Question about training fastspeech2 with v1.yaml vs y2.yaml

[JunjunCui]

Hi~ Thanks for your contribution. It's a real meaningful work. I trained a fastspeech2 model on own dataset for Chinese language. I trained well on v2.yaml, but I couldn't converge good on v1.yaml，this is their loss trajectory. Would you mind suggesting me to improve the v1 fastspeech2 .

[dathudeptrai]

Hi~ Thanks for your contribution. It's a real meaningful work. I trained a fastspeech2 model on own dataset for Chinese language. I trained well on v2.yaml, but I couldn't converge good on v1.yaml，this is their loss trajectory. Would you mind suggesting me to improve the v1 fastspeech2 .

i'm not sure what is a problem but the v should be better than v2 :)). But if v2 is better in ur case, so let use v2 and just don't care about v1 :))

[JunjunCui]

Hi~ Thanks for your contribution. It's a real meaningful work. I trained a fastspeech2 model on own dataset for Chinese language. I trained well on v2.yaml, but I couldn't converge good on v1.yaml，this is their loss trajectory. Would you mind suggesting me to improve the v1 fastspeech2 .

i'm not sure what is a problem but the v should be better than v2 :)). But if v2 is better in ur case, so let use v2 and just don't care about v1 :))

Thanks for your reply. There is a little badcase in my results on V2, so I want to try V1, but the convergence is worse.

[dathudeptrai]

@JunjunCui try to use smaller initalize learning rate for v1.

[JunjunCui]

@JunjunCui try to use smaller initalize learning rate for v1.

Thanks! i will try it

[JunjunCui]

@dathudeptra I used a smaller initial learning rate on V1 and got a good convergence, thank you!!

[dathudeptrai]

@JunjunCui :D