coderpiaobozhe / classifier-free-diffusion-guidance-pytorch Goto Github PK

Hi! I'm excited to find such excellent work, but unfortunately, I failed to obtain conditional images following your instruction. I have run the train.py for 1000 epochs, but I found output like img1 which seems random after I set --label range

Therefore, I modified part of the code in sample.py shown in img2 so that I can output single-kind pictures ranging from 0 to 9

and ran orders like " python sample.py --label 0 --device cuda:1 --epc 999 --w 5.0 --T 1000", but still got random images shown in img3.

I wonder if you could help out with this problem. Thanks a lot!

Thanks for your impressive work! I am still wondering if you have tested the cosine noise scheduler. I have tried the cosine noise scheduler but the result is kind of wired.

make: *** No rule to make target 'train'. Stop.
When I run the make train command, this is what I get.
I also tried running it with python train.py to no avail.

Once we trained the model, is there a way to obtain log_prob(x) or log_prob(x|y)?

Can you tell me why cat tilde_betas and betas? Why not just use betas？

Hi, Thanks alot for the great work.
Can you also share you training log?
How many epochs did you train the model for?
what loss did you ultimately settle for?
Thanks a lot

This should be filled with a value close to 1 instead of 0？

I tried twice, set epoch=20 and 300, after traing, the sampled result are all random noise.
I didn't change any other parameter, is there anything wrong with the training process?

I run the code just without editing and meet this error.
What is the problem..?

Traceback (most recent call last):
File "train.py", line 9, in
from diffusion import GaussianDiffusion
File "/workspace/sanghyu.yoon/nfsdata/home/sanghyu.yoon/code/research_23/classifier-free-diffusion-guidance-Pytorch/diffusion.py", line 8, in
class GaussianDiffusion(nn.Module):
File "/workspace/sanghyu.yoon/nfsdata/home/sanghyu.yoon/code/research_23/classifier-free-diffusion-guidance-Pytorch/diffusion.py", line 79, in GaussianDiffusion
def q_mean_variance(self, x_0:torch.Tensor, t:torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
TypeError: 'type' object is not subscriptable

I try an implementation of the classifier-free diffusion which is mainly adapted from the classifier-guided diffusion and using your training hyperparameters on CiFar10. But I got a FID score between 4.1 and 4.2 when the w was set to 0.1 and 0.2. It's worse than the original DDPM. Could you please tell me your sample quality? Thank you very much!