Train on custom input -> noise as output about dalle2-pytorch HOT 7 CLOSED

Yeah no problem.
I have nearly 200k images to train from.

Because of that size it took nearly 24h for 3 epochs on a A100 GPU.

It took maybe 5-8 epochs each to see some quite good results on that low resolution.
If you want a batter resolution you maybe need to add a second UNet to enhance it.

from dalle2-pytorch.

Thank you very much for your answer, which gave me a great help!!! :)

from dalle2-pytorch.

Update:
Got it running, but receive pure noise as output. Any tips?

from dalle2-pytorch.

hey, did you get the correct output without noise? I still dont know how to run the code, would you please provide your train code? thanks!!

from dalle2-pytorch.

Hello, yes.

I'm still an amateur, but I got it working.

Attached is basically my whole model. There are definitely a few inconsistencies, but after I put DallE together, it produces images as I envisioned them.

You can read about how to "assemble" DallE here: #282

You load the trainer checkpoints as in the code I provided and add a ["model"] after it as follows: torch.load(checkpoint_path)['model']

Thats it.

Model.zip

from dalle2-pytorch.

Thank you very much! It helps me a lot!!
By the way, how many images does your dataset have? How many epochs does it take to get a good result?
I was a complete newcomer and had no idea about any of this.
It would be great if you could answer me!!^ ^

from dalle2-pytorch.

FIY: had an Error in the Priortrainer. You need to insert text_loader = iter(text_dataset) image_loader = iter(image_dataset) one loop above. After: for epoch in range(num_epochs):

Otherwise you would just train your Prior from one image the whole time...

from dalle2-pytorch.