nuclearboy95 / anomaly-detection-deep-svdd-tensorflow Goto Github PK

Hi,
Thanks for your implementaion of the deep svdd. Could you show us the average auc score of your implementation on cifar10. I think it's also very important for others to reuse your work.

Regards,
fsk

Hi,
Thanks for your implementaion of the deep svdd.
I have a question for implementation of the Deep-SVDD loss function in this code.

I think the "class DeepSVDD (line29-line36)" in 'dsvdd.deepSVDD' denote the two diiferent loss function for Deep-SVDD.

        with task('Build graph'):
            self.x = tf.placeholder(tf.float32, [None] + list(input_shape))
            self.latent_op = self.keras_model(self.x)
            self.dist_op = tf.reduce_sum(tf.square(self.latent_op - self.c), axis=-1)
            

            if self.objective == 'soft-boundary':
                self.score_op = self.dist_op - self.R ** 2
                penalty = tf.maximum(self.score_op, tf.zeros_like(self.score_op))
                self.loss_op = self.R ** 2 + (1 / self.nu) * penalty

            else:  # one-class
                self.score_op = self.dist_op
                self.loss_op = self.score_op

            opt = tf.train.AdamOptimizer(lr)
            self.train_op = opt.minimize(self.loss_op)

but this function's objective function looks different from the paper's equation (3) and equation (4) which is missing the network weight decay regularizer.
Is it a right implementation for eq(3) and eq(4)? Or "the network weight decay is exist at other part of the code which
was ignored by me.

Thanks again!
Sincerely!

Hi,

I would like to test your algorithm for anomaly detection.Could you define the arguments needed by DeepSVDD, especially "input_shape" and "representation_dim" so that I can pass adequate arguments.

Regards

Extract from deepSVDD.py :

class DeepSVDD:
def init(self, keras_model, input_shape=(28, 28, 1), objective='one-class',
nu=0.1, representation_dim=32, batch_size=128, lr=1e-3):