# machine-learning
- m
mega b
09/23/2021, 9:09 PMwith tacotron2? - u
user
09/23/2021, 9:09 PMye - m
mega b
09/23/2021, 9:09 PM14 - u
user
09/23/2021, 9:10 PMk, thank u - u
user
09/23/2021, 9:17 PMalright, I've planned it, and am probably going to do a large flamingo voice model - u
user
09/23/2021, 9:17 PMthe final one - u
user
09/23/2021, 9:21 PMLuckily, there's plenty of data - u
user
09/23/2021, 9:21 PMlike, 3 years worth of possible data - b
beepbopbus
09/23/2021, 10:26 PMim gonna train a talknet model with 6 minutes of data (122 wavs) for 500 epochs - b
beepbopbus
09/23/2021, 10:26 PMcuz the 250 epochs one sounded robotic - b
beepbopbus
09/23/2021, 10:26 PMand bad - u
user
09/23/2021, 10:52 PMcuda ran out of memory - u
user
09/23/2021, 10:52 PMman colab hates me - u
user
09/23/2021, 10:53 PMI can't train even if I lower batch size - u
user
09/23/2021, 10:54 PMI'm forced to reduce the amount of voice clips which cause worse sounding models - u
user
09/23/2021, 10:54 PM - u
user
09/23/2021, 11:00 PMI found that it usually works at 150 voice clips - h
hecko
09/23/2021, 11:05 PMthat is extremely odd, total clip count should not matter at all maybe it's just one or a few particularly long clip(s) overfilling the vram - u
user
09/23/2021, 11:05 PMoh, probably ye - u
user
09/23/2021, 11:06 PMthat does make sense - u
user
09/23/2021, 11:46 PMkaggle automatically gives p100 I think so this is pog - f
F.B
09/23/2021, 11:47 PManother zuzu model be like - b
beepbopbus
09/23/2021, 11:47 PMit worked well :) - b
beepbopbus
09/23/2021, 11:47 PMit the loss kept going down which is good i think - c
Couch
09/23/2021, 11:58 PMyour model is forgerring - c
Couch
09/23/2021, 11:58 PM* * - f
F.B
09/23/2021, 11:59 PMbrub - c
ColdFir3
09/24/2021, 12:13 AMI mean f it here's the kaggle notebook guys, to use a GPU you must confirm your account with your number - c
ColdFir3
09/24/2021, 12:13 AMyour dataset must be a zip containing a folder called "wavs" with the wav files