akanimax / msg-stylegan-tf Goto Github PK

Hi， I am using your MSG-GAN code ， and it very human-readable and good for user
I tried to read the msg-stylegan but， tensorflow version is so difficult to read ，maybe it's efficient？？

If you have proved the effectiveness of the method, I wonder if it is convenient to provide pytorch version of the code for more people to use

Thank you

I want to use the pre-trained model but I am not able to do so using CPU. I don't have a GPU and its raising error.

Hi,
Thanks for creating this repository! I'm implementing the MSG-GAN method for my own project and I'm trying to use WGAN-GP loss, but I'm finding it difficult since there are multiple images. Could you point to the WGAN-GP loss implementation with multiple images in this repo so that I can use it for reference?
Thanks,
Tharun Mohandoss

Dear all,
I would like to kindly ask further explanation about the architectural design of the network, for instance:

• Why are there multiple skip connections and how the whole thing is trained?
• Do I have to concatenate both the multi-scale layers from the generator and random real images instances while evaluating the discriminator? How do I do that?
• There's a "combinational layer" in yellow, what's that?
  I have indeed read the original paper but still, there are some things that I cannot understand so I hope you will kindly answer me.

Looking forward to hearing from you soon.

Sincerely

I'm currently working on an own implementation of your great paper during which I came up with the question, if the second generator pass in each training step is actually necessary. What I mean is, that we generate fake samples once to optimize D and then once again to do so for G. But for the second case, the weights of G actually haven't changed so the drawn fakes won't differ either. It would be different, if we used different latent samples for each pass, but this does not seem to be the case.

From my current understanding, we could simply retain the gradients from the first pass and avoid one costly generator call, reducing the computational resources by 25% or something per training iteration. But I have to admit, that I'm more familiar with your PyTorch implementation of the paper, which I took as a reference, to clarify on this question. Nonetheless, I'm asking here, as it seems to be "official" repository.

To test the model can i give the input my images which are not in the train dataset will that work?

tensorflow.python.framework.errors_impl.InvalidArgumentError: Cannot assign a device for operation G_synthesis/noise0/Initializer/random_normal/RandomStandardNormal: Could not satisfy explicit device specification '' because the node {{colocation_node G_synthesis/noise0/Initializer/random_normal/RandomStandardNormal}} was colocated with a group of nodes that required incompatible device '/device:GPU:0'. All available devices [/job:localhost/replica:0/task:0/device:CPU:0, /job:localhost/replica:0/task:0/device:XLA_GPU:0, /job:localhost/replica:0/task:0/device:XLA_GPU:1, /job:localhost/replica:0/task:0/device:XLA_GPU:2, /job:localhost/replica:0/task:0/device:XLA_CPU:0].

Hi, I am new to python and machine learning. I find this idea very useful to calculate the gradient at not just the final layer but at the intermediate layers of GAN too. If I am not wrong it means that you calculate loss between let's say: Gen3 (generator intermediate layer 3) and Dis3 (discriminator intermediate layer 3) and calculate the gradient update for that layer only. In simple words the loss is being calculated at each layer. Please correct me if I am wrong.

The problem is I am unable to find this implementation in the code provided. Could you please help in understaing the confusion?

Hi,

Thanks for sharing the implementation! I'm wondering that how many epochs you used for training LSUN church?

Thanks

Hello, I tried to resume my training from the previous learning, but cannot make it. Please can you give me a brief guide? Thank you! :)