tkazusa/cvae-gan

cvae-gan's Issues

y_fake_f & y_fake_p

Hi, thanks for your code for CVAE-GAN, it does help me a lot!
However, there is a question: what are the 'f' and 'p' in 'y_fake_f' and 'y_fake_p' (DisciminatorLossLayer()) means? I think the 'p' maybe means predict, but I have no idea about the 'f'.
Hope you will answer me, thanks again!

class DiscriminatorLossLayer(Layer):
    __name__ = 'discriminator_loss_layer'

    def __init__(self, **kwargs):
        self.is_placeholder = True
        super(DiscriminatorLossLayer, self).__init__(**kwargs)

    def lossfun(self, y_real, y_fake_f, y_fake_p):
        y_pos = K.ones_like(y_real)
        y_neg = K.zeros_like(y_real)
        loss_real = keras.metrics.binary_crossentropy(y_pos, y_real)
        loss_fake_f = keras.metrics.binary_crossentropy(y_neg, y_fake_f)
        loss_fake_p = keras.metrics.binary_crossentropy(y_neg, y_fake_p)
        return K.mean(loss_real + loss_fake_f + loss_fake_p)

    def call(self, inputs):
        y_real = inputs[0]
        y_fake_f = inputs[1]
        y_fake_p = inputs[2]
        loss = self.lossfun(y_real, y_fake_f, y_fake_p)
        self.add_loss(loss, inputs=inputs)

        return y_real

why return x value, not actual loss in loss functions?

Thank you for sharing this code. I'm just confused about the loss calculation classes. In each call function of them, instead of returning the calculated loss, you returned the input x values. Is there any specific reason for this? Or is there anything I misunderstood about Keras call functions?

class ClassifierLossLayer(Layer):
    __name__ = 'classifier_loss_layer'

    def __init__(self, **kwargs):
        self.is_placeholder = True
        super(ClassifierLossLayer, self).__init__(**kwargs)

    def lossfun(self, c_true, c_pred):
        return K.mean(keras.metrics.categorical_crossentropy(c_true, c_pred))

    def call(self, inputs):
        c_true = inputs[0]
        c_pred = inputs[1]
        loss = self.lossfun(c_true, c_pred)
        self.add_loss(loss, inputs=inputs)

        return c_true

ImportError: No module named 'data'

Hi,

I get an error when running train_CVAE-GAN.py
"
File "train_CVAE-GAN.py", line 13, in
from data.gravure import load_data
ImportError: No module named 'data'
"
Do you know how to fix it?

Python tensorflow Version

Hi,
Can you tell me the Python and Tensorflow version that you ran on?
Thank you very much!

load_data

Edit: Nevermind

Problems about GD and GC

In the original paper, the author use GD and GC to match the mean feature of xr and xp. However, I don't see any relations between xr and xp(xr is the real image and xp are generated images from random classes). Can you explain that to me? Thanks a lot!