leafinity / keras_ganomaly Goto Github PK

hi, I have a question need your help.

In your demo code

why np.expand_dims(new_x, axis=-1) / 127 - 1 in function reshape_x

please help me

thanks

def get_data_generator(data, batch_size=32):
    datalen = len(data)
    cnt = 0
    while True:
        idxes = np.arange(datalen)
        np.random.shuffle(idxes)
        cnt += 1
        for i in range(int(np.ceil(datalen/batch_size))):
            train_x = np.take(data, idxes[i*batch_size: (i+1) * batch_size], axis=0)
            y = np.ones(len(train_x))
            yield train_x, [y, y, y]

Why are the y ones instead of zeros? The loss have to close to zero, right?
Do I misunderstand something?

hi, i have some question on advloss

in your code, you use K.mean to calculating the average value of gan feature before it subtract from ori feature.
like this
"K.mean(K.square(ori_feature - K.mean(gan_feature, axis=0)))"

Is it necessary for calculating the average value of gan feature?
Can i subtract gan feature from ori feature without doing it?
like this
"K.mean(K.square(ori_feature - gan_feature))"

thanks

Hello!
I am trying to implement ganomaly network on different kind of data but I get error: 'int' object has no attribute 'mean'. problem raises in loss functions when we want to use keras backend as K and take mean over differences. As I know it may happen because of backend either mean is taking an int as input. But I am using same shape for my data as MNIST data to train generator. I don't know what is going wrong.
Please help me.
Thanks