cGAN with multi-class labels about keras-gan HOT 5 CLOSED

Oh, I see what you mean. Hm. I haven't tried this but I think you can do:

        noise = Input(shape=(self.latent_dim,))
        label = Input(shape=(self.label_dim,), dtype='int32')

        label_embedding = Embedding(self.num_classes, self.latent_dim, input_length=self.label_dim)(label) # output shape (batch_size, self.label_dim, self.latent_dim)

        model_input = multiply([noise, label_embedding]) # output shape (batch_size, self.label_dim, self.latent_dim)
        model_input = Flatten()(model_input) # output shape (batch_size, self.latent_dim * self.label_dim)

Then you would have to set model.add(Dense(256, input_dim=self.latent_dim)) to model.add(Dense(256, input_dim=self.latent_dim * self.label_dim)) on line 60.

Instead of flattening the tensor of shape (batch_size, self.label_dim, self.latent_dim) it should also be possible to multiply along dimension 1, to get a tensor of shape (batch_size, self.latent_dim):

        ...
        model_input = dot(model_input, axes=1) # output shape (batch_size, self.latent_dim)

I would try both ways and see which yields the best results.

from keras-gan.

Sure, in the current implementation of cGAN the generator and discriminator expects a label input of shape (x,), where x is a integer value between 0 and self.num_classes - 1. So for MNIST this integer value will be between 0 and 9. But this can be adapted for any classification task.

from keras-gan.

The problem is label shape is hardcoded as label = Input(shape=(1,), dtype='int32') so trying to use any shape greater than 1 will cause an error during the model_input = multiply([noise, label_embedding]) operation. The error will be something like ValueError: Operands could not be broadcast together with shapes (x,) (y,)

from keras-gan.

Ah yes that appears to work. Then for the discriminator network it's the same code just swap self.latent_dim with np.prod(self.img_shape) and rework the training loop as necessary and it's good to go.

from keras-gan.

Is it common to use Embedding for CGAN? why not just append label to flattened image?

Related:
#143

from keras-gan.