why autoregressive doesn't need revertible? about flowtron HOT 3 CLOSED

Thank you, I got it later! Did you see the glow-TTS and Flow-TTS two papers？ They are non-AR and around 20 times faster than Tacotron2. I wander the speed of Flowtron, how about it? Thank you~~ 发件人: Rafael Valle <[email protected]> 答复: NVIDIA/flowtron <[email protected]> 日期: 2020年5月25日 星期一 00:47 收件人: NVIDIA/flowtron <[email protected]> 抄送: "zewangzhang(张泽旺)" <[email protected]>, Author <[email protected]> 主题: Re: [NVIDIA/flowtron] why autoregressive doesn't need revertible? (#21)(Internet mail) From the paper between Eqs. 8 and 9. Here NN() can be any autoregressive causal transformation. This can be achieved by time-wise concatenation of a 0-valued vector to the input provided to NN(). The affine coupling layer preserves invertibility for the overall network, even though NN() does not need to be invertible. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub<#21 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/APPCP3ECM2RUZ2ZUZWAX3G3RTFFRJANCNFSM4NIZPC4Q>.