Fix relative cav

[monz]

• the paper states in section 3.6 for relative tcav, construct a positive and negative set, e.g., for concepts P_dot, P_stripe, P_mesh this would look like, P_stripe (positive) and {P_dot U p_mesh} (negative). However, the code was giving sets like, P_stripe (pos), P_dot (neg). Now, the complement set is used for relative tcav as described in the paper.
• when computing relative tcav, plot function failed because concepts are also 'random' concepts and is_random_concept(concept) is always true, therefore. Now, plot for relative tcav handles 'random' concepts correctly.

Hope I did understand the paper correctly and gave a proper fix.

[BeenKim]

Thank you for submitting this RP @monz ! Two things if you are able 1) mind changing the relative_cav variable to be is_relative_cav (to avoid confusion that the variable IS the cav) 2) this is big enough change that I feel a few tests could be very helpful. Would you mind adding a few simple ones to our test deck? Thank you! been

[monz]

Hi @BeenKim, please have a look to the added tests, hope this is what you expected, kind regards, Markus.

[monz]

Hi @BeenKim I know its kind of a big PR, however, I am curious what you think about it. Kind Regards, Markus.

[BeenKim]

Hi Markus!

Thank you for incorporating our request! I'm excited to review this, but having trouble finding time to test the new changes end-to-end. Hopeful we'll get to this soon. Cheers Been

On Fri, Jun 11, 2021 at 12:19 AM Markus Monz @.***> wrote:

Hi @BeenKim https://github.com/BeenKim I know its kind of a big PR, however, I am curious what you think about it. Kind Regards, Markus.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/tensorflow/tcav/pull/111#issuecomment-859342328, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAP35E35ABRBEIG4H6FP5U3TSG2GHANCNFSM44NZHAOA .