Why is my experimental cifar10 result far worse than your paper shows?

[jianyuheng]

Hi, Sir! Thanks for your excellent idea! I download your code and plan to recreate your experiment results. But when i set the parameters with 'sh scripts/cifar10_svdd.sh gpu cifar 0 adam 0.0001 150 0.1 1 1 0 exp 3 1 -1 ' , I got a bad result Train objective: 1.69638 Train accuracy: 90.02% Val objective: 1.75854 Val accuracy: 88.90% Test objective: 3.78288 Test accuracy: 31.02% Test AUC: 60.71%

I think that something wrong happened, but I dont know how to improve the result. I am looking forward to your reply!

[lukasruff]

Hi jyhengcoder,

thanks for your interest and the kind words.

My guess is that this is the result without AE pretraining, i.e. Deep SVDD network weight initialization from the encoder weights of a corresponding autoencoder?

You can use the in_name argument of the cifar10_svdd.sh script to load the encoder weights from an autoencoder.

For example, first run sh scripts/cifar10_cae.sh gpu cifar_cae 0 adam 0.0001 350 3 1 -1 to train the convolutional autoencoder we used for pretraining in our paper, and then run sh scripts/cifar10_svdd.sh gpu cifar 0 adam 0.0001 150 0.1 1 1 0 cifar_cae/weights_final 3 1 -1 to train the OC Deep SVDD model.

Let me know if you were able to replicate our results.