kakoedlinnoeslovo / center_loss Goto Github PK

Hi, I have a question:
def call(self, x, mask=None): ## x[0] is Nx2, x[1] is Nx10 onehot, self.centers is 10x2 delta_centers = K.dot(K.transpose(x[1]), (K.dot(x[1], self.centers) - x[0])) # 10x2 denominator = K.sum(K.transpose(x[1]), axis=1, keepdims=True) + 1 delta_centers /= denominator new_centers = self.centers - self.alpha_center * delta_centers self.add_update((self.centers, new_centers), x) self.result = (K.dot(x[1], self.centers) - x[0]) self.result = K.sum(self.result ** 2, axis=1, keepdims=True) return self.result
the call function has already contained the center loss. But in the below:
self.model.compile(optimizer=optimizer, loss=[losses.categorical_crossentropy, self.center_loss], loss_weights=[1, self.lambda_centerloss])

the compile get center_loss again and again, is that true?

Hi, thanks for your implement, but I have some problem about this, hope you can help me.
First, why it said "Train on 18000 samples, validate on 10000 samples", because I think we input 60000 samples for training?
Second, If we want to know the accuracy, we should only concern the val_dense_2_acc or we should concern both val_dense_2_acc and val_centerlosslayer_acc? And why val_centerlosslayer_acc always 1.0?

Hi Kakoedlinnoeslovo,

Thanks for the code. I am trying to implement center-loss in Tensorflow 2.0 and found that some arguments in self.add_update are deprecated. It does not throw any error, but the centers are never updated.
You can find the description here:
https://github.com/tensorflow/tensorflow/blob/v2.1.0/tensorflow/python/keras/engine/base_layer.py#L1170-L1259

Can you provide a work around for this?

Thank you