I need that.

Detective Cromwell looking good.

Superlit bro

Stonks

From Killer Bean.

Extactly.

I could see that for some architectures, but for the original question of > Is it possible to train a TacoTron2 model with DeepSpeed? The answer is yes, it can be done but it would be worthless for 99% of people or just outright make everything slower. It's a 99% Autoregressive architecture without any feed-forward style training so you shouldn't run segments of the model on their own or communication times would make the segments entirely unusable with tacotron2.

Congratz!!! 🙌 🥳

This is what happens if I turn off raw input.

raw input on.

This is bugging me now. There's no way I can make my very first Talknet model. Just like I said, I'm only into making Tacotron 2 models.

I'm new to this, does this look like a promising model so far?

You can do it! It's not that complicated

Damn, my model turned out really bad, just stuttering. This happened last time and I made some adjustments and it was a bit better, but not enough

Not sure what went wrong

He kinda grunts too

Warning! Reached max decoder steps

what does that mean

It's okay if it's fine for you, but guess what? I tried it, but I always get the stupid error that says I need to "iNsTaLl DePeNdEnCiEs UsInG eItHeR !pIp Or !ApT"...

That's why making Tacotron 2 models (and YTP's too) is my thing. And besides, I don't make Talknet models, I request my datasets to people who can make Talknet versions of it.

(open original)

using updated dall-e mini model on huggingface

from 32 images

are you using the fixed notebook? https://colab.research.google.com/drive/1KSeOCgyFXTYgVIvvPhOMVXlLvMXwiQMM?usp=sharing

Ooh, forgot about that one. I guess it might work better than the one I just used?

it's specifically made because the old one got broken

It means the model never reached the end of the input text and had to abort before it filled the graphics card with audio

Ok, thanks, I guess that means I need to take a look at redoing model and making sure my transcripts and clips are good