MIG and SAP and DCI

[amasawa]

May I ask how to design the MIG, SAP, and DCI in your paper? The sequence data do not have the ground truth factor.

[stathius]

The data we have used have used in the paper have the ground truth. If you create the data with the scripts we provided, that should be the case as well.

…

On Sat, 16 Dec 2023 at 10:45, BerenWu ***@***.***> wrote: May I ask how to design the MIG, SAP, and DCI in your paper? The sequence data do not have the ground truth. — Reply to this email directly, view it on GitHub <#1>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAPQ4BIW25XFVKXYBR6AAS3YJV3VDAVCNFSM6AAAAABAXQBP6OVHI2DSMVQWIX3LMV43ASLTON2WKOZSGA2DINZQGMYTQMQ> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[amasawa]

Thank you for your quick response. May I inquire about details about the quantitative disentanglement in these types (SAP MIG DCI)? I couldn't find them in the code repository. Taking MIG as an example, since betaVAE, different disentangled models seem to have variations in their estimation of MIG. For instance:

• controlVAE
• gcae.

[stathius]

I used disent for computing metrics. The only change needed for MIG was to replace the discrete entropy with the continuous one.