Hi,

I think you have an error in:

while it should be q(x_t|x_0).

I would like to know if the fourth part of this series is available now. If it is available, can you put it here or share the link with us?

Why do you shift alphas_prod_p in diffusion_02_model.ipynb?

alphas_prod_p = torch.cat([torch.tensor([1]).float(), alphas_prod[:-1]], 0)

Hello, thank you for your tutorial. It helps me a lot to understand the model. I have got a small question.
Since we set the n_steps to 100, the x_seq contains a random noise and 100 following x for each denoising step. When we plot, we plot the x distribution after every 10 steps. Why in the figure, the labels are from 100 to 1000 rather than 10 to 100? Did I miss something?

Thanks for the very nice tutorial.

I found that you may want to use normal distribution for noise term but use "rand" instand of "randn"? In the 6-th block of https://github.com/acids-ircam/diffusion_models/blob/main/diffusion_02_model.ipynb

Bests!

Which one's right? I guess there should be a '-' sign

def denoising_score_matching(scorenet, samples, key, sigma=0.01):
    noise = jax.random.normal(key, samples.shape)
    perturbed_samples = samples + noise * sigma
    target = -noise / sigma
    scores = scorenet(perturbed_samples)
    loss = 1 / 2. * ((scores - target) ** 2).sum(axis=-1).mean(axis=0)
    return loss

denoising_score_matching(model, data[:10], jax.random.PRNGKey(0))

Here target = -noise/sigma should be replaced with target = -noise

In notebooks I see updates of EMA model def update, where we change shallow model, but def ema method was never called. It should be called ones per N epochs or it will be used at the end of all training process? As I can see, def update does not affect actual model and smoothing is not happen.