r0nn13/conditional-dcgan-keras

Keras implementation of the conditional GAN.

conditional-dcgan-keras's Issues

How to change the image sizes?

Hi,

Thanks for sharing! and I have two questions,
-I have 2 classes and for each one 2k images with same size 250x250, I wont to resize them but I get error, How come over with this???

• I would like to generator just give me one images with the size not concatinating the small size in one image, How can I do that?

Thanks in Advance,
Maryam

Error in mergin

Hi,
I've got an error related to the merge part of generative network:
"
ValueError: "concat" mode can only merge layers with matching output shapes except for the concat axis. Layer shapes: [(None, 1024, 2, 512), (None, 512, 2, 2)]
"
Is a bug in you code?
beside this I've got an error for using "mode" in "BatchNormalization"

BatchNormalization(mode=2) on keras 2.1.5

In the sampled code for the keras implementation in the generator there is the batchnormalization step that it is coded as follow:

e1 = BatchNormalization(mode=2)(e1)

However, when the mode = 2 then there is an emerging error from keras. I noticed that when i set the mode=0 the whole thing is working. However, what it the difference between mode=2 and mode =0. And if it is crucial difference how can I have mode = 2. I guess that error has to do with the keras version (I am using keras 2.1.5).

Add class information of the images into the loss function

Is it possible to take into account as a conditional information also the actual labels of the image 2 image translation? As I understood my input image Xa from the a belongs to A domain is transformed and mapped to Xb where b belongs to B domain and also Xb belongs to X_train while the Xa belongs to y_train. Is it possible to add the actual classes of the X_a and X_b into the loss function of the discriminator?

Discriminator loss function

I am trying to understand the loss function of discriminator:

  `def discriminator_loss(y_true,y_pred):
              return K.mean(K.binary_crossentropy(K.flatten(y_pred), K.concatenate([K.ones_like(K.flatten(y_pred[:BATCH_SIZE,:,:,:])),K.zeros_like(K.flatten(y_pred[:BATCH_SIZE,:,: :])) ]) ), axis=-1)`

I am wandering why the y_true is not used at all and only y_pred is used twice. Is it kind of mistake? They way that the discriminator is trained is:

 `# Training D:
 real_pairs = np.concatenate((X_train[index * BATCH_SIZE:(index + 1) * BATCH_SIZE, :, :, :], image_batch),axis=1)
 fake_pairs = np.concatenate((X_train[index * BATCH_SIZE:(index + 1) * BATCH_SIZE, :, :, :], generated_images), axis=1)

  X = np.concatenate((real_pairs, fake_pairs))
  y = np.concatenate((np.ones((BATCH_SIZE, 1, 64, 64)), np.zeros((BATCH_SIZE, 1, 64, 64))))
  d_loss = discriminator.train_on_batch(X, y)`

Is it that I need to add also the y_true in the discriminator or no?

Error in merging layers

Hello, first I was using keras 2.0. Then i realized thousands of things have changed in 2.0. So I downgrade keras to 1.0.8. Now I am getting the following error.

Exception: "concat" mode can only merge layers with matching output shapes except for the concat axis. Layer shapes: [(None, 512, 1, 1), (None, 512, 2, 2)]

Is there any problem in your code? Or is it some kind of version problem?

All zero targets?

https://github.com/r0nn13/conditional-dcgan-keras/blob/master/conditional_gan.py#L193

This creates an array of all zeros. Surely half should be zero and the other half ones?

Implementation using as a keras backend tensorflow

Is it possible to run the code using keras-tensorflow? If so what should be the arguments in that case?