(try to) unload voicefixer if it raises an error during loading voicefixer

cleanup block that makes embedding latents for random/microphone happen, remove builtin voice options from voice list to avoid duplicates

maybe remedy an issue that crops up if you have a non-wav and non-json file in a results folder (assuming)

when you have three separate machines running and you test one one, but you accidentally revert changes because you then test on another

I hate commas make tuples

you might be wondering why so many small commits instead of rolling the HEAD back one to just combine them, i don't want to force push and roll back the paperspace i'm testing in

typo

forgot to save the simplified training input json first before touching any of the settings that dump to the yaml

added VRAM validation for a given batch:gradient accumulation size ratio (based emprically off of 6GiB, 16GiB, and 16x2GiB, would be nice to have more data on what's safe)

oops

forgot template

big cleanup to make my life easier when i add more parameters

oops

actually make using adamw_zero optimizer for multi-gpus work

disable validation if validation dataset not found, clamp validation batch size to validation dataset size instead of simply reusing batch size, switch to adamw_zero optimizier when training with multi-gpus (because the yaml comment said to and I think it might be why I'm absolutely having garbage luck training this japanese dataset)

lazy wrap around the voicefixer block because sometimes it just an heros itself despite having a specific block to load it beforehand

fixed notebooks, provided paperspace notebook

made validation working (will document later)

oops

set validation to save rate and validation file if exists (need to test later)

added helper script to cull short enough lines from training set as a validation set (if it yields good results doing validation during training, i'll add it to the web ui)

fixed update checker for dlas+tortoise-tts

added override for large training plots

set NaN alarm to show the iteration it happened it

added update checking for dlas and tortoise-tts, caching voices (for a given model and voice name) so random latents will remain the same

fixes

brain dead

fried my brain trying to nail out bugs involving using solely ar model=auto

cleaned up some model loading logic, added 'auto' mode for AR model (deduced by current voice)

added (yet another) experimental voice latent calculation mode (when chunk size is 0 and theres a dataset generated, itll leverage it by padding to a common size then computing them, should help avoid splitting mid-phoneme)

brain worms and headaches

download dvae.pth for the people who managed to somehow put the web UI into a state where it never initializes TTS at all somehow

moved (actually not working) setting to use BigVGAN to a dropdown to select between vocoders (for when slotting in future ones), and ability to load a new vocoder while TTS is loaded

reworked generating metadata to embed, should now store overrided settings

added settings editing (will add a guide on what to do later, and an example)

forgot to re-append the existing transcription when skipping existing (have to go back again and do the first 10% of my giant dataset

Im not too sure if manually invoking gc actually closes all the open files from whisperx (or ROCm), but it seems to have gone away longside setting 'ulimit -Sn' to half the output of 'ulimit -Hn'

and it turned out I wasn't even using the aligned segments, kmsing now that I have to *redo* my dataset again

Added option to skip transcribing if it exists in the output text file, because apparently whisperx will throw a "max files opened" error when using ROCm because it does not close some file descriptors if you're batch-transcribing or something, so poor little me, who's retranscribing his japanese dataset for the 305823042th time woke up to it partially done i am so mad I have to wait another few hours for it to continue when I was hoping to wake up to it done

added api_name for generation, added whisperx backend, relocated use whispercpp option to whisper backend list

stretch loss plot to target iteration just so its not so misleading with the scale

UI cleanup, actually fix syncing the epoch counter (i hope), setting auto-suggest voice chunk size whatever to 0 will just split based on the average duration length, signal when a NaN info value is detected (there's some safeties in the training, but it will inevitably fuck the model)

(should) fix reported epoch metric desyncing from defacto metric, fixed finding next milestone from wrong sign because of 2AM brain

brianworms

typo

oops

print in e-notation for LR

remove redundant stats, add showing LR

forgot to remove a debug print

whispercpp actually works now (language loading was weird, slicing needed to divide time by 100), transcribing audio checks for silence and discards them