p2333 / max-mahalanobis-training Goto Github PK

Hi, I have a question in the advtest_other.py

row 283: adv_x = att.generate(x_place, is_MMC=is_MMC, **att_params)

But the att is from the cleverhans library. the official API doesn't have this argument. So I think you modify the source code of cleverhans a bit.

Can you upload the modified source file? I hope it's just one file.

In your paper, you have very dense clusters in your t-SNE visualization plot.

I can generate very dense logits in 10-dimension to the true center.
the features/outputs after the LDA like [100+, 100+, 100+, 0.1-, 100+, 100+, ...]. It's very close to the center in 10d.
But when I apply t-SNE (with sklearn library), it generates something like sticks/bands.

So any suggestion to generate the t-SNE visualization better as yours?

#Load centers in MMC
kernel_dict = loadmat('kernel_paras/meanvar1_featuredim'+str(feature_dim)+'_class'+str(num_class)+'.mat')
mean_logits_np = kernel_dict['mean_logits'] #num_class X num_dense
mean_logits = FLAGS.mean_var * tf.constant(mean_logits_np,dtype=tf.float32)

can give some code implementation?