7bf1dc45f850b8ae1b5a1dd4f475f8b6-Supplemental-Conference.pdf

In this appendix, we provide pseudo-code algorithms explaining how to build the metric from a29 trained VAEandhowtousetheproposed sampling process. B.2.1 TheHMCsampler43 In the sampling process we propose to rely on the Hamiltonian Monte Carlo sampler to sample44 fromtheRiemanian uniformdistribution. Moreover,sinceG(z)issmooth andhas66 a closed form, it can be differentiated with respect toz pretty easily. Figure 5: Closest element inthetraining set(Near.) Each model is trained on each label of the train set and used to generate 2k samples per89 class.