378a063b8fdb1db941e34f4bde584c7d-AuthorFeedback.pdf

EBMs are faster to train than other likelihood models, with fewer parameters, but are more expensive3 thanGANbasedmodels(duetoLangevindynamicssampling),andslowertosample. ForEBM, sampling time depends on steps of sampling. Wefurther plan torelease13 source code for every section to further clarify experi-14 ments,aswellaspretrainedmodels.15 Wehaveaddthistothepaper.34 Mode Evaluation (R3) Following R3, we evaluated the number of modes under the stacked MNIST task.