Create Colab notebook with `denoising_diffusion_pytorch` example about xtal2png HOT 7 CLOSED

Not working on Colab lucidrains/denoising-diffusion-pytorch#33

from xtal2png.

Working on this as a script for now: https://github.com/sparks-baird/xtal2png/blob/main/notebooks/ddpm.py

Has been producing less favorable results #79 #80. The priority is to get bookends attached #12 so that hyperparameter optimization can take place on several models.

from xtal2png.

Hi, i'm working con Colab too. How can i print inception score (IS), FID score and NLL (Negative Log Likelihoods) during the training for comparing them to the values written in the article?

from xtal2png.

@vinz97 by the article do you mean the blog post? I'd suggest asking this as an issue on https://github.com/lucidrains/denoising-diffusion-pytorch. I briefly perused the code and previous issues and didn't find the answer. As a side note, for FID you might consider using https://github.com/GaParmar/clean-fid. For inception score, you might consider https://github.com/sbarratt/inception-score-pytorch, though they recommend against using it as a performance metric for generative models. For NLL,
learned_gaussian_diffusion.py#L119-L120 is probably relevant. Keep in mind that it saves checkpoint models, so you could also post-process, loading each model, and calculating this information as needed unless real-time visualization is critical for you.

You might also consider using imagen-pytorch in image-to-image mode.

from xtal2png.

@vinz97 btw I answered assuming you weren't asking about anything specific to xtal2png. If that wasn't the case, please let me know.

from xtal2png.

@sgbaird Yeah sorry for the confusing question, i was asking about the code https://github.com/lucidrains/denoising-diffusion-pytorch.
I'm using the code written in the README file of the repository (under the USAGE paragraph, the second one). My aim is to training the CIFAR10 dataset and trying to print the Inception score, FID and NLL for comparing them to the values written in this article.
But i can use only Colab, so the training is really slowly, especially for 700k steps

from xtal2png.

See https://github.com/sparks-baird/xtal2png/blob/main/scripts/denoising_diffusion_pytorch_example.py and https://github.com/sparks-baird/xtal2png/blob/main/scripts/denoising_diffusion_pytorch_pretrained_sample.py instead

from xtal2png.