Li Haipeng
Li Haipeng
Thanks for your job, it's really excellent! I created a PullRequest to update the database, and hope it helps ;)
yes quite a few, e.g. 1. use gyroscope to run in realtime and high quality https://graphics.stanford.edu/papers/stabilization/ 2. estimate meshflow model to online dvs http://www.liushuaicheng.org/eccv2016/ 3. compute a mixtures model to...
Got the same issue on cifar10, based on settings (bs=1024): ``` step,fid,kid 10000,54.75194549560547,0.024161577224731445 20000,38.80967330932617,0.01055598258972168 30000,35.451629638671875,0.008167028427124023 40000,32.956600189208984,0.005323648452758789 50000,32.06228256225586,0.004504680633544922 ```
It mainly depends on your batch size and input resolution, such as ``` image_size=128, train_batch_size=14, it takes about 11G GPU memory ```
@lucidrains Okay, Let's find time to implement this feature
did you produce the colored noise such as:  I think I have also failed to produce a normal image
[issue62](https://github.com/lucidrains/denoising-diffusion-pytorch/issues/134) as mentioned, I solved the problem, hope it help
I got the same question In my mind, `posterior_variance` is the one that we need. as 'pred_img = x_start + posterior_variance * noise'
Thanks for your correction. We will fix this error. ；）
Hi, 1. I think it doesn't matter what cuda version you use, have a try with any matching cuda version. 2. The code seems to be correct, the problem is...