# machine-learning
- o
OccultMC
09/18/2022, 12:41 PMnvm cant find it - o
Odcub Rodriguez
09/18/2022, 1:26 PMHow do I use Automatic Transcribing (compatible with GPU and CPU)? - o
OccultMC
09/18/2022, 1:29 PMread the instructions on the colab and make sure Runtime is set on GPU - o
OccultMC
09/18/2022, 1:29 PMRuntime > change runtime > gpu - o
OccultMC
09/19/2022, 5:23 AM - o
OccultMC
09/19/2022, 5:23 AMbeat that 😎 - p
PixPrucer
09/19/2022, 6:25 AMThis looks like a synthesizer - a
a nickname?
09/19/2022, 6:25 AMWow - p
PixPrucer
09/19/2022, 6:25 AMI wonder what did you train to get the attention graph to look so. Dead - j
Justin
09/19/2022, 6:26 AMhaters gonna say it's photoshop - a
a nickname?
09/19/2022, 6:28 AMHow? - a
a nickname?
09/19/2022, 6:28 AMIt kinda looks legit to me - o
OccultMC
09/19/2022, 6:35 AM😎 - c
cadeotsg
09/19/2022, 7:14 AMsex - p
PixPrucer
09/19/2022, 7:14 AMThis is threatening - a
a nickname?
09/19/2022, 7:14 AM??? - p
PixPrucer
09/19/2022, 7:15 AMoh no I just realised the joke - p
PixPrucer
09/19/2022, 7:15 AMYou are so evil for that - a
a nickname?
09/19/2022, 7:15 AMAlso Justin just saying this cause it’s true you jacoblenstar waltman13 dawn de era and scruffygamer make awesome models - o
OccultMC
09/19/2022, 9:41 AMDoes it get any better than this? 🥺 - p
PixPrucer
09/19/2022, 10:33 AMI need to further fine-tune the HiFi-GAN model but it's working 🐉 - m
mepc36
09/19/2022, 1:46 PMCan anyone tell me what the metrics in the below log line mean?iter: 587 (0.76 s) | lr: 0.0005 | loss_mel: -1.775 | loss_prior_mel: 0.482 | loss_ctc: 2.670 | loss_duration: 0.317 | loss_f0: 0.001 | loss_energy: 0.001 | loss_vpred: 0.599 | binarization_loss: 0.541
is iteration anditer
is learn rate, but I don't know what loss_mel is.lr - h
hecko
09/19/2022, 2:01 PMloss_mel is how different the mel spectrogram (internal version of audio) is from the training data - m
mepc36
09/19/2022, 2:02 PMso I'd want that number to be as low as possible, right? - h
hecko
09/19/2022, 2:02 PMyeah - h
hecko
09/19/2022, 2:02 PMthough it's odd that it's negative - u
{K EY1} (Kei)
09/19/2022, 2:02 PMIn theory, yes But in practice, if it's too low it'll overfit, which is where the model can only produce the training data - m
mepc36
09/19/2022, 2:03 PMDoes positive/negative indicate overfit/underfit, respectively? - u
{K EY1} (Kei)
09/19/2022, 2:03 PMNo - m
mepc36
09/19/2022, 2:03 PMthanks for the help you guys!