnist.py, the result is awful. It seems Generator can not learn much from training, G_loss keeps a high value while D_loss is always much samller.
I am a little doubt with the effectiveness of this pytorch version code

Please forgive me asking a stupid question.
The CoGAN can be trained end-to-end.
But according to GAN principle, the parameters of Discriminators updated by gradient descent and the parameters of Generators updated by gradient ascent.
I don't understand how do you implement these update steps.
Could you figure out for me? Which parts of your code fix these problem?
Thank you.

Hello,

I am trying to run the code. Can you please tell me how to run this and specify any dependencies? There was no README file included with the code.

Thank You,
Maria

Which components did you add to Caffe?

I would like to train this net with your caffe net. Is there any instructions to this?

I cannot find the download link. Also, the link I found somewhere else is not valid anymore.

method of weight sharing could be applied in natural language?

Hi There,

For CoGAN learn to generate corresponding faces with eye-glasses and without eye-glasses.
Any help is much appreciated

How to train and test on my own dataset?

I get *** OSError: Not a gzipped file (b'\n\n') when loading the usps dataset in cogan_pytorch.
Thanks.

Hi, Great way of doing GANs in caffe with the toggle params inside the layer!

I have a minor question: In your scripts, you have specified the learning rates for the batch norm parameters as 1 during training. In traditional caffe-docs, the lr_params for batch norm mean/variance parameters are set to zero since these statistics are accumulated.

I am wondering why you have used the mean variance parameters as learnable params ?

I am using pytorch version to train the mnistedge, but after more than 20000 iterations, the results are like random noise. Can someone offered help?!