-
https://git.ecker.tech/ aims to provide a place to share my efforts while maintaining true ownership of my code, as I do not trust GitHub.
XMR: 4B9TQdkAkBFYrbj5ztvTx89e5LpucPeTSPzemCihdDi9EBnx7btn8RDNZTBz2zihWsjMnDkzn5As1LU6gLv3KQy8BLsZ8SG
- Joined on
2022-10-10
Right, I forgot to freeze the gradio version, as I was getting errors out the ass on something entirely unrelated. My last known version was 3.23.0, so:
pip3 install -U gradio<=3.23.0
s…
I'll just do my weekly evaluation a little bit ahead of time.
I think the AR fried.
Despite the loss slowly going down, the range between the metrics are even more chaotic, and the…
Your batch size isn't evenly divisible by your gradient accumulation size. Stick to even numbers for both values.
Huh, funny. I believe it was the other day it crossed my mind to try and figure out that exact same thing: how DLAS would handle it's input prompt shuffling for better training. Being in the land…
Dreaded progress report:
- as noted with the dip at the end, that's from me removing LibriTTS's
train-other-500subset, and the…
Fixed in commit cbe21745df58ce220631ac34525e17b10c225c61. Apologies.
I noticed in the VALL-E implementation I did hotfix this with text = [ text ], and I suppose I neglected to copy that part…
Oh right, I forgot that instantiating the phonemizer backend yourself rather than using its API to instantiate one-use backends requires it to be by array rather than string. I'll see what I can…
You'll need to do something like (after activating the venv):
pip3 install -U numpy==1.23.5
I could have sworn I had the version frozen for this repo, but I can't for the life of me…
Oops, somehow I had that change make its way upstream. For my VALL-E training, I needed it to show by iteration step rather than epoch count, and my local change on my training machine somehow…
those metrics did look pretty bad
They still look pretty bad 1.5 epochs in, but it at least seems to be showing it's "learning" from the gradient norms getting smaller, and a random spike in…
Oh, I never actually touched those threads on /g/. I only shilled my original TorToiSe fork through a rentry on /v/ when the 11.AI threads there were on a downward spiral from the changes.
I'm…
a63bdbed33
Reverted a bunch of brainworm changes, kept the helpful ones (deleting densors explicitly)
Mmm... I think I fucked up the training script.
There's been no movement for the past few days, and I removed the train-other-500 dataset and, while the loss / accuracies moved, it still…
the audio quality is nowhere near Tortoise. It is also pretty unstable. It might be just me being stupid but I have my qualm that this model will never be like good old
That's pretty much how…
So what's the limiting factor in just using that 60k hour datase
In order:
- disk space
- not a big deal, I can move things around on my bunch of spinning rust, of transcribe in pieces,…
ef322110d3
-7 hours wasted because it crashed on eval and it kept crashed on eval a half hour after i went to bed