# tacotron-2-support
- m
mega b
09/15/2022, 1:32 AMYes - u
{K EY1} (Kei)
09/15/2022, 2:35 AMDamn What happened wrong during training? - h
haru0l
09/15/2022, 10:21 AMprobably model somehow didnt save properly when ending training - u
{K EY1} (Kei)
09/15/2022, 1:42 PMDamn - m
mepc36
09/15/2022, 1:48 PMSomeone mind helping me understand what these "loss" logs mean for the accuracy/quality of the model? It's the output from a radtts model training. As in, which of these do I care about? Do these numbers indicate a bad/good model?Copy codeiteration: 464 (3.30 s) | lr: 0.001 | loss_mel: -1.067 | loss_prior_mel: 0.517 | loss_ctc: 3.740 | loss_duration: 0.286 | loss_f0: 0.005 | loss_energy: 0.005 | loss_vpred: 1.372 | binarization_loss: 0.791 - m
mepc36
09/15/2022, 1:48 PMIts from iteration #464 in epoch #1 so it's early yet - u
{K EY1} (Kei)
09/15/2022, 1:52 PMIt all looks normal to me, but i'm not sure what loss to look at in order to know when to stop I'm not experienced with radtts training - m
mepc36
09/15/2022, 2:08 PMThats my question too, here's the repo: https://github.com/NVIDIA/radtts - m
mepc36
09/15/2022, 2:08 PMSo is there no universal measurement of loss (like 5, or 0, or negative 1 million), at which we can say, "This model is a good model?" - m
mepc36
09/15/2022, 2:08 PMI read this article but it didn't answer those questions: https://towardsdatascience.com/understanding-loss-functions-the-smart-way-904266e9393 - u
{K EY1} (Kei)
09/15/2022, 2:16 PMYeah i'm not sure - u
{K EY1} (Kei)
09/15/2022, 2:16 PMAre you running tensorboard? - u
{K EY1} (Kei)
09/15/2022, 2:17 PMYou can maybe check the graphs, or audio inference if there is any - m
mepc36
09/15/2022, 3:07 PMNot sure, ill check... I guess I'm really just looking for a standardized tool that will tell me how good my model is, any ideas? - m
mepc36
09/15/2022, 3:08 PMLike thes graphs that the tacotron2 notebook in #841437191073955920 publishes during training: https://colab.research.google.com/drive/1CIPTTj94EocZe2w5-zCDC3G44K9OaFal - o
Odcub Rodriguez
09/16/2022, 12:09 AMshould I request a model? or.. what? - u
(Dawn) Will Draw Fictional Women
09/16/2022, 12:14 AMno? - o
Odcub Rodriguez
09/16/2022, 12:19 AMwhy no with a question mark? - o
OccultMC
09/16/2022, 2:35 AM"He found his art never progressed when he literally used his sweat and tears." - o
OccultMC
09/16/2022, 2:35 AMwhat are some good tips with scaling a model - o
OccultMC
09/16/2022, 2:35 AMor some tips on making the voice more smoother - p
PUMPKINEATER
09/16/2022, 3:35 AMThat is TortoiseTTS? - u
(Dawn) Will Draw Fictional Women
09/16/2022, 3:36 AMwhy would he post tortoise clips in fucking tacotron support - p
PUMPKINEATER
09/16/2022, 3:37 AMIdk it just sounds like tortoisetts - t
tylerdurdenceketi
09/16/2022, 8:35 AMHello, I get "ValueError: Expected more than 1 value per channel when training, got input size torch.Size([1, 512, 1])" error Can you help me? - m
mepc36
09/16/2022, 1:59 PMAnyone know if the legacy tacotron2 notebook in notebooks is trained from a warmstart or not? This one: https://colab.research.google.com/drive/1CIPTTj94EocZe2w5-zCDC3G44K9OaFal - m
mepc36
09/16/2022, 2:03 PMThe answer is "no", according to line 615 in the .py you get from downloading that notebook in .py format - u
{K EY1} (Kei)
09/16/2022, 2:04 PMIt looks to me like it's warm started It has a warm start argument when you start the training step - m
mepc36
09/16/2022, 2:07 PMyeah all the vars are in global scope though, so I'm 99% sure that the invocation of
in the training step is just passing in that warm_start var whose value istrain(..., warmstart_arg_here)
. It didn't get overwritten in between the var definition and the train function call I don't thinkFalse - u
{K EY1} (Kei)
09/16/2022, 2:08 PMWell, if it is warm starting you'll see a message that says "warm starting from checkpoint {whatever}"