# machine-learning
- h
hecko
11/18/2021, 5:44 AMwow that kinda,, sucks - h
hecko
11/18/2021, 5:44 AMlike the voice itself is nice but the prosody is awful - u
user
11/18/2021, 8:37 AM@User is it nomal that validation loss only shows when there is a checkpoint? - u
user
11/18/2021, 8:43 AMIt only shows the validation loss when it reaches a checkpoint - u
user
11/18/2021, 8:43 AMAnd that's the only time it updates on the tensorboard - b
burritohead
11/18/2021, 11:34 AMIs there a guide that spells out how to train locally? Basically all the stuff you all went through with sked - u
user
11/18/2021, 11:46 AMUnfortunately no and its fucking pain - u
user
11/18/2021, 11:46 AMThere's only like a 50% chance of it working first try - m
mega b
11/18/2021, 3:55 PMI think it's aiming for voice accuracy - u
user
11/18/2021, 4:07 PMi have 38 samples - h
hecko
11/18/2021, 4:08 PMcheck your transcripts - h
hecko
11/18/2021, 4:09 PMit's possible that they're misaligned somewhere - h
hecko
11/18/2021, 4:09 PMalso i think batch size 30 means the last 8 will be ignored because they can't form a complete batch, not the main issue here but just so you know - u
user
11/18/2021, 4:10 PMwait so what batch size do i need - u
user
11/18/2021, 4:10 PMdefinitely not i jsut checked - m
matty1406
11/18/2021, 4:10 PMhow many audio files do you have? - h
hecko
11/18/2021, 4:11 PM38 or 19 i think? whichever fits in vram - m
matty1406
11/18/2021, 4:12 PMso what hecko said - h
hecko
11/18/2021, 4:33 PMvery odd - h
hecko
11/18/2021, 4:34 PMmaybe it's training from nothing?? - h
hecko
11/18/2021, 4:34 PMas in not the ljspeech base model - u
user
11/18/2021, 4:35 PMit definitely is - u
user
11/18/2021, 4:35 PMits using that mnodel - u
user
11/18/2021, 4:35 PMpython train.py --output_directory=outdir --log_directory=logdir --warm_start -c tacotron2_statedict.pt --hparams=training_files=filelists/list.txt,validation_files=filelists/list.txt,batch_size=26,learning_rate=1e-5,epochs=1000,iters_per_checkpoint=50 - u
user
11/18/2021, 4:35 PMis this command good - u
user
11/18/2021, 4:36 PMepochs is only 1000 because i dont want it to stop - m
mega b
11/18/2021, 4:59 PMtry arpabet transcriptions - j
Justin
11/18/2021, 5:50 PM - h
HoodyP
11/18/2021, 6:33 PMActually, I did train on my RTX 3050 before... So, I think yours could work... I actually trained it on Anaconda/Miniconda with a virtual enviroment for Python 2.7 - u
user
11/18/2021, 6:34 PMHow did you do it?