actually accumulate derivatives when estimating milestones and final loss by using half of the log

oops v2

5 hour sleep brained

oops

reordered things so it uses fresh data and not last-updated data

added '''estimating''' iterations until milestones (lr=[1, 0.5, 0.1] and final lr, very, very inaccurate because it uses instantaneous delta lr, I'll need to do a riemann sum later

forgot to try and have it try and auto-detect for openai/whisper when no language is specified

added option to set worker size in training config generator (because the default is overkill), for whisper transcriptions, load a specialized language model if it exists (for now, only english), output transcription to web UI when done transcribing

sloppy fix to actually kill children when using multi-GPU distributed training, set GPU training count based on what CUDA exposes automatically so I don't have to keep setting it to 2

forgot to add 'bs / gradient accum < 2 clamp validation logic

renamed mega batch factor to an actual real term: gradient accumulation factor, fixed halting training not actually killing the training process and freeing up resources, some logic cleanup for gradient accumulation (so many brain worms and wrong assumptions from testing on low batch sizes) (read the training section in the wiki for more details)

fixed user inputted LR schedule not actually getting used (oops)

auto-suggested voice chunk size is based on the total duration of the voice files divided by 10 seconds, added setting to adjust the auto-suggested division factor (a really oddly worded one), because I'm sure people will OOM blindly generating without adjusting this slider

oops

validated whispercpp working, fixed args.listen not being saved due to brainworms

added option to specify using BigVGAN as the vocoder for mrq/tortoise-tts

oops

experimental multi-gpu training (Linux only, because I can't into batch files)

added button to just load a training set's loss information, added installing broncotc/bitsandbytes-rocm when running setup-rocm.sh

added loading/saving of voice latents by model hash, so no more needing to manually regenerate every time you change models

oops

leverage tensorboard to parse tb_logger files when starting training (it seems to give a nicer resolution of training data, need to see about reading it directly while training)

added new training tunable: loss_text_ce_loss weight, added option to specify source model in case you want to finetune a finetuned model (for example, train a Japanese finetune on a large dataset, then finetune for a specific voice, need to truly validate if it produces usable output), some bug fixes that came up for some reason now and not earlier

oops

added to embedded metadata: datetime, model path, model hash

show different losses, rewordings

oops

added graph to chart loss_gpt_total rate, added option to prune X number of previous models/states, something else

I don't remember.

Added option to disable bitsandbytesoptimizations for systems that do not support it (systems without a Turing-onward Nvidia card), saves use of float16 and bitsandbytes for training into the config json

actually fixed the training output text parser

oops, epoch increments twice

csome adjustments to the training output parser, now updates per iteration for really large batches (like the one I'm doing for a dataset size of 19420)

fixed some files not copying for bitsandbytes (I was wrong to assume it copied folders too), fixed stopping generating and training, some other thing that I forgot since it's been slowly worked on in my small free times

ops

finally added model refresh button, also searches in the training folder for outputted models so you don't even need to copy them

god i finally found some time and focus: reworded print/save freq per epoch => print/save freq (in epochs), added import config button to reread the last used settings (will check for the output folder's configs first, then the generated ones) and auto-grab the last resume state (if available), some other cleanups i genuinely don't remember what I did when I spaced out for 20 minutes

forgot to mult by batch size

fixed the brain worm discrepancy between epochs, iterations, and steps

oops

huge success

how did this revert...

oops

properly placed the line toe xtract starting iteration

oops

clamp batch size to sample count when generating for the sickos that want that, added setting to remove non-final output after a generation, something else I forgot already

explicitly use python3 because some OSs will not have python alias to python3, allow batch size 1

Added very experimental float16 training for cards with not enough VRAM (10GiB and below, maybe) \!NOTE\! this is VERY EXPERIMETNAL, I have zero free time to validate it right now, I'll do it later

wrap checking for metadata when adding a voice in case it throws an error

fixes