clvrai / bicyclegan-tensorflow Goto Github PK

Is there sth. wrong to substitute Leaky Relu with Relu ? I change it because tensorflow1.0 seems hasn't the Leaky Relu function and the loss is negative number.

@gitlim your work is very awesome. but when i want to extract the features of the end of the encoder path in the generator, the features' shape is (1 1 1 512). I find that all the value are 0. why ?
thanks !

Hello my name is Alejandro Barredo Arrieta and I am currently working on a research project related with GAN architectures. Would it be possible for you to share the model's checkpoint.
Thank you in advance.

When calculate loss_vae_gan, why try to minimize the difference between D_real and 0.9? why not 1.0?

Although we can pass a batch size argument in the main bicycle-gan.py code In the data_loader.py there is no implementation to fetch the respective batches..rather it always fetches a batch size of 1. Hence leading to this error.

Cannot feed value of shape (1, x) for Tensor u'input_images:0', which has shape '(2, x)'

Whenever a non-default batch is taken.

load dataset
error：
if images[0].shape[0] == 256: IndexError: list index out of range

Is the sh. broken?

It is a great work for your implementation~~
I think there is smoething going wrong in this code..
in ops line 98:
def residual(input, num_filters, name, is_train, reuse, norm, pad='REFLECT',
bias=False)
But in encoder line 57:
E = ops.residual(E, n, 'res{}_{}'.format(n, i + 1), self._is_train,
self._reuse, norm=self.norm, activation='leaky',
bias=True)
The anctivation layer in the reidual block is 'relu' in your ops. What is going on?...I feel confused~
Maybe the line 57 in encoder need to be : E = ops.residual(E, n, 'res{}{}'.format(n, i + 1), self._is_train,
self._reuse, norm=self._norm,
bias=True)

Hello,
In the original paper the authors state they keep the encoder fixed when performing the update step for cLR-GAN. However, in your code it seems that it is being updated. Are you aware of this?

Thanks

Hi,

The resnet block is incorrectly implemented the input is always 64 channel while the residual block channel size keeps changing.

If you are looking for pull requests, please let me know.