Advanced Training Page
Last updated
Last updated
This might be advanced for you, but if you don't want to go over the advanced tutorial, now is your chance to click this button:
One of the youtuber/uberduck contributer, "Adam is cool and stuff" has made colab notebook so we could use it for training.
You can check out his videos:
For now, let's do opening this colab page and see what we could do.
The next thing you'll need to do, and what you need is to copy to drive so that you could save some of your settings and stuff.
Next up, we'll have to check which GPU we have. Here's a list thing that supports GPU's:
P100
V100
T4
If you have K80 or P4 however, it's not gonna work. We need to change GPUs, BUT HOW?!?
Easy. just go to runtime and select factory reset:
If this window doesn't pop up:
That means it's not factory resetting and also didn't pop up asking are we sure we want to factory reset,
Try clearing cookies, disabling adblock, or maybe use a different browser.
Once you got GPU from the list above, we'll move on to the next step.
Running this cell asks you to log in to your google account and copy-pasting code in there. This will have access to your drive so that it could save your voice model into your drive (up to 300 MB)
Self-explanatory, it downloads tacotron.
It creates wavs folder, what you wanna do is click the folder icon:
Once you have clicked it you should see it on your right side:
What you'll do next is click tacotron2 and then you upload your finished audio files wavs in there:
and also upload your text file into filelists.
Here's the next thing you should do, run this first cell, this will load tacotron in. Name model into whatever you like and then click play on 2nd cell. On the 3rd cell, you have to rename PUTFILENAME into your text document name (list.txt) and click play.
Okay now here comes the hard part:
I'll try to explain as much as I can.
batch_size - i recommend putting 20. DON'T YOU EVER PUT TOO MUCH OR YOU'LL END UP OUT OF MEMORY.
epochs - This might be complicated, but it determines how many times ai will be trained. For example, I recommend putting 500. this AI will train 500 times, and then it will stop automatically. Try experimenting with it, I'm pretty much sure when you have some files like, maybe 100, then use 100 epochs. But correct me if I'm wrong.
Please note that if you put too much epoch, you either will get kicked out from google colab after 12 hours or you've just overtraining the model. So do so stop it manually after it's done.
Those other things are not important unless you are smart than that. But anyway, once you have done, run this cell.
Now run this first cell, this will generate mels (aka .npy files)
Then run in to check data, if it's checked like this:
It should be good to go.
Finally, run the third cell and let it train:
You may notice that graph kinda screwed up, but let it train and you should have it like this:
Once that done, do stop the training by clicking the stop button on the training cell. BUT DO IT AT ONCE OR YOU'LL SCREW UP!
If you have issues with it, please check the troubleshooting page
Congratulations, you have now done training the model. You are now ready to listen to what your model sounds like.
